diff --git a/src/rt/atomic.rs b/src/rt/atomic.rs
index 7f63417..392f707 100644
--- a/src/rt/atomic.rs
+++ b/src/rt/atomic.rs
@@ -118,6 +118,12 @@ pub(super) struct State {
     /// Last time the atomic was accessed for a store or rmw operation.
     last_non_load_access: Option<Access>,
 
+    /// Last time the atomic was accessed using an RMW access that did not fail.
+    ///
+    /// This is used for applying the RMW atomicity constraint to other stores
+    /// (so they don't squeeze in between the read and write part of the RMW).
+    last_successful_rmw: VersionVec,
+
     /// Currently tracked stored values. This is the `MAX_ATOMIC_HISTORY` most
     /// recent stores to the atomic cell in loom execution order.
     stores: [Store; MAX_ATOMIC_HISTORY],
@@ -413,6 +419,7 @@ impl State {
             is_mutating: false,
             last_access: None,
             last_non_load_access: None,
+            last_successful_rmw: VersionVec::new(),
             stores: Default::default(),
             cnt: 0,
         };
@@ -482,6 +489,9 @@ impl State {
             }
         }
 
+        // RMW Atomicity.
+        modification_order.join(&self.last_successful_rmw);
+
         sync.sync_store(threads, ordering);
 
         let mut first_seen = FirstSeen::new();
@@ -535,6 +545,10 @@ impl State {
                 let sync = self.stores[index].sync;
                 self.store(threads, sync, next, success);
 
+                // RMW Atomicity. Future stores must happen after this, because we
+                // already observed a value.
+                self.last_successful_rmw.join(&threads.active().causality);
+
                 Ok(prev)
             }
             Err(e) => {
@@ -546,7 +560,7 @@ impl State {
 
     fn apply_load_coherence(&mut self, threads: &mut thread::Set, index: usize) {
         for i in 0..self.stores.len() {
-            // Skip if the is current.
+            // Skip if it is the current load.
             if index == i {
                 continue;
             }
@@ -730,18 +744,9 @@ impl State {
         //
         // Add all stores **unless** a newer store has already been seen by the
         // current thread's causality.
-        'outer: for i in 0..self.stores.len() {
-            let store_i = &self.stores[i];
-
-            if i >= cnt {
-                // Not a real store
-                continue;
-            }
-
-            for j in 0..self.stores.len() {
-                let store_j = &self.stores[j];
-
-                if i == j || j >= cnt {
+        'outer: for (i, store_i) in self.stores.iter().take(cnt).enumerate() {
+            for (j, store_j) in self.stores.iter().take(cnt).enumerate() {
+                if i == j {
                     continue;
                 }
 
@@ -784,18 +789,9 @@ impl State {
 
         // Unlike `match_load_to_stores`, rmw operations only load "newest"
         // stores, in terms of modification order.
-        'outer: for i in 0..self.stores.len() {
-            let store_i = &self.stores[i];
-
-            if i >= cnt {
-                // Not a real store
-                continue;
-            }
-
-            for j in 0..self.stores.len() {
-                let store_j = &self.stores[j];
-
-                if i == j || j >= cnt {
+        'outer: for (i, store_i) in self.stores.iter().take(cnt).enumerate() {
+            for (j, store_j) in self.stores.iter().take(cnt).enumerate() {
+                if i == j {
                     continue;
                 }
 
diff --git a/tests/atomic.rs b/tests/atomic.rs
index 73a8636..3792f03 100644
--- a/tests/atomic.rs
+++ b/tests/atomic.rs
@@ -1,8 +1,9 @@
 #![deny(warnings, rust_2018_idioms)]
 
-use loom::sync::atomic::AtomicUsize;
+use loom::sync::atomic::{AtomicBool, AtomicUsize};
 use loom::thread;
 
+use std::collections::HashSet;
 use std::sync::atomic::Ordering::{AcqRel, Acquire, Relaxed, Release};
 use std::sync::Arc;
 
@@ -111,7 +112,10 @@ fn compare_and_swap_reads_old_values() {
 
 #[test]
 fn fetch_add_atomic() {
-    loom::model(|| {
+    let solutions = std::sync::Arc::new(std::sync::Mutex::new(HashSet::new()));
+    let solutions1 = std::sync::Arc::clone(&solutions);
+
+    loom::model(move || {
         let a1 = Arc::new(AtomicUsize::new(0));
         let a2 = a1.clone();
 
@@ -119,7 +123,142 @@ fn fetch_add_atomic() {
 
         let v1 = a1.fetch_add(1, Relaxed);
         let v2 = th.join().unwrap();
+        let v3 = a1.load(Relaxed);
+
+        solutions1.lock().unwrap().insert((v1, v2, v3));
+    });
+
+    let solutions = solutions.lock().unwrap();
+
+    assert!(solutions.contains(&(0, 1, 2)));
+    assert!(solutions.contains(&(1, 0, 2)));
+    assert_eq!(solutions.len(), 2);
+}
+
+#[test]
+fn store_does_not_squeeze_in_rmw() {
+    let solutions = std::sync::Arc::new(std::sync::Mutex::new(HashSet::new()));
+    let solutions1 = std::sync::Arc::clone(&solutions);
+
+    loom::model(move || {
+        let a1 = Arc::new(AtomicUsize::new(0));
+        let a2 = a1.clone();
+
+        let th = thread::spawn(move || {
+            a1.store(1, Relaxed);
+        });
+
+        let b1 = a2.swap(2, Relaxed);
+
+        a2.store(3, Relaxed);
+
+        let b2 = a2.swap(4, Relaxed);
+
+        th.join().unwrap();
+
+        let b3 = a2.load(Relaxed);
+
+        solutions1.lock().unwrap().insert((b1, b2, b3));
+    });
+
+    let solutions = solutions.lock().unwrap();
+
+    // store(1) before swap(2).
+    assert!(solutions.contains(&(1, 3, 4)));
+
+    // store(1) after swap(2), before store(3).
+    assert!(solutions.contains(&(0, 3, 4)));
+
+    // store(1) after store(3), before swap(4).
+    assert!(solutions.contains(&(0, 1, 4)));
+
+    // store(1) after swap(4) (but before join).
+    assert!(solutions.contains(&(0, 3, 1)));
+
+    assert_eq!(solutions.len(), 4);
+}
+
+#[test]
+fn store_oncurrent_failed_rmw() {
+    let solutions = std::sync::Arc::new(std::sync::Mutex::new(HashSet::new()));
+    let solutions1 = std::sync::Arc::clone(&solutions);
+
+    loom::model(move || {
+        let a1 = Arc::new(AtomicUsize::new(0));
+        let a2 = a1.clone();
+
+        let th = thread::spawn(move || {
+            a1.store(1, Relaxed);
+        });
+
+        let b1 = a2.compare_exchange(0, 2, Relaxed, Relaxed);
+
+        th.join().unwrap();
+
+        let b2 = a2.load(Relaxed);
+
+        solutions1.lock().unwrap().insert((b1, b2));
+    });
+
+    let solutions = solutions.lock().unwrap();
+
+    // store(1) before compare_exchange(0, 2).
+    assert!(solutions.contains(&(Err(1), 1)));
+
+    // store(1) after compare_exchange(0, 2).
+    assert!(solutions.contains(&(Ok(0), 1)));
+
+    assert_eq!(solutions.len(), 2, "{:?}", solutions);
+}
+
+#[test]
+fn unordered_stores() {
+    let solutions = std::sync::Arc::new(std::sync::Mutex::new(HashSet::new()));
+    let solutions1 = std::sync::Arc::clone(&solutions);
+
+    loom::model(move || {
+        let a1 = Arc::new(AtomicUsize::new(0));
+        let a2 = a1.clone();
+
+        let th = thread::spawn(move || {
+            a1.store(1, Relaxed);
+        });
+
+        a2.store(2, Relaxed);
+
+        th.join().unwrap();
+
+        let b = a2.load(Relaxed);
+
+        solutions1.lock().unwrap().insert(b);
+    });
+
+    let solutions = solutions.lock().unwrap();
+
+    assert!(solutions.contains(&1));
+    assert!(solutions.contains(&2));
+
+    assert_eq!(solutions.len(), 2);
+}
+
+// See issue https://github.com/tokio-rs/loom/issues/254
+#[test]
+fn concurrent_rmw_store() {
+    loom::model(move || {
+        let flag = Arc::new(AtomicBool::new(false));
+
+        let th = thread::spawn({
+            let flag = flag.clone();
+            move || flag.store(true, Relaxed) // a.k.a. unlock()
+        });
+
+        if flag.swap(false, Relaxed) {
+            // a.k.a. if !try_lock { return }
+            return;
+        }
+
+        th.join().unwrap();
 
-        assert_ne!(v1, v2);
+        assert!(flag.load(Relaxed)); // a.k.a. is_locked().
     });
 }