diff --git a/fish-rust/src/fd_monitor.rs b/fish-rust/src/fd_monitor.rs
index 90c00580a..78d002e41 100644
--- a/fish-rust/src/fd_monitor.rs
+++ b/fish-rust/src/fd_monitor.rs
@@ -4,7 +4,7 @@ use std::sync::{Arc, Mutex};
 use std::time::{Duration, Instant};
 
 pub use self::fd_monitor_ffi::ItemWakeReason;
-use self::fd_monitor_ffi::{new_fd_event_signaller, FdEventSignaller};
+pub use self::fd_monitor_ffi::{new_fd_event_signaller, FdEventSignaller};
 use crate::fd_readable_set::FdReadableSet;
 use crate::fds::AutoCloseFd;
 use crate::ffi::void_ptr;
diff --git a/fish-rust/src/threads.rs b/fish-rust/src/threads.rs
index 571662083..0107a4de6 100644
--- a/fish-rust/src/threads.rs
+++ b/fish-rust/src/threads.rs
@@ -2,8 +2,12 @@
 //! ported directly from the cpp code so we can use rust threads instead of using pthreads.
 
 use crate::flog::{FloggableDebug, FLOG};
+use once_cell::race::OnceBox;
+use std::num::NonZeroU64;
 use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::{Arc, Mutex};
 use std::thread::{self, ThreadId};
+use std::time::{Duration, Instant};
 
 impl FloggableDebug for ThreadId {}
 
@@ -22,6 +26,39 @@ static THREAD_ASSERTS_CFG_FOR_TESTING: AtomicBool = AtomicBool::new(false);
 /// This allows us to notice when we've forked.
 static IS_FORKED_PROC: AtomicBool = AtomicBool::new(false);
 
+/// Maximum number of threads for the IO thread pool.
+const IO_MAX_THREADS: usize = 1024;
+
+/// How long an idle [`ThreadPool`] thread will wait for work (against the condition variable)
+/// before exiting.
+const IO_WAIT_FOR_WORK_DURATION: Duration = Duration::from_millis(500);
+
+/// The iothreads [`ThreadPool`] singleton. Used to lift I/O off of the main thread and used for
+/// completions, etc.
+static IO_THREAD_POOL: OnceBox<Mutex<ThreadPool>> = OnceBox::new();
+
+/// The event signaller singleton used for completions and queued main thread requests.
+static NOTIFY_SIGNALLER: once_cell::sync::Lazy<&'static crate::fd_monitor::FdEventSignaller> =
+    once_cell::sync::Lazy::new(|| unsafe {
+        // This is leaked to avoid C++-side destructors. When ported fully to rust, we won't need to
+        // leak anything.
+        let signaller = crate::fd_monitor::new_fd_event_signaller();
+        let signaller_ref: &crate::fd_monitor::FdEventSignaller = signaller.as_ref().unwrap();
+        let result = std::mem::transmute(signaller_ref);
+        std::mem::forget(signaller);
+        result
+    });
+
+/// A [`ThreadPool`] or [`Debounce`] work request.
+type WorkItem = Box<dyn FnOnce() + 'static + Send>;
+
+/// The queue of [`WorkItem`]s to be executed on the main thread. This is added to from
+/// [`Debounce::enqueue_main_thread()`] and read from in `iothread_service_main()`.
+///
+/// Since items are enqueued from various background threads then read by the main thread, the work
+/// items must implement `Send`.
+static MAIN_THREAD_QUEUE: Mutex<Vec<WorkItem>> = Mutex::new(Vec::new());
+
 /// Initialize some global static variables. Must be called at startup from the main thread.
 pub fn init() {
     unsafe {
@@ -38,6 +75,10 @@ pub fn init() {
         let result = libc::pthread_atfork(None, None, Some(child_post_fork));
         assert_eq!(result, 0, "pthread_atfork() failure: {}", errno::errno());
     }
+
+    IO_THREAD_POOL
+        .set(Box::new(Mutex::new(ThreadPool::new(1, IO_MAX_THREADS))))
+        .expect("IO_THREAD_POOL has already been initialized!");
 }
 
 #[inline(always)]
@@ -153,6 +194,419 @@ pub fn spawn<F: FnOnce() + Send + 'static>(callback: F) -> bool {
     result
 }
 
+/// Data shared between the thread pool [`ThreadPool`] and worker threads [`WorkerThread`].
+#[derive(Default)]
+struct ThreadPoolProtected {
+    /// The queue of outstanding, unclaimed work requests
+    pub request_queue: std::collections::VecDeque<WorkItem>,
+    /// The number of threads that exist in the pool
+    pub total_threads: usize,
+    /// The number of threads waiting for more work (i.e. idle threads)
+    pub waiting_threads: usize,
+}
+
+/// Data behind an [`Arc`] to share between the [`ThreadPool`] and [`WorkerThread`] instances.
+#[derive(Default)]
+struct ThreadPoolShared {
+    /// The mutex to access shared state between [`ThreadPool`] and [`WorkerThread`] instances. This
+    /// is accessed both standalone and via [`cond_var`](Self::cond_var).
+    mutex: Mutex<ThreadPoolProtected>,
+    /// The condition variable used to wake up waiting threads. This is tied to [`mutex`](Self::mutex).
+    cond_var: std::sync::Condvar,
+}
+
+pub struct ThreadPool {
+    /// The data which needs to be shared with worker threads.
+    shared: Arc<ThreadPoolShared>,
+    /// The minimum number of threads that will be kept waiting even when idle in the pool.
+    soft_min_threads: usize,
+    /// The maximum number of threads that will be created to service outstanding work requests, by
+    /// default. This may be bypassed.
+    max_threads: usize,
+}
+
+impl std::fmt::Debug for ThreadPool {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("ThreadPool")
+            .field("min_threads", &self.soft_min_threads)
+            .field("max_threads", &self.max_threads)
+            .finish()
+    }
+}
+
+impl ThreadPool {
+    /// Construct a new `ThreadPool` instance with the specified min and max num of threads.
+    pub fn new(soft_min_threads: usize, max_threads: usize) -> Self {
+        ThreadPool {
+            shared: Default::default(),
+            soft_min_threads,
+            max_threads,
+        }
+    }
+
+    /// Enqueue a new work item onto the thread pool.
+    ///
+    /// The function `func` will execute on one of the pool's background threads. If `cant_wait` is
+    /// set, the thread limit may be disregarded if extant threads are busy.
+    ///
+    /// Returns the number of threads that were alive when the work item was enqueued.
+    pub fn perform<F: FnOnce() + 'static + Send>(&mut self, func: F, cant_wait: bool) -> usize {
+        let work_item = Box::new(func);
+        self.perform_inner(work_item, cant_wait)
+    }
+
+    fn perform_inner(&mut self, f: WorkItem, cant_wait: bool) -> usize {
+        enum ThreadAction {
+            None,
+            Wake,
+            Spawn,
+        }
+
+        let local_thread_count;
+        let thread_action = {
+            let mut data = self.shared.mutex.lock().expect("Mutex poisoned!");
+            local_thread_count = data.total_threads;
+            data.request_queue.push_back(f);
+            FLOG!(
+                iothread,
+                "enqueuing work item (count is ",
+                data.request_queue.len(),
+                ")"
+            );
+            if data.waiting_threads >= data.request_queue.len() {
+                // There are enough waiting threads, wake one up.
+                ThreadAction::Wake
+            } else if cant_wait || data.total_threads < self.max_threads {
+                // No threads are idle waiting but we can or must spawn a new thread to service the
+                // request.
+                data.total_threads += 1;
+                ThreadAction::Spawn
+            } else {
+                // There is no need to do anything because we've reached the max number of threads.
+                ThreadAction::None
+            }
+        };
+
+        // Act only after unlocking the mutex.
+        match thread_action {
+            ThreadAction::None => (),
+            ThreadAction::Wake => {
+                // Wake a thread if we decided to do so.
+                FLOG!(iothread, "notifying thread ", std::thread::current().id());
+                self.shared.cond_var.notify_one();
+            }
+            ThreadAction::Spawn => {
+                // Spawn a thread. If this fails, it means there are already a bunch of worker
+                // threads and it is very unlikely that they are all about to exit so one is likely
+                // able to handle the incoming request. This means we can ignore the failure with
+                // some degree of confidence. (This is also not an error we expect to routinely run
+                // into under normal, non-resource-starved circumstances.)
+                if self.spawn_thread() {
+                    FLOG!(iothread, "pthread spawned");
+                } else {
+                    // We failed to spawn a thread; decrement the thread count.
+                    self.shared
+                        .mutex
+                        .lock()
+                        .expect("Mutex poisoned!")
+                        .total_threads -= 1;
+                }
+            }
+        }
+
+        local_thread_count
+    }
+
+    /// Attempt to spawn a new worker thread.
+    fn spawn_thread(&mut self) -> bool {
+        let shared = Arc::clone(&self.shared);
+        let soft_min_threads = self.soft_min_threads;
+        self::spawn(move || {
+            let worker = WorkerThread {
+                shared,
+                soft_min_threads,
+            };
+
+            worker.run();
+        })
+    }
+}
+
+pub struct WorkerThread {
+    /// The data shared with the [`ThreadPool`].
+    shared: Arc<ThreadPoolShared>,
+    /// The soft min number of threads for the associated [`ThreadPool`].
+    soft_min_threads: usize,
+}
+
+impl WorkerThread {
+    /// The worker loop entry point for this thread.
+    fn run(mut self) {
+        while let Some(work_item) = self.dequeue_work_or_commit_to_exit() {
+            FLOG!(
+                iothread,
+                "pthread ",
+                std::thread::current().id(),
+                " got work"
+            );
+
+            // Perform the work
+            work_item();
+        }
+
+        FLOG!(
+            iothread,
+            "pthread ",
+            std::thread::current().id(),
+            " exiting"
+        );
+    }
+
+    /// Dequeue a work item (perhaps waiting on the condition variable) or commit to exiting by
+    /// reducing the active thread count.
+    fn dequeue_work_or_commit_to_exit(&mut self) -> Option<WorkItem> {
+        let mut data = self.shared.mutex.lock().expect("Mutex poisoned!");
+
+        // If the queue is empty, check to see if we should wait. We should wait if our exiting
+        // would drop us below our soft thread count minimum.
+        if data.request_queue.is_empty()
+            && data.total_threads == self.soft_min_threads
+            && IO_WAIT_FOR_WORK_DURATION > Duration::ZERO
+        {
+            data.waiting_threads += 1;
+            data = self
+                .shared
+                .cond_var
+                .wait_timeout(data, IO_WAIT_FOR_WORK_DURATION)
+                .expect("Mutex poisoned!")
+                .0;
+            data.waiting_threads -= 1;
+        }
+
+        // Now that we've (perhaps) waited, see if there's something on the queue.
+        let result = data.request_queue.pop_front();
+
+        // If we are returning None then ensure we balance the thread count increment from when we
+        // were created. This has to be done here in this awkward place because we've already
+        // committed to exiting - we will never pick up more work. So we need to make sure to
+        // decrement the thread count while holding the lock as we have effectively already exited.
+        if result.is_none() {
+            data.total_threads -= 1;
+        }
+
+        return result;
+    }
+}
+
+/// Returns a [`MutexGuard`](std::sync::MutexGuard) containing the IO [`ThreadPool`].
+fn borrow_io_thread_pool() -> std::sync::MutexGuard<'static, ThreadPool> {
+    IO_THREAD_POOL
+        .get()
+        .unwrap()
+        .lock()
+        .expect("Mutex poisoned!")
+}
+
+/// Enqueues work on the IO thread pool singleton.
+pub fn iothread_perform(f: impl FnOnce() + 'static + Send) {
+    let mut thread_pool = borrow_io_thread_pool();
+    thread_pool.perform(f, false);
+}
+
+/// Enqueues priority work on the IO thread pool singleton, disregarding the thread limit.
+///
+/// It does its best to spawn a thread if all other threads are occupied. This is primarily for
+/// cases where deferring creation of a new thread might lead to a deadlock.
+pub fn iothread_perform_cant_wait(f: impl FnOnce() + 'static + Send) {
+    let mut thread_pool = borrow_io_thread_pool();
+    thread_pool.perform(f, true);
+}
+
+pub fn iothread_service_main_with_timeout(timeout: Duration) {
+    if crate::fd_readable_set::is_fd_readable(
+        i32::from(NOTIFY_SIGNALLER.read_fd()),
+        timeout.as_millis() as u64,
+    ) {
+        iothread_service_main();
+    }
+}
+
+pub fn iothread_service_main() {
+    self::assert_is_main_thread();
+
+    // Note: the order here is important. We must consume events before handling requests, as
+    // posting uses the opposite order.
+    NOTIFY_SIGNALLER.try_consume();
+
+    // Move the queue to a local variable. The MAIN_THREAD_QUEUE lock is not held after this.
+    let queue = std::mem::take(&mut *MAIN_THREAD_QUEUE.lock().expect("Mutex poisoned!"));
+
+    // Perform each completion in order.
+    for func in queue {
+        (func)();
+    }
+}
+
+/// Does nasty polling via select() and marked as unsafe because it should only be used for testing.
+pub unsafe fn iothread_drain_all() {
+    while borrow_io_thread_pool()
+        .shared
+        .mutex
+        .lock()
+        .expect("Mutex poisoned!")
+        .total_threads
+        > 0
+    {
+        iothread_service_main_with_timeout(Duration::from_millis(1000));
+    }
+}
+
+/// `Debounce` is a simple class which executes one function on a background thread while enqueing
+/// at most one more. Subsequent execution requests overwrite the enqueued one. It takes an optional
+/// timeout; if a handler does not finish within the timeout then a new thread is spawned to service
+/// the remaining request.
+///
+/// Debounce implementation note: we would like to enqueue at most one request, except if a thread
+/// hangs (e.g. on fs access) then we do not want to block indefinitely - such threads are called
+/// "abandoned". This is implemented via a monotone uint64 counter, called a token. Every time we
+/// spawn a thread, we increment the token. When the thread has completed running a work item, it
+/// compares its token to the active token; if they differ then this thread was abandoned.
+#[derive(Clone)]
+pub struct Debounce {
+    timeout: Duration,
+    /// The data shared between [`Debounce`] instances.
+    data: Arc<Mutex<DebounceData>>,
+}
+
+/// The data shared between [`Debounce`] instances.
+struct DebounceData {
+    /// The (one or none) next enqueued request, overwritten each time a new call to
+    /// [`perform()`](Self::perform) is made.
+    next_req: Option<WorkItem>,
+    /// The non-zero token of the current non-abandoned thread or `None` if no thread is running.
+    active_token: Option<NonZeroU64>,
+    /// The next token to use when spawning a thread.
+    next_token: NonZeroU64,
+    /// The start time of the most recently spawned thread or request (if any).
+    start_time: Instant,
+}
+
+impl Debounce {
+    pub fn new(timeout: Duration) -> Self {
+        Self {
+            timeout,
+            data: Arc::new(Mutex::new(DebounceData {
+                next_req: None,
+                active_token: None,
+                next_token: NonZeroU64::new(1).unwrap(),
+                start_time: Instant::now(),
+            })),
+        }
+    }
+
+    /// Run an iteration in the background with the given thread token. Returns `true` if we handled
+    /// a request or `false` if there were no requests to handle (in which case the debounce thread
+    /// exits).
+    ///
+    /// Note that this method is called from a background thread.
+    fn run_next(&self, token: NonZeroU64) -> bool {
+        let request = {
+            let mut data = self.data.lock().expect("Mutex poisoned!");
+            if let Some(req) = data.next_req.take() {
+                data.start_time = Instant::now();
+                req
+            } else {
+                // There is no pending request. Mark this token as no longer running.
+                if Some(token) == data.active_token {
+                    data.active_token = None;
+                }
+                return false;
+            }
+        };
+
+        // Execute request after unlocking the mutex.
+        (request)();
+        return true;
+    }
+
+    /// Enqueue `handler` to be performed on a background thread. If another function is already
+    /// enqueued, this overwrites it and that function will not be executed.
+    ///
+    /// The result is a token which is only of interest to the test suite.
+    pub fn perform(&self, handler: impl FnOnce() + 'static + Send) -> NonZeroU64 {
+        let h = Box::new(handler);
+        self.perform_inner(h)
+    }
+
+    /// Enqueue `handler` to be performed on a background thread with [`Completion`] `completion`
+    /// to be performed on the main thread. If a function is already enqueued, this overwrites it
+    /// and that function will not be executed.
+    ///
+    /// If the function executes within the optional timeout then `completion` will be invoked on
+    /// the main thread with the result of the evaluated `handler`.
+    ///
+    /// The result is a token which is only of interest to the test suite.
+    pub fn perform_with_completion<H, R, C>(&self, handler: H, completion: C) -> NonZeroU64
+    where
+        H: FnOnce() -> R + 'static + Send,
+        C: FnOnce(R) + 'static + Send,
+        R: 'static + Send,
+    {
+        let h = Box::new(move || {
+            let result = handler();
+            let c = Box::new(move || {
+                (completion)(result);
+            });
+            Self::enqueue_main_thread_result(c);
+        });
+        self.perform_inner(h)
+    }
+
+    fn perform_inner(&self, handler: WorkItem) -> NonZeroU64 {
+        let mut spawn = false;
+        let active_token = {
+            let mut data = self.data.lock().expect("Mutex poisoned!");
+            data.next_req = Some(handler);
+            // If we have a timeout and our running thread has exceeded it, abandon that thread.
+            if data.active_token.is_some()
+                && !self.timeout.is_zero()
+                && (Instant::now() - data.start_time > self.timeout)
+            {
+                // Abandon this thread by dissociating its token from this [`Debounce`] instance.
+                data.active_token = None;
+            }
+            if data.active_token.is_none() {
+                // We need to spawn a new thread. Mark the current time so that a new request won't
+                // immediately abandon us and start a new thread too.
+                spawn = true;
+                data.active_token = Some(data.next_token);
+                data.next_token = data.next_token.checked_add(1).unwrap();
+                data.start_time = Instant::now();
+            }
+            data.active_token.expect("Something should be active now.")
+        };
+
+        // Spawn after unlocking the mutex above.
+        if spawn {
+            // We need to clone the Arc to get it to last for the duration of the 'static lifetime.
+            let debounce = self.clone();
+            iothread_perform(move || {
+                while debounce.run_next(active_token) {
+                    // Keep thread alive/busy.
+                }
+            });
+        }
+
+        active_token
+    }
+
+    /// Static helper to add a [`WorkItem`] to [`MAIN_THREAD_ID`] and signal [`NOTIFY_SIGNALLER`].
+    fn enqueue_main_thread_result(f: WorkItem) {
+        MAIN_THREAD_QUEUE.lock().expect("Mutex poisoned!").push(f);
+        NOTIFY_SIGNALLER.post();
+    }
+}
+
 #[test]
 /// Verify that spawing a thread normally via [`std::thread::spawn()`] causes the calling thread's
 /// sigmask to be inherited by the newly spawned thread.