diff --git a/src/next.rs b/src/next.rs
index 09877e2..6f18644 100644
--- a/src/next.rs
+++ b/src/next.rs
@@ -2,11 +2,11 @@ use std::{
     time::{Instant, Duration},
     collections::VecDeque,
     marker::PhantomData,
-    sync::{Mutex, Condvar, Arc, atomic::AtomicUsize}, fmt, error,
+    sync::{Mutex, Condvar, Arc, MutexGuard},
+    fmt,
+    error,
 };
 
-// TODO: shut down debouncer on drop
-
 /// Creates a new debouncer for deduplicating groups of "raw" events which occur at a similar time.
 /// The debouncer is comprised of two halves; a [`DebouncerTx`](DebouncerTx) for sending raw events
 /// to the debouncer, and a [`DebouncerRx`](DebouncerRx) for receiving grouped (debounced) events
@@ -52,14 +52,7 @@ pub fn debouncer<R, D, F>(debounce_time: Duration, fold: F)
 where
     F: Fn(Option<D>, R) -> D,
 {
-    let shared_state = Arc::new(Debouncer {
-        state: Mutex::new(DebouncerState::new()),
-        debounce_time,
-        queue_wait_cvar: Condvar::new(),
-        event_ready_wait_cvar: Condvar::new(),
-        tx_count: AtomicUsize::new(1),
-        rx_count: AtomicUsize::new(1),
-    });
+    let shared_state = Arc::new(Debouncer::new(debounce_time, 1, 1));
     
     let tx = DebouncerTx {
         debouncer: shared_state.clone(),
@@ -89,12 +82,16 @@ where
     }
 }
 
-// FIXME: increment some sort of tx count
 impl<R, D, F> Clone for DebouncerTx<R, D, F>
 where
     F: Clone,
 {
     fn clone(&self) -> Self {
+        self.debouncer
+            .lock_state()
+            .add_tx()
+            .expect("debouncer tx count should not overflow");
+        
         Self {
             debouncer: self.debouncer.clone(),
             fold: self.fold.clone(),
@@ -103,10 +100,20 @@ where
     }
 }
 
-// FIXME: check tx count and shut down if zero
 impl<R, D, F> Drop for DebouncerTx<R, D, F> {
     fn drop(&mut self) {
-        todo!()
+        let remaining_tx = {
+            let mut state_guard = self.debouncer.lock_state();
+            state_guard.remove_tx()
+        };
+
+        if remaining_tx == 0 {
+            // There may be some rx threads waiting on the condvars, so notify them to stop them
+            // from waiting. Upon waking up, the rx threads will see that the tx count is 0 and
+            // switch to their "shutdown" behaviour.
+            self.debouncer.event_ready_wait_cvar.notify_all();
+            self.debouncer.queue_wait_cvar.notify_all();
+        }
     }
 }
 
@@ -124,19 +131,25 @@ impl<D> DebouncerRx<D> {
     }
 }
 
-// FIXME: increment some sort of rx count
 impl<D> Clone for DebouncerRx<D> {
     fn clone(&self) -> Self {
+        self.debouncer
+            .lock_state()
+            .add_rx()
+            .expect("debouncer rx count should not overflow");
+
         Self {
             debouncer: self.debouncer.clone(),
         }
     }
 }
 
-// FIXME: check rx count and shut down if zero
 impl<D> Drop for DebouncerRx<D> {
     fn drop(&mut self) {
-        todo!()
+        // Decrement the rx count. We don't need to notify any threads waiting on condvars if the
+        // count reaches 0, because only rx threads wait on the condvars and we know there are no
+        // more rx threads (because the count just reached 0!)
+        self.debouncer.lock_state().remove_rx();
     }
 }
 
@@ -172,11 +185,25 @@ struct Debouncer<T> {
     debounce_time: Duration,
     queue_wait_cvar: Condvar,
     event_ready_wait_cvar: Condvar,
-    tx_count: AtomicUsize,
-    rx_count: AtomicUsize,
 }
 
 impl<T> Debouncer<T> {
+    fn new(debounce_time: Duration, tx_count: usize, rx_count: usize) -> Self {
+        Self {
+            state: Mutex::new(DebouncerState::new(tx_count, rx_count)),
+            debounce_time,
+            queue_wait_cvar: Condvar::new(),
+            event_ready_wait_cvar: Condvar::new(),
+        }
+    }
+
+    fn lock_state(&self) -> MutexGuard<DebouncerState<T>> {
+        match self.state.lock() {
+            Ok(guard) => guard,
+            Err(err) => err.into_inner(),
+        }
+    }
+    
     fn push<R, F>(&self, raw_event: R, fold: F) -> Result<(), SendError<R>>
     where
         F: Fn(Option<T>, R) -> T,
@@ -184,11 +211,11 @@ impl<T> Debouncer<T> {
         let now = Instant::now();
 
         let push_outcome = {
-            let mut state_guard = self.state.lock().unwrap();
+            let mut state_guard = self.lock_state();
             
-            // Return an error if the debouncer is closed. Include the raw event in the error so
-            // that it isn't lost.
-            if state_guard.shutdown {
+            // Return an error if there are no rxs left to send to. Include the raw event in the
+            // error so that it isn't lost.
+            if state_guard.has_no_rxs() {
                 return Err(SendError(raw_event));
             }
 
@@ -212,10 +239,12 @@ impl<T> Debouncer<T> {
             Shutdown,
         }
 
-        let mut state_guard = self.state.lock().unwrap();
+        let mut state_guard = self.lock_state();
 
         let result = 'result: {
-            if state_guard.shutdown {
+            // Check that there are any txs left so we don't get stuck waiting forever if the queue
+            // is empty (as a tx is the only thing that can wake us up).
+            if state_guard.has_no_txs() {
                 break 'result PopOutcome::Shutdown;
             }
 
@@ -225,17 +254,23 @@ impl<T> Debouncer<T> {
             //    wait for it to be ready.
             // 3. Before we can wake up from waiting, some other thread pops the now-ready `x`, and
             //    the queue is now empty.
-            // 4. We wake up and must wait for the queue to become non-empty again.  
+            // 4. We wake up and must wait for the queue to become non-empty again.
+            //
+            // Additionally, the loop also handles spurious returns from the condvar wait.
             'pop_event_outer_loop: loop {
                 // If there are no accumulators in the queue, wait for one to be pushed.
                 if state_guard.acc_queue.is_empty() {
                     // Park the thread. We will be woken up again either by a new accumulator being
                     // pushed to the queue, or by the debouncer being shut down.
+                    // 
+                    // This may return spuriously so the queue may still be empty when we wake up,
+                    // but if this happens we will retry the wait when we attempt and fail to peek
+                    // the queue later.
                     state_guard = self.queue_wait_cvar.wait(state_guard).unwrap();
     
                     // We may have been unparked because someone wants to shut down the debouncer,
                     // so check the shutdown flag.
-                    if state_guard.shutdown {
+                    if state_guard.has_no_txs() {
                         break PopOutcome::Shutdown;
                     }
                 }
@@ -253,7 +288,9 @@ impl<T> Debouncer<T> {
                         // popped.
                         let Some(peeked_acc) = state_guard.acc_queue.peek_oldest() else {
                             // If there is no accumulator for us to pop from the queue, go back to
-                            // waiting for the queue to be non-empty.
+                            // waiting for the queue to be non-empty. This could happen because the
+                            // previous condvar wait returned spuriously, or because another thread
+                            // popped the accumulator before us.
                             continue 'pop_event_outer_loop;
                         };
         
@@ -270,13 +307,19 @@ impl<T> Debouncer<T> {
                             // Wait the amount of time between now and the `ready_time` of the
                             // accumulator. This is done using a condvar so the sleep can be
                             // interrupted if someone wants to shut down the debouncer.
+                            //
+                            // This may return spuriously so we may wake up before the time has
+                            // elapsed and without being notified by anyone. However, since we go
+                            // back so the start of the loop after this, we will see that there is
+                            // still time remaining before the accumulator becomes ready, and
+                            // therefore we will resume waiting.
                             (state_guard, _) = self.event_ready_wait_cvar
                                 .wait_timeout(state_guard, wait_time)
                                 .unwrap();
     
                             // Again, we may have been unparked because someone wants to shut down the
                             // debouncer, so check the shutdown flag and return if it is set.
-                            if state_guard.shutdown {
+                            if state_guard.has_no_txs() {
                                 break PopOutcome::Shutdown;
                             }
         
@@ -317,11 +360,11 @@ impl<T> Debouncer<T> {
     }
 
     fn try_pop(&self) -> Result<Option<T>, ReceiveError> {
-        let mut state_guard = self.state.lock().unwrap();
+        let mut state_guard = self.lock_state();
 
         // If the debouncer has been shut down, return any remaining accumulators then start
         // returning errors once the accumulators have been depleted.
-        if state_guard.shutdown {
+        if state_guard.has_no_txs() {
             return state_guard
                 .acc_queue
                 .pop_oldest_acc_discard_none()
@@ -356,16 +399,53 @@ impl<T> Debouncer<T> {
 
 struct DebouncerState<T> {
     acc_queue: EventAccQueue<T>,
-    shutdown: bool,
+    // These could alternatively be `AtomicUsize`s which live outside the mutex, but we opt to have
+    // them inside the mutex instead to avoid a confusing mixture of mutexes / condvars and
+    // atomics.
+    tx_count: usize,
+    rx_count: usize,
 }
 
 impl<T> DebouncerState<T> {
-    fn new() -> Self {
+    fn new(tx_count: usize, rx_count: usize) -> Self {
         Self {
             acc_queue: EventAccQueue::new(),
-            shutdown: false,
+            tx_count,
+            rx_count,
         }
     }
+
+    fn has_no_txs(&self) -> bool {
+        self.tx_count == 0
+    }
+
+    fn add_tx(&mut self) -> Result<(), CountOverflowError> {
+        self.tx_count = self.tx_count.checked_add(1)
+            .ok_or(CountOverflowError)?;
+
+        Ok(())
+    }
+
+    fn remove_tx(&mut self) -> usize {
+        self.tx_count = self.tx_count.saturating_sub(1);
+        self.tx_count
+    }
+
+    fn has_no_rxs(&self) -> bool {
+        self.rx_count == 0
+    }
+
+    fn add_rx(&mut self) -> Result<(), CountOverflowError> {
+        self.rx_count = self.rx_count.checked_add(1)
+            .ok_or(CountOverflowError)?;
+
+        Ok(())
+    }
+
+    fn remove_rx(&mut self) -> usize {
+        self.rx_count = self.rx_count.saturating_sub(1);
+        self.rx_count
+    }
 }
 
 struct EventAccQueue<T> {
@@ -468,3 +548,5 @@ impl<T> EventAcc<T> {
     }
 }
 
+#[derive(Debug)]
+struct CountOverflowError;