Saturday, November 21, 2015

Starvation and Fairness

 If a thread is not granted CPU time because other threads grab it all, it is called "starvation". The thread is "starved to death" because other threads are allowed the CPU time instead of it. The solution to starvation is called "fairness" - that all threads are fairly granted a chance to execute.
Causes of Starvation in Java

The following three common causes can lead to starvation of threads in Java:

    Threads with high priority swallow all CPU time from threads with lower priority.

    Threads are blocked indefinately waiting to enter a synchronized block, because other threads are constantly allowed access before it.

    Threads waiting on an object (called wait() on it) remain waiting indefinitely because other threads are constantly awakened instead of it.


Threads with high priority swallow all CPU time from threads with lower priority

You can set the thread priority of each thread individually. The higher the priority the more CPU time the thread is granted. You can set the priority of threads between 1 and 10. Exactly how this is interpreted depends on the operating system your application is running on. For most applications you are better off leaving the priority unchanged.
Threads are blocked indefinitely waiting to enter a synchronized block

Java's synchronized code blocks can be another cause of starvation. Java's synchronized code block makes no guarantee about the sequence in which threads waiting to enter the synchronized block are allowed to enter. This means that there is a theoretical risk that a thread remains blocked forever trying to enter the block, because other threads are constantly granted access before it. This problem is called "starvation", that a thread is "starved to death" by because other threads are allowed the CPU time instead of it.
Threads waiting on an object (called wait() on it) remain waiting indefinitely

The notify() method makes no guarantee about what thread is awakened if multiple thread have called wait() on the object notify() is called on. It could be any of the threads waiting. Therefore there is a risk that a thread waiting on a certain object is never awakened because other waiting threads are always awakened instead of it.
Implementing Fairness in Java

While it is not possible to implement 100% fairness in Java we can still implement our synchronization constructs to increase fairness between threads.

First lets study a simple synchronized code block:

public class Synchronizer{

  public synchronized void doSynchronized(){
    //do a lot of work which takes a long time
  }

}

If more than one thread call the doSynchronized() method, some of them will be blocked until the first thread granted access has left the method. If more than one thread are blocked waiting for access there is no guarantee about which thread is granted access next.
Using Locks Instead of Synchronized Blocks

To increase the fairness of waiting threads first we will change the code block to be guarded by a lock rather than a synchronized block:

public class Synchronizer{
  Lock lock = new Lock();

  public void doSynchronized() throws InterruptedException{
    this.lock.lock();
      //critical section, do a lot of work which takes a long time
    this.lock.unlock();
  }

}

Notice how the doSynchronized() method is no longer declared synchronized. Instead the critical section is guarded by the lock.lock() and lock.unlock() calls.

A simple implementation of the Lock class could look like this:

public class Lock{
  private boolean isLocked      = false;
  private Thread  lockingThread = null;

  public synchronized void lock() throws InterruptedException{
    while(isLocked){
      wait();
    }
    isLocked      = true;
    lockingThread = Thread.currentThread();
  }

  public synchronized void unlock(){
    if(this.lockingThread != Thread.currentThread()){
      throw new IllegalMonitorStateException(
        "Calling thread has not locked this lock");
    }
    isLocked      = false;
    lockingThread = null;
    notify();
  }
}

If you look at the Synchronizer class above and look into this Lock implementation you will notice that threads are now blocked trying to access the lock() method, if more than one thread calls lock() simultanously. Second, if the lock is locked, the threads are blocked in the wait() call inside the while(isLocked) loop in the lock() method. Remember that a thread calling wait() releases the synchronization lock on the Lock instance, so threads waiting to enter lock() can now do so. The result is that multiple threads can end up having called wait() inside lock().

If you look back at the doSynchronized() method you will notice that the comment between lock() and unlock() states, that the code in between these two calls take a "long" time to execute. Let us further assume that this code takes long time to execute compared to entering the lock() method and calling wait() because the lock is locked. This means that the majority of the time waited to be able to lock the lock and enter the critical section is spent waiting in the wait() call inside the lock() method, not being blocked trying to enter the lock() method.

As stated earlier synchronized blocks makes no guarantees about what thread is being granted access if more than one thread is waiting to enter. Nor does wait() make any guarantees about what thread is awakened when notify() is called. So, the current version of the Lock class makes no different guarantees with respect to fairness than synchronized version of doSynchronized(). But we can change that.

The current version of the Lock class calls its own wait() method. If instead each thread calls wait() on a separate object, so that only one thread has called wait() on each object, the Lock class can decide which of these objects to call notify() on, thereby effectively selecting exactly what thread to awaken.
A Fair Lock

Below is shown the previous Lock class turned into a fair lock called FairLock. You will notice that the implementation has changed a bit with respect to synchronization and wait() / notify() compared to the Lock class shown earlier.

Exactly how I arrived at this design beginning from the previous Lock class is a longer story involving several incremental design steps, each fixing the problem of the previous step: Nested Monitor Lockout, Slipped Conditions, and Missed Signals. That discussion is left out of this text to keep the text short, but each of the steps are discussed in the appropriate texts on the topic ( see the links above). What is important is, that every thread calling lock() is now queued, and only the first thread in the queue is allowed to lock the FairLock instance, if it is unlocked. All other threads are parked waiting until they reach the top of the queue.

public class FairLock {
    private boolean           isLocked       = false;
    private Thread            lockingThread  = null;
    private List waitingThreads =
            new ArrayList();

  public void lock() throws InterruptedException{
    QueueObject queueObject           = new QueueObject();
    boolean     isLockedForThisThread = true;
    synchronized(this){
        waitingThreads.add(queueObject);
    }

    while(isLockedForThisThread){
      synchronized(this){
        isLockedForThisThread =
            isLocked || waitingThreads.get(0) != queueObject;
        if(!isLockedForThisThread){
          isLocked = true;
           waitingThreads.remove(queueObject);
           lockingThread = Thread.currentThread();
           return;
         }
      }
      try{
        queueObject.doWait();
      }catch(InterruptedException e){
        synchronized(this) { waitingThreads.remove(queueObject); }
        throw e;
      }
    }
  }

  public synchronized void unlock(){
    if(this.lockingThread != Thread.currentThread()){
      throw new IllegalMonitorStateException(
        "Calling thread has not locked this lock");
    }
    isLocked      = false;
    lockingThread = null;
    if(waitingThreads.size() > 0){
      waitingThreads.get(0).doNotify();
    }
  }
}

public class QueueObject {

  private boolean isNotified = false;

  public synchronized void doWait() throws InterruptedException {
    while(!isNotified){
        this.wait();
    }
    this.isNotified = false;
  }

  public synchronized void doNotify() {
    this.isNotified = true;
    this.notify();
  }

  public boolean equals(Object o) {
    return this == o;
  }
}

First you might notice that the lock() method is no longer declared synchronized. Instead only the blocks necessary to synchronize are nested inside synchronized blocks.

FairLock creates a new instance of QueueObject and enqueue it for each thread calling lock(). The thread calling unlock() will take the top QueueObject in the queue and call doNotify() on it, to awaken the thread waiting on that object. This way only one waiting thread is awakened at a time, rather than all waiting threads. This part is what governs the fairness of the FairLock.

Notice how the state of the lock is still tested and set within the same synchronized block to avoid slipped conditions.

Also notice that the QueueObject is really a semaphore. The doWait() and doNotify() methods store the signal internally in the QueueObject. This is done to avoid missed signals caused by a thread being preempted just before calling queueObject.doWait(), by another thread which calls unlock() and thereby queueObject.doNotify(). The queueObject.doWait() call is placed outside the synchronized(this) block to avoid nested monitor lockout, so another thread can actually call unlock() when no thread is executing inside the synchronized(this) block in lock() method.

Finally, notice how the queueObject.doWait() is called inside a try - catch block. In case an InterruptedException is thrown the thread leaves the lock() method, and we need to dequeue it.
A Note on Performance

If you compare the Lock and FairLock classes you will notice that there is somewhat more going on inside the lock() and unlock() in the FairLock class. This extra code will cause the FairLock to be a sligtly slower synchronization mechanism than Lock. How much impact this will have on your application depends on how long time the code in the critical section guarded by the FairLock takes to execute. The longer this takes to execute, the less significant the added overhead of the synchronizer is. It does of course also depend on how often this code is called.

Deadlock Prevention

I would recommend before going through this Blog first you need to go through Deadlock in Java to have a deep understanding of Deadlocks

In some situations it is possible to prevent deadlocks. I'll describe three techniques in this text:
  1. Lock Ordering
  2. Lock Timeout
  3. Deadlock Detection

Lock Ordering

Deadlock occurs when multiple threads need the same locks but obtain them in different order.
If you make sure that all locks are always taken in the same order by any thread, deadlocks cannot occur. Look at this example:
Thread 1:

  lock A 
  lock B


Thread 2:

   wait for A
   lock C (when A locked)


Thread 3:

   wait for A
   wait for B
   wait for C
If a thread, like Thread 3, needs several locks, it must take them in the decided order. It cannot take a lock later in the sequence until it has obtained the earlier locks.
For instance, neither Thread 2 or Thread 3 can lock C until they have locked A first. Since Thread 1 holds lock A, Thread 2 and 3 must first wait until lock A is unlocked. Then they must succeed in locking A, before they can attempt to lock B or C.
Lock ordering is a simple yet effective deadlock prevention mechanism. However, it can only be used if you know about all locks needed ahead of taking any of the locks. This is not always the case.

Lock Timeout

Another deadlock prevention mechanism is to put a timeout on lock attempts meaning a thread trying to obtain a lock will only try for so long before giving up. If a thread does not succeed in taking all necessary locks within the given timeout, it will backup, free all locks taken, wait for a random amount of time and then retry. The random amount of time waited serves to give other threads trying to take the same locks a chance to take all locks, and thus let the application continue running without locking.
Here is an example of two threads trying to take the same two locks in different order, where the threads back up and retry:
Thread 1 locks A
Thread 2 locks B

Thread 1 attempts to lock B but is blocked
Thread 2 attempts to lock A but is blocked

Thread 1's lock attempt on B times out
Thread 1 backs up and releases A as well
Thread 1 waits randomly (e.g. 257 millis) before retrying.

Thread 2's lock attempt on A times out
Thread 2 backs up and releases B as well
Thread 2 waits randomly (e.g. 43 millis) before retrying.
In the above example Thread 2 will retry taking the locks about 200 millis before Thread 1 and will therefore likely succeed at taking both locks. Thread 1 will then wait already trying to take lock A. When Thread 2 finishes, Thread 1 will be able to take both locks too (unless Thread 2 or another thread takes the locks in between).
An issue to keep in mind is, that just because a lock times out it does not necessarily mean that the threads had deadlocked. It could also just mean that the thread holding the lock (causing the other thread to time out) takes a long time to complete its task.
Additionally, if enough threads compete for the same resources they still risk trying to take the threads at the same time again and again, even if timing out and backing up. This may not occur with 2 threads each waiting between 0 and 500 millis before retrying, but with 10 or 20 threads the situation is different. Then the likeliness of two threads waiting the same time before retrying (or close enough to cause problems) is a lot higher.
A problem with the lock timeout mechanism is that it is not possible to set a timeout for entering a synchronized block in Java. You will have to create a custom lock class or use one of the Java 5 concurrency constructs in the java.util.concurrency package. Writing custom locks isn't difficult but it is outside the scope of this text. Later texts in the Java concurrency trails will cover custom locks.

Deadlock Detection

Deadlock detection is a heavier deadlock prevention mechanism aimed at cases in which lock ordering isn't possible, and lock timeout isn't feasible.
Every time a thread takes a lock it is noted in a data structure (map, graph etc.) of threads and locks. Additionally, whenever a thread requests a lock this is also noted in this data structure.
When a thread requests a lock but the request is denied, the thread can traverse the lock graph to check for deadlocks. For instance, if a Thread A requests lock 7, but lock 7 is held by Thread B, then Thread A can check if Thread B has requested any of the locks Thread A holds (if any). If Thread B has requested so, a deadlock has occurred (Thread A having taken lock 1, requesting lock 7, Thread B having taken lock 7, requesting lock 1).
Of course a deadlock scenario may be a lot more complicated than two threads holding each others locks. Thread A may wait for Thread B, Thread B waits for Thread C, Thread C waits for Thread D, and Thread D waits for Thread A. In order for Thread A to detect a deadlock it must transitively examine all requested locks by Thread B. From Thread B's requested locks Thread A will get to Thread C, and then to Thread D, from which it finds one of the locks Thread A itself is holding. Then it knows a deadlock has occurred.
Below is a graph of locks taken and requested by 4 threads (A, B, C and D). A data structure like this that can be used to detect deadlocks.



So what do the threads do if a deadlock is detected?
One possible action is to release all locks, backup, wait a random amount of time and then retry. This is similar to the simpler lock timeout mechanism except threads only backup when a deadlock has actually occurred. Not just because their lock requests timed out. However, if a lot of threads are competing for the same locks they may repeatedly end up in a deadlock even if they back up and wait.

A better option is to determine or assign a priority of the threads so that only one (or a few) thread backs up. The rest of the threads continue taking the locks they need as if no deadlock had occurred. If the priority assigned to the threads is fixed, the same threads will always be given higher priority. To avoid this you may assign the priority randomly whenever a deadlock is detected.

Deadlock in Java

Thread Deadlock

A deadlock is when two or more threads are blocked waiting to obtain locks that some of the other threads in the deadlock are holding. Deadlock can occur when multiple threads need the same locks, at the same time, but obtain them in different order.

For instance, if thread 1 locks A, and tries to lock B, and thread 2 has already locked B, and tries to lock A, a deadlock arises. Thread 1 can never get B, and thread 2 can never get A. In addition, neither of them will ever know. They will remain blocked on each their object, A and B, forever. This situation is a deadlock.

The situation is illustrated below:

Thread 1  locks A, waits for B
Thread 2  locks B, waits for A

Here is an example of a TreeNode class that call synchronized methods in different instances:

public class TreeNode {

  TreeNode parent   = null; 
  List     children = new ArrayList();

  public synchronized void addChild(TreeNode child){
    if(!this.children.contains(child)) {
      this.children.add(child);
      child.setParentOnly(this);
    }
  }
 
  public synchronized void addChildOnly(TreeNode child){
    if(!this.children.contains(child){
      this.children.add(child);
    }
  }
 
  public synchronized void setParent(TreeNode parent){
    this.parent = parent;
    parent.addChildOnly(this);
  }

  public synchronized void setParentOnly(TreeNode parent){
    this.parent = parent;
  }
}

If a thread (1) calls the parent.addChild(child) method at the same time as another thread (2) calls the child.setParent(parent) method, on the same parent and child instances, a deadlock can occur. Here is some pseudo code that illustrates this:

Thread 1: parent.addChild(child); //locks parent
          --> child.setParentOnly(parent);

Thread 2: child.setParent(parent); //locks child
          --> parent.addChildOnly()

First thread 1 calls parent.addChild(child). Since addChild() is synchronized thread 1 effectively locks the parent object for access from other treads.

Then thread 2 calls child.setParent(parent). Since setParent() is synchronized thread 2 effectively locks the child object for acces from other threads.

Now both child and parent objects are locked by two different threads. Next thread 1 tries to call child.setParentOnly() method, but the child object is locked by thread 2, so the method call just blocks. Thread 2 also tries to call parent.addChildOnly() but the parent object is locked by thread 1, causing thread 2 to block on that method call. Now both threads are blocked waiting to obtain locks the other thread holds.

Note: The two threads must call parent.addChild(child) and child.setParent(parent) at the same time as described above, and on the same two parent and child instances for a deadlock to occur. The code above may execute fine for a long time until all of a sudden it deadlocks.

The threads really need to take the locks *at the same time*. For instance, if thread 1 is a bit ahead of thread2, and thus locks both A and B, then thread 2 will be blocked already when trying to lock B. Then no deadlock occurs. Since thread scheduling often is unpredictable there is no way to predict *when* a deadlock occurs. Only that it *can* occur.

More Complicated Deadlocks

Deadlock can also include more than two threads. This makes it harder to detect. Here is an example in which four threads have deadlocked:

Thread 1  locks A, waits for B
Thread 2  locks B, waits for C
Thread 3  locks C, waits for D
Thread 4  locks D, waits for A

Thread 1 waits for thread 2, thread 2 waits for thread 3, thread 3 waits for thread 4, and thread 4 waits for thread 1.
Database Deadlocks

A more complicated situation in which deadlocks can occur, is a database transaction. A database transaction may consist of many SQL update requests. When a record is updated during a transaction, that record is locked for updates from other transactions, until the first transaction completes. Each update request within the same transaction may therefore lock some records in the database.

If multiple transactions are running at the same time that need to update the same records, there is a risk of them ending up in a deadlock.

For example

Transaction 1, request 1, locks record 1 for update
Transaction 2, request 1, locks record 2 for update
Transaction 1, request 2, tries to lock record 2 for update.
Transaction 2, request 2, tries to lock record 1 for update.

Since the locks are taken in different requests, and not all locks needed for a given transaction are known ahead of time, it is hard to detect or prevent deadlocks in database transactions.


For How to Prevent Deadlock in Java Go through  Deadlock Prevention

Thread Pools in Java

 Thread Pools are useful when you need to limit the number of threads running in your application at the same time. There is a performance overhead associated with starting a new thread, and each thread is also allocated some memory for its stack etc.

Instead of starting a new thread for every task to execute concurrently, the task can be passed to a thread pool. As soon as the pool has any idle threads the task is assigned to one of them and executed. Internally the tasks are inserted into a Blocking Queue which the threads in the pool are dequeuing from. When a new task is inserted into the queue one of the idle threads will dequeue it successfully and execute it. The rest of the idle threads in the pool will be blocked waiting to dequeue tasks.

Thread pools are often used in multi threaded servers. Each connection arriving at the server via the network is wrapped as a task and passed on to a thread pool. The threads in the thread pool will process the requests on the connections concurrently. A later trail will get into detail about implementing multithreaded servers in Java.

Java 5 comes with built in thread pools in the java.util.concurrent package, so you don't have to implement your own thread pool. You can read more about it in my text on the java.util.concurrent.ExecutorService. Still it can be useful to know a bit about the implementation of a thread pool anyways.

Here is a simple thread pool implementation. Please note that this implementation uses my own BlockingQueue class as explained in my Blocking Queues tutorial. In a real life implementation you would probably use one of Java's built-in blocking queues instead.

public class ThreadPool {

    private BlockingQueue taskQueue = null;
    private List threads = new ArrayList();
    private boolean isStopped = false;

    public ThreadPool(int noOfThreads, int maxNoOfTasks){
        taskQueue = new BlockingQueue(maxNoOfTasks);

        for(int i=0; i            threads.add(new PoolThread(taskQueue));
        }
        for(PoolThread thread : threads){
            thread.start();
        }
    }

    public synchronized void  execute(Runnable task) throws Exception{
        if(this.isStopped) throw
            new IllegalStateException("ThreadPool is stopped");

        this.taskQueue.enqueue(task);
    }

    public synchronized void stop(){
        this.isStopped = true;
        for(PoolThread thread : threads){
           thread.doStop();
        }
    }

}

public class PoolThread extends Thread {

    private BlockingQueue taskQueue = null;
    private boolean       isStopped = false;

    public PoolThread(BlockingQueue queue){
        taskQueue = queue;
    }

    public void run(){
        while(!isStopped()){
            try{
                Runnable runnable = (Runnable) taskQueue.dequeue();
                runnable.run();
            } catch(Exception e){
                //log or otherwise report exception,
                //but keep pool thread alive.
            }
        }
    }

    public synchronized void doStop(){
        isStopped = true;
        this.interrupt(); //break pool thread out of dequeue() call.
    }

    public synchronized boolean isStopped(){
        return isStopped;
    }
}

The thread pool implementation consists of two parts. A ThreadPool class which is the public interface to the thread pool, and a PoolThread class which implements the threads that execute the tasks.

To execute a task the method ThreadPool.execute(Runnable r) is called with a Runnable implementation as parameter. The Runnable is enqueued in the blocking queue internally, waiting to be dequeued.

The Runnable will be dequeued by an idle PoolThread and executed. You can see this in the PoolThread.run() method. After execution the PoolThread loops and tries to dequeue a task again, until stopped.

To stop the ThreadPool the method ThreadPool.stop() is called. The stop called is noted internally in the isStopped member. Then each thread in the pool is stopped by calling doStop() on each thread. Notice how the execute() method will throw an IllegalStateException if execute() is called after stop() has been called.

The threads will stop after finishing any task they are currently executing. Notice the this.interrupt() call in PoolThread.doStop(). This makes sure that a thread blocked in a wait() call inside the taskQueue.dequeue() call breaks out of the wait() call, and leaves the dequeue() method call with an InterruptedException thrown. This exception is caught in the PoolThread.run() method, reported, and then the isStopped variable is checked. Since isStopped is now true, the PoolThread.run() will exit and the thread dies. 


For Difference between cachedThreadPool and fixedThreadPool refer CachedThreadPool vs FixedThreadPool

Read / Write Locks in Java

 A read / write lock is more sophisticated lock than the Lock implementations shown in the text Locks in Java. Imagine you have an application that reads and writes some resource, but writing it is not done as much as reading it is. Two threads reading the same resource does not cause problems for each other, so multiple threads that want to read the resource are granted access at the same time, overlapping. But, if a single thread wants to write to the resource, no other reads nor writes must be in progress at the same time. To solve this problem of allowing multiple readers but only one writer, you will need a read / write lock.

Java 5 comes with read / write lock implementations in the java.util.concurrent package. Even so, it may still be useful to know the theory behind their implementation.
Read / Write Lock Java Implementation

First let's summarize the conditions for getting read and write access to the resource:
Read Access       If no threads are writing, and no threads have requested write access.
Write Access       If no threads are reading or writing.

If a thread wants to read the resource, it is okay as long as no threads are writing to it, and no threads have requested write access to the resource. By up-prioritizing write-access requests we assume that write requests are more important than read-requests. Besides, if reads are what happens most often, and we did not up-prioritize writes, starvation could occur. Threads requesting write access would be blocked until all readers had unlocked the ReadWriteLock. If new threads were constantly granted read access the thread waiting for write access would remain blocked indefinately, resulting in starvation. Therefore a thread can only be granted read access if no thread has currently locked the ReadWriteLock for writing, or requested it locked for writing.

A thread that wants write access to the resource can be granted so when no threads are reading nor writing to the resource. It doesn't matter how many threads have requested write access or in what sequence, unless you want to guarantee fairness between threads requesting write access.

With these simple rules in mind we can implement a ReadWriteLock as shown below:

public class ReadWriteLock{

  private int readers       = 0;
  private int writers       = 0;
  private int writeRequests = 0;

  public synchronized void lockRead() throws InterruptedException{
    while(writers > 0 || writeRequests > 0){
      wait();
    }
    readers++;
  }

  public synchronized void unlockRead(){
    readers--;
    notifyAll();
  }

  public synchronized void lockWrite() throws InterruptedException{
    writeRequests++;

    while(readers > 0 || writers > 0){
      wait();
    }
    writeRequests--;
    writers++;
  }

  public synchronized void unlockWrite() throws InterruptedException{
    writers--;
    notifyAll();
  }
}

The ReadWriteLock has two lock methods and two unlock methods. One lock and unlock method for read access and one lock and unlock for write access.

The rules for read access are implemented in the lockRead() method. All threads get read access unless there is a thread with write access, or one or more threads have requested write access.

The rules for write access are implemented in the lockWrite() method. A thread that wants write access starts out by requesting write access (writeRequests++). Then it will check if it can actually get write access. A thread can get write access if there are no threads with read access to the resource, and no threads with write access to the resource. How many threads have requested write access doesn't matter.

It is worth noting that both unlockRead() and unlockWrite() calls notifyAll() rather than notify(). To explain why that is, imagine the following situation:

Inside the ReadWriteLock there are threads waiting for read access, and threads waiting for write access. If a thread awakened by notify() was a read access thread, it would be put back to waiting because there are threads waiting for write access. However, none of the threads awaiting write access are awakened, so nothing more happens. No threads gain neither read nor write access. By calling noftifyAll() all waiting threads are awakened and check if they can get the desired access.

Calling notifyAll() also has another advantage. If multiple threads are waiting for read access and none for write access, and unlockWrite() is called, all threads waiting for read access are granted read access at once - not one by one.
Read / Write Lock Reentrance

The ReadWriteLock class shown earlier is not reentrant. If a thread that has write access requests it again, it will block because there is already one writer - itself. Furthermore, consider this case:

    Thread 1 gets read access.

    Thread 2 requests write access but is blocked because there is one reader.

    Thread 1 re-requests read access (re-enters the lock), but is blocked because there is a write request

In this situation the previous ReadWriteLock would lock up - a situation similar to deadlock. No threads requesting neither read nor write access would be granted so.

To make the ReadWriteLock reentrant it is necessary to make a few changes. Reentrance for readers and writers will be dealt with separately.
Read Reentrance

To make the ReadWriteLock reentrant for readers we will first establish the rules for read reentrance:

    A thread is granted read reentrance if it can get read access (no writers or write requests), or if it already has read access (regardless of write requests).

To determine if a thread has read access already a reference to each thread granted read access is kept in a Map along with how many times it has acquired read lock. When determing if read access can be granted this Map will be checked for a reference to the calling thread. Here is how the lockRead() and unlockRead() methods looks after that change:

public class ReadWriteLock{

  private Map readingThreads =
      new HashMap();

  private int writers        = 0;
  private int writeRequests  = 0;

  public synchronized void lockRead() throws InterruptedException{
    Thread callingThread = Thread.currentThread();
    while(! canGrantReadAccess(callingThread)){
      wait();                                                                  
    }

    readingThreads.put(callingThread,
       (getAccessCount(callingThread) + 1));
  }


  public synchronized void unlockRead(){
    Thread callingThread = Thread.currentThread();
    int accessCount = getAccessCount(callingThread);
    if(accessCount == 1){ readingThreads.remove(callingThread); }
    else { readingThreads.put(callingThread, (accessCount -1)); }
    notifyAll();
  }


  private boolean canGrantReadAccess(Thread callingThread){
    if(writers > 0)            return false;
    if(isReader(callingThread) return true;
    if(writeRequests > 0)      return false;
    return true;
  }

  private int getReadAccessCount(Thread callingThread){
    Integer accessCount = readingThreads.get(callingThread);
    if(accessCount == null) return 0;
    return accessCount.intValue();
  }

  private boolean isReader(Thread callingThread){
    return readingThreads.get(callingThread) != null;
  }

}

As you can see read reentrance is only granted if no threads are currently writing to the resource. Additionally, if the calling thread already has read access this takes precedence over any writeRequests.
Write Reentrance

Write reentrance is granted only if the thread has already write access. Here is how the lockWrite() and unlockWrite() methods look after that change:

public class ReadWriteLock{

    private Map readingThreads =
        new HashMap();

    private int writeAccesses    = 0;
    private int writeRequests    = 0;
    private Thread writingThread = null;

  public synchronized void lockWrite() throws InterruptedException{
    writeRequests++;
    Thread callingThread = Thread.currentThread();
    while(! canGrantWriteAccess(callingThread)){
      wait();
    }
    writeRequests--;
    writeAccesses++;
    writingThread = callingThread;
  }

  public synchronized void unlockWrite() throws InterruptedException{
    writeAccesses--;
    if(writeAccesses == 0){
      writingThread = null;
    }
    notifyAll();
  }

  private boolean canGrantWriteAccess(Thread callingThread){
    if(hasReaders())             return false;
    if(writingThread == null)    return true;
    if(!isWriter(callingThread)) return false;
    return true;
  }

  private boolean hasReaders(){
    return readingThreads.size() > 0;
  }

  private boolean isWriter(Thread callingThread){
    return writingThread == callingThread;
  }
}

Notice how the thread currently holding the write lock is now taken into account when determining if the calling thread can get write access.
Read to Write Reentrance

Sometimes it is necessary for a thread that have read access to also obtain write access. For this to be allowed the thread must be the only reader. To achieve this the writeLock() method should be changed a bit. Here is what it would look like:

public class ReadWriteLock{

    private Map readingThreads =
        new HashMap();

    private int writeAccesses    = 0;
    private int writeRequests    = 0;
    private Thread writingThread = null;

  public synchronized void lockWrite() throws InterruptedException{
    writeRequests++;
    Thread callingThread = Thread.currentThread();
    while(! canGrantWriteAccess(callingThread)){
      wait();
    }
    writeRequests--;
    writeAccesses++;
    writingThread = callingThread;
  }

  public synchronized void unlockWrite() throws InterruptedException{
    writeAccesses--;
    if(writeAccesses == 0){
      writingThread = null;
    }
    notifyAll();
  }

  private boolean canGrantWriteAccess(Thread callingThread){
    if(isOnlyReader(callingThread))    return true;
    if(hasReaders())                   return false;
    if(writingThread == null)          return true;
    if(!isWriter(callingThread))       return false;
    return true;
  }

  private boolean hasReaders(){
    return readingThreads.size() > 0;
  }

  private boolean isWriter(Thread callingThread){
    return writingThread == callingThread;
  }

  private boolean isOnlyReader(Thread thread){
      return readers == 1 && readingThreads.get(callingThread) != null;
      }
 
}

Now the ReadWriteLock class is read-to-write access reentrant.
Write to Read Reentrance

Sometimes a thread that has write access needs read access too. A writer should always be granted read access if requested. If a thread has write access no other threads can have read nor write access, so it is not dangerous. Here is how the canGrantReadAccess() method will look with that change:

public class ReadWriteLock{

    private boolean canGrantReadAccess(Thread callingThread){
      if(isWriter(callingThread)) return true;
      if(writingThread != null)   return false;
      if(isReader(callingThread)  return true;
      if(writeRequests > 0)       return false;
      return true;
    }

}

Fully Reentrant ReadWriteLock

Below is the fully reentran ReadWriteLock implementation. I have made a few refactorings to the access conditions to make them easier to read, and thereby easier to convince yourself that they are correct.

public class ReadWriteLock{

  private Map readingThreads =
       new HashMap();

   private int writeAccesses    = 0;
   private int writeRequests    = 0;
   private Thread writingThread = null;


  public synchronized void lockRead() throws InterruptedException{
    Thread callingThread = Thread.currentThread();
    while(! canGrantReadAccess(callingThread)){
      wait();
    }

    readingThreads.put(callingThread,
     (getReadAccessCount(callingThread) + 1));
  }

  private boolean canGrantReadAccess(Thread callingThread){
    if( isWriter(callingThread) ) return true;
    if( hasWriter()             ) return false;
    if( isReader(callingThread) ) return true;
    if( hasWriteRequests()      ) return false;
    return true;
  }


  public synchronized void unlockRead(){
    Thread callingThread = Thread.currentThread();
    if(!isReader(callingThread)){
      throw new IllegalMonitorStateException("Calling Thread does not" +
        " hold a read lock on this ReadWriteLock");
    }
    int accessCount = getReadAccessCount(callingThread);
    if(accessCount == 1){ readingThreads.remove(callingThread); }
    else { readingThreads.put(callingThread, (accessCount -1)); }
    notifyAll();
  }

  public synchronized void lockWrite() throws InterruptedException{
    writeRequests++;
    Thread callingThread = Thread.currentThread();
    while(! canGrantWriteAccess(callingThread)){
      wait();
    }
    writeRequests--;
    writeAccesses++;
    writingThread = callingThread;
  }

  public synchronized void unlockWrite() throws InterruptedException{
    if(!isWriter(Thread.currentThread()){
      throw new IllegalMonitorStateException("Calling Thread does not" +
        " hold the write lock on this ReadWriteLock");
    }
    writeAccesses--;
    if(writeAccesses == 0){
      writingThread = null;
    }
    notifyAll();
  }

  private boolean canGrantWriteAccess(Thread callingThread){
    if(isOnlyReader(callingThread))    return true;
    if(hasReaders())                   return false;
    if(writingThread == null)          return true;
    if(!isWriter(callingThread))       return false;
    return true;
  }


  private int getReadAccessCount(Thread callingThread){
    Integer accessCount = readingThreads.get(callingThread);
    if(accessCount == null) return 0;
    return accessCount.intValue();
  }


  private boolean hasReaders(){
    return readingThreads.size() > 0;
  }

  private boolean isReader(Thread callingThread){
    return readingThreads.get(callingThread) != null;
  }

  private boolean isOnlyReader(Thread callingThread){
    return readingThreads.size() == 1 &&
           readingThreads.get(callingThread) != null;
  }

  private boolean hasWriter(){
    return writingThread != null;
  }

  private boolean isWriter(Thread callingThread){
    return writingThread == callingThread;
  }

  private boolean hasWriteRequests(){
      return this.writeRequests > 0;
  }

}

Calling unlock() From a finally-clause

When guarding a critical section with a ReadWriteLock, and the critical section may throw exceptions, it is important to call the readUnlock() and writeUnlock() methods from inside a finally-clause. Doing so makes sure that the ReadWriteLock is unlocked so other threads can lock it. Here is an example:

lock.lockWrite();
try{
  //do critical section code, which may throw exception
} finally {
  lock.unlockWrite();
}

This little construct makes sure that the ReadWriteLock is unlocked in case an exception is thrown from the code in the critical section. If unlockWrite() was not called from inside a finally-clause, and an exception was thrown from the critical section, the ReadWriteLock would remain write locked forever, causing all threads calling lockRead() or lockWrite() on that ReadWriteLock instance to halt indefinately. The only thing that could unlock the ReadWriteLockagain would be if the ReadWriteLock is reentrant, and the thread that had it locked when the exception was thrown, later succeeds in locking it, executing the critical section and calling unlockWrite() again afterwards. That would unlock the ReadWriteLock again. But why wait for that to happen, if it happens? Calling unlockWrite() from a finally-clause is a much more robust solution.