Race Condition Examples

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22

# VULNERABLE: Counter race in a web server request counter
 
class RequestMetrics:
    def __init__(self):
        self.request_count = 0
        self.total_response_time = 0.0
    
    def record_request(self, response_time_ms):
        # RACE: Read-modify-write on request_count
        self.request_count = self.request_count + 1
        
        # RACE: Read-modify-write on total_response_time
        self.total_response_time = self.total_response_time + response_time_ms
    
    def get_average_response_time(self):
        # RACE: Reading two related values non-atomically
        # count could increase between reading total and count
        if self.request_count == 0:
            return 0
        return self.total_response_time / self.request_count  # May be inconsistent!
 
# Multiple threads handling requests simultaneously will lose updates

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26

// VULNERABLE: Classic bank account race
 
public class BankAccount {
    private double balance;
    
    public BankAccount(double initialBalance) {
        this.balance = initialBalance;
    }
    
    public void withdraw(double amount) {
        // RACE: Check-then-act without atomicity
        if (balance >= amount) {
            // VULNERABILITY WINDOW: Another thread could withdraw here
            balance = balance - amount;  // May overdraw!
        }
    }
    
    public void transfer(BankAccount destination, double amount) {
        // RACE: Multiple operations that should be atomic
        if (this.balance >= amount) {
            this.balance -= amount;        // What if we fail after this?
            destination.balance += amount;  // And before this?
        }
        // Non-atomic transfer: money could disappear if interrupted!
    }
}

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37

// VULNERABLE: Classic broken double-checked locking
 
public class Singleton {
    private static Singleton instance;  // NOT volatile!
    
    public static Singleton getInstance() {
        if (instance == null) {                    // First check (no lock)
            synchronized (Singleton.class) {
                if (instance == null) {            // Second check (with lock)
                    instance = new Singleton();    // PROBLEM HERE
                }
            }
        }
        return instance;
    }
    
    private Singleton() {
        // Complex initialization...
    }
}
 
/*
 * WHY THIS FAILS:
 * 
 * The line "instance = new Singleton()" is NOT atomic. It involves:
 * 1. Allocate memory for Singleton
 * 2. Invoke constructor to initialize fields
 * 3. Assign reference to instance
 * 
 * The compiler/CPU can REORDER steps 2 and 3!
 * 
 * Thread A: Allocates memory, assigns to instance (step 3)
 * Thread A: ...about to run constructor (step 2)
 * Thread B: if (instance == null) -> FALSE (sees non-null reference)
 * Thread B: return instance  // Returns UNINITIALIZED object!
 * Thread B: Tries to use object, crashes or corrupts data
 */

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24

# VULNERABLE: Lazy cache population race
 
class ExpensiveComputationCache:
    def __init__(self):
        self._cache = {}
    
    def get_result(self, key):
        if key not in self._cache:
            # RACE: Multiple threads may compute simultaneously
            result = self._expensive_compute(key)  # Takes 5 seconds
            
            # RACE: Multiple threads may write different results
            # if computation is not deterministic
            self._cache[key] = result
        
        return self._cache[key]
    
    def _expensive_compute(self, key):
        # Expensive operation...
        return computed_result
 
# Problem 1: Multiple threads waste resources computing the same key
# Problem 2: In Python 3.5, dict operations weren't thread-safe
# Problem 3: If _expensive_compute has side effects, they happen multiple times

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36

// VULNERABLE: Concurrent ArrayList modification
 
List<String> list = new ArrayList<>();
 
// Thread 1: Adding elements
void addItems() {
    for (int i = 0; i < 10000; i++) {
        list.add("item-" + i);  // RACE: ArrayList is not thread-safe
    }
}
 
// Thread 2: Iterating
void processItems() {
    for (String item : list) {    // RACE: ConcurrentModificationException
        process(item);             // Or worse: silent corruption
    }
}
 
// Thread 3: Removing elements
void removeItems() {
    for (int i = 0; i < list.size(); i++) {  // RACE: size() changes during loop
        if (shouldRemove(list.get(i))) {
            list.remove(i);  // RACE: indices shift, may skip/double-process
            i--;             // Naive "fix" still races
        }
    }
}
 
/*
 * FAILURE MODES:
 * - ArrayIndexOutOfBoundsException
 * - ConcurrentModificationException
 * - Null elements where there shouldn't be
 * - Missing elements (never processed)
 * - Corrupted internal array (size != actual elements)
 */

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31

// VULNERABLE: HashMap corruption leading to infinite loop
 
Map<Integer, String> map = new HashMap<>();
 
// Multiple threads inserting simultaneously
void insertMany() {
    for (int i = 0; i < 100000; i++) {
        // RACE: HashMap.put() is NOT thread-safe
        map.put(i, "value-" + i);
    }
}
 
/*
 * WHAT CAN HAPPEN:
 * 
 * HashMap internally uses a linked list for hash collisions.
 * Concurrent insertions can corrupt the linked list into a CYCLE.
 * 
 * When get() or iteration traverses a cyclic list:
 * while (e != null) { e = e.next; }  // INFINITE LOOP
 * 
 * The thread hangs forever, consuming 100% CPU.
 * 
 * This is not hypothetical—it has crashed production systems,
 * including instances of the HotSpot JVM itself during startup.
 * 
 * Java HashMap javadoc explicitly warns:
 * "If multiple threads access a hash map concurrently, and at least 
 * one of the threads modifies the map structurally, it must be 
 * synchronized externally."
 */

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28

// VULNERABLE: Naive put-if-absent
 
Map<String, ExpensiveObject> cache = Collections.synchronizedMap(new HashMap<>());
 
public ExpensiveObject getOrCreate(String key) {
    // Each operation is atomic, but the COMBINATION is not!
    
    if (!cache.containsKey(key)) {     // Check
        // VULNERABILITY WINDOW
        // Another thread could insert the same key here!
        ExpensiveObject obj = new ExpensiveObject();  // Wasted work
        cache.put(key, obj);            // Act: may overwrite other thread's value
    }
    
    return cache.get(key);  // RACE: May return null if removed between put and get!
}
 
/*
 * PROBLEMS:
 * 1. Two threads may both see key missing, both create objects
 * 2. One thread's object gets overwritten and lost
 * 3. If ExpensiveObject has side effects, they happen twice
 * 4. Even worse: cache.get(key) could return null if another
 *    thread removed the key between put() and get()
 *
 * FIX: Use computeIfAbsent() which is atomic:
 *   cache.computeIfAbsent(key, k -> new ExpensiveObject())
 */

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27

# VULNERABLE: Directory creation race
 
import os
 
def ensure_directory(path):
    if not os.path.exists(path):     # Check
        # VULNERABILITY WINDOW
        # Another process could create the directory OR
        # create a SYMLINK to somewhere else!
        os.mkdir(path)               # Act: may fail or create in wrong location
    
    # Assume directory now exists and we control it
    write_to_directory(path)  # DANGEROUS if it's not what we expected
 
# PROBLEMS:
# 1. FileExistsError if another process creates first
# 2. Security: attacker creates symlink, we write to wrong location
# 3. Race between existence check and our use of the directory
 
# SAFER APPROACH:
def safe_ensure_directory(path):
    try:
        os.mkdir(path)  # Atomic: creates or fails
    except FileExistsError:
        pass  # Already exists, that's fine
    
    # But still should verify it's a directory we own!

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40

// VULNERABLE: Connection state machine race
 
public class Connection {
    private enum State { DISCONNECTED, CONNECTING, CONNECTED, DISCONNECTING }
    private State state = State.DISCONNECTED;
    
    public void connect() {
        if (state == State.DISCONNECTED) {
            state = State.CONNECTING;     // RACE: Two threads could both see DISCONNECTED
            
            performConnect();             // Network operation takes time
            
            state = State.CONNECTED;      // RACE: What if disconnect() called during connect?
        }
    }
    
    public void disconnect() {
        if (state == State.CONNECTED) {
            state = State.DISCONNECTING;  // RACE: connect() might be in progress
            
            performDisconnect();
            
            state = State.DISCONNECTED;   // RACE: Might overwrite CONNECTING state
        }
    }
    
    public void send(byte[] data) {
        if (state == State.CONNECTED) {   // RACE: State could change after this check
            doSend(data);                 // Send on disconnected socket!
        }
    }
}
 
/*
 * FAILURE MODES:
 * - Two simultaneous connects create two sockets (resource leak)
 * - Disconnect during connect leaves inconsistent state
 * - Send after disconnect check passes, socket already closed
 * - State ends up in impossible combination (CONNECTING + doDisconnect happening)
 */

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42

// VULNERABLE: Flag-based synchronization (broken)
 
// Thread 1: Producer
int data;
int ready = 0;
 
void producer() {
    data = 42;      // Write data
    ready = 1;      // Signal that data is ready
}
 
// Thread 2: Consumer
void consumer() {
    while (ready == 0) {
        // Spin, waiting for ready flag
    }
    printf("%d
", data);  // Expect 42
}
 
/*
 * WHAT CAN GO WRONG:
 *
 * 1. COMPILER REORDERING:
 *    Compiler may reorder "data = 42" after "ready = 1"
 *    (They're independent writes to the compiler)
 *
 * 2. CPU REORDERING:
 *    Store to 'data' may be buffered, while store to 'ready'
 *    propagates to consumer first
 *
 * 3. CONSUMER SEES:
 *    ready = 1, data = 0 (old value)
 *    Prints 0 instead of 42
 *
 * 4. WORSE ON WEAK MEMORY MODELS:
 *    On ARM, this can RELIABLY fail without memory barriers
 *
 * FIX: Use atomics with proper memory ordering:
 *    atomic_store_explicit(&ready, 1, memory_order_release);
 *    atomic_load_explicit(&ready, memory_order_acquire);
 */

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44

// VULNERABLE: Use-after-free race
 
struct Connection {
    int socket_fd;
    char buffer[1024];
    // ...
};
 
struct Connection* global_conn = NULL;
 
// Thread 1: Processing loop
void process_connection() {
    while (global_conn != NULL) {
        // Use the connection
        read(global_conn->socket_fd, global_conn->buffer, 1024);
        process_data(global_conn->buffer);
    }
}
 
// Thread 2: Cleanup on error
void cleanup_on_error() {
    if (global_conn != NULL) {
        struct Connection* temp = global_conn;
        global_conn = NULL;    // Signal to other thread
        
        // BUG: Thread 1 might be in middle of using global_conn!
        // The NULL check in Thread 1's while loop may have passed
        // BEFORE we set global_conn = NULL
        
        close(temp->socket_fd);  // Close while Thread 1 might be reading!
        free(temp);              // Free while Thread 1 holds a reference!
    }
}
 
/*
 * RACE TIMELINE:
 * T1: Check global_conn != NULL -> TRUE
 * T1: Enters loop body
 * T2: Sets global_conn = NULL
 * T2: Closes socket, frees memory
 * T1: read(global_conn->socket_fd, ...) // CRASH: accessing freed memory
 *
 * Result: Undefined behavior, crash, or security vulnerability
 */

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44

// VULNERABLE: Naive reference counting
 
class RefCounted {
    int refcount = 1;  // Non-atomic!
    
public:
    void addRef() {
        refcount++;    // RACE: Not atomic
    }
    
    void release() {
        refcount--;    // RACE: Two threads may both decrement from 2 to 1
        if (refcount == 0) {
            delete this;  // RACE: Might double-free
        }
    }
};
 
/*
 * RACE SCENARIO:
 * 
 * refcount = 2 (two threads hold references)
 * 
 * Thread A: reads refcount (2)
 * Thread B: reads refcount (2)
 * Thread A: computes 2 - 1 = 1
 * Thread B: computes 2 - 1 = 1
 * Thread A: writes refcount = 1
 * Thread B: writes refcount = 1
 * 
 * refcount = 1 (should be 0!)
 * Object never freed -> memory leak
 * 
 * WORSE SCENARIO:
 * 
 * Thread A: decrements refcount to 0
 * Thread B: about to decrement
 * Thread A: calls delete this
 * Thread B: decrements freed memory
 * Thread B: refcount appears to be -1 or anything
 * Thread B: may also call delete this -> double free!
 *
 * FIX: Use atomic operations for refcount
 */

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39

# VULNERABLE: Inventory oversell in distributed e-commerce
 
class InventoryService:
    def try_purchase(self, product_id, quantity):
        # Read current inventory from database
        current_stock = db.query(
            "SELECT quantity FROM inventory WHERE product_id = ?",
            product_id
        )
        
        if current_stock >= quantity:
            # RACE WINDOW: Many concurrent requests pass this check
            # All see the same stock level, all think they can proceed
            
            # Update inventory
            db.execute(
                "UPDATE inventory SET quantity = quantity - ? WHERE product_id = ?",
                quantity, product_id
            )
            
            # Create order
            create_order(product_id, quantity)
            return True
        
        return False
 
# SCENARIO:
# Product has 1 unit in stock
# 10 concurrent purchase requests arrive
# All 10 read current_stock = 1
# All 10 pass the if check
# All 10 decrement inventory (quantity goes to -9!)
# All 10 create orders
# Result: 10 orders fulfilled from 1 unit of inventory
 
# FIX: Use atomic conditional update:
#   UPDATE inventory SET quantity = quantity - ?
#   WHERE product_id = ? AND quantity >= ?
# Then check rows_affected to see if it succeeded

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31

# VULNERABLE: Distributed like counter race
 
class LikeService:
    def add_like(self, post_id, user_id):
        # Check if user already liked (prevent double-like)
        already_liked = redis.sismember(f"likes:{post_id}", user_id)
        
        if not already_liked:
            # RACE: Two requests from same user could both reach here
            
            # Add to set of users who liked
            redis.sadd(f"likes:{post_id}", user_id)
            
            # Increment like count
            redis.incr(f"like_count:{post_id}")
            
            # RACE RESULT: Like count and set member count may differ!
            # The set correctly has user once (idempotent)
            # But we may have incremented the counter twice
    
    def remove_like(self, post_id, user_id):
        # Check if user actually liked
        did_like = redis.sismember(f"likes:{post_id}", user_id)
        
        if did_like:
            # RACE: Unlike and Like could interleave badly
            redis.srem(f"likes:{post_id}", user_id)
            redis.decr(f"like_count:{post_id}")
 
# FIX: Use Lua script for atomic check-and-modify in Redis
# Or use transactions (MULTI/EXEC) when supported