Chapter Seven -- Synchronization Examples -- Lecture Notes

• 7.0 Objectives
  
  
  • Explain the bounded-buffer, readers-writers, and dining-philosophers synchronization problems.
  • Describe specific tools used by Linux and Windows to solve process synchronization problems.
  • Illustrate how POSIX and Java can be used to solve process synchronization problems.
  • Design and develop solutions to process synchronization problems using POSIX and Java APIs. (We won't do any actual design work)
  
  
• 7.1 Classic Problems of Synchronization
  
  The authors present "classic" problems that are often used to test new synchronization tools.
  
  
  • 7.1.1 The Bounded Buffer Problem
    
    We can solve the bounded buffer problem by encapsulating the functionality of the counter in semaphores, as illustrated by the code below.

    
    ---------------------
        /*  All these variables and types are shared */
    #define BUFFER_SIZE 10
    typedef struct 
    { 
       /* declare desired fields for the buffer item type */
    } item ;
    
    item buffer[BUFFER_SIZE] ;   // the buffer implemented as an array of items
    
    int in=0,                  /* in == where the next item is added */
    out=0;                     /* out == where the next item is removed */
    semaphore full(0),         /* full->value == the number of full buffers */
    empty(BUFFER_SIZE) ;       /* empty->value == the number of empty buffers */
    ---------------------
    Producer's Code
    do
    {
          /* produce an item in nextp */
       wait (empty) ;
       buffer[in]= nextp ;
       in = (in + 1) % BUFFER_SIZE ;
       signal(full) ;
    } while (true) ;
    ---------------------
    Consumer's Code
    do
    {
       wait (full) ;
       nextc = buffer[out] ;
       out = (out + 1) % BUFFER_SIZE ;
       signal(empty) ;
           /* consume nextc */
    } while (true) ;
    ---------------------
    

  • 7.1.2 The Readers-Writers Problem
    
    

    

    
    
    Several processes share access to a file. Some of the processes (readers) never do anything but read the file. Others (writers) may perform writes on the file when they access it. The problem is to synchronize the processes so that readers can share the file concurrently but writers get exclusive access. There are different versions of the problem, for example:
    
    

    

    

    • The first readers-writers problem: no reader will be kept waiting unless a writer has already obtained permission to access the file.
    • The second readers-writers problem: once a writer is ready, that writer performs its access as soon as possible.
    
    A solution to either of the two problems above may result in starvation. (In the example code provided in the text, writers can starve.)
    
    A compromise might be a protocol that:
    • allows all processes to be served in FIFO order, and
    • allows groups of readers between writers to access the file concurrently.
    
    Many operating systems now make reader-writer locks available to system programmers. Processes that want only to read from a file can acquire a (shared) read-lock. Processes that want to write request an exclusive writer's lock.
    
    
  • 7.1.3 The Dining-Philosophers Problem
    
    

    

    One way to understand this problem is to think of five processes and five disk drives arranged in a circle, with the drives and the processes alternating. Each process occasionally needs to get exclusive access to the disk drives on its 'left' and 'right' in order to transfer some data from one to the other. The problem is to synchronize the processes so they get the access they need without experiencing indefinite postponement. The protocol has to resolve competition between pairs of processes that have to share the disk drive between them.
    
    
    • 7.1.3.1 Semaphore Solution
      
      

      

      The example protocol proposed in this section allows deadlock to occur, an extreme form of starvation.
      
      There are several ways to solve the problem. Resource ordering provides a very simple solution. A satisfactory solution to the problem would
      
      
      1. make sure only one process has possession of any particular disk drive at any one time.
      2. allow only processes that are vying for disk drive D or are in possession of D to participate in the decision as to who gets possession of D next, and make sure this decision is not postponed indefinitely.
      3. make sure there exists a bound on the number of times that other processes are allowed to make a data transfer after one process P has made a request to begin a transfer and before P is allowed to perform the transfer.
         
         
  • 7.1.3.2 Monitor Solution
    
    
    • The text shows some code that employs a monitor and some client code to create a partial solution to the dining philosophers problem. (Mutual exclusion is assured, and deadlock is impossible, but starvation can occur.
      
      
    • In order to eat, a philosopher P has to wait for this event: E = [both neighbors of P are not eating]. There is no built-in, up-front limit to how many times a philosopher different from P will eat before E happens. In other words, P can be postponed indefinitely.
  
  
• 7.2 Synchronization within the Kernel
  
  
  • We look at examples of synchronization tools used in Windows and in Linux.
    
    
  • 7.2.1 Synchronization in Windows
    
    
    • The Windows kernel is multithreaded. It supports real-time applications and multiple processors.
      
      
    • On a uni-processor it relies to some extent on masking interrupts to insure exclusive access to global shared data.
      
      
    • On a multiprocessor, it uses spinlocks to guard short sections of code accessing global resources. The OS ensures that a thread holding a spinlock will not be preempted. (This helps allow the thread to quickly finish using the resource, and release it.)
      
      
    • Threads outside the kernel can use dispatcher objects to synchronize.
      • synchronization can take the form of mutex, semaphore, event, or timer.
      • Events are used much like conditions -- to wait for a desired condition to occur.
      • Timers are to notify thread(s) that a specific amount of time has expired.
      
      
    • There's a Windows synchronization tool available in user-mode called a critical-section object. If waiting is required, it starts off busy-waiting like a spinlock, but after waiting "too long" it blocks. This method avoids the overhead of context switching when waits are short, and also avoids excessive waste of CPU cycles.
    
    
  • 7.2.2 Synchronization in Linux
    
    
    • The Linux kernel was nonpreemptive through much of its development (like many versions of unix). That means a kernel process could not be preempted. Since version 2.6, the Linux kernel is fully preemptive.
      
      
    • Linux supports atomic integers.
    • Linux supports mutex locks that have a blocking wait operation.
    • Linux has spinlocks for waits of short duration.
    • Linux has semaphores for longer waits.
    • Reader-writer versions of both spinlocks and semaphores are available.
      
      
    • On uni-processors, Linux enforces exclusive access by temporarily disabling preemption of the kernel.
  
  
• 7.3 POSIX Synchronization
  
  Almost all of the synchronization tools discussed in the previous section are only available for use by kernel code. The POSIX API discussed in this section provides tools for user-level application code. (However, bear in mind that POSIX tools must be implemented using kernel-level synchronization tools.)
  
  
  • The Pthreads API offers mutex locks, condition variables, and read-write locks
    
    
  • Semaphores and spinlocks are often available as extensions to the standard package.
  
  
  • 7.3.1 POSIX Mutex Locks
    
    The POSIX API offers mutex locks that have a blocking wait operation. They are used to solve critical section problems in this manner:

              #include <pthread.h>
              pthread_mutex_t mutex ;
              pthread_mutex_init (&mutex, NULL) ;
              ...
                /* other stuff in the program */
              ...
              pthread_mutex_lock (&mutex) ;
                /* critical section */
              pthread_mutex_unlock (&mutex) ;

  • 7.3.2 POSIX Semaphores
    
    Semaphores and spinlocks are often available as extensions to the standard package. There are named and unnamed semaphores.
    
    
    • 7.3.2.1 POSIX Named Semaphores
      
      Named semaphores are given a name when created, and it is easy for multiple processes to share a named semaphore, because the system makes the semaphore available to any process that 'knows' the name.
      
      MacOS and Linux have named POSIX semaphores.
      
      
    • 7.3.2.2 POSIX Unnamed Semaphores
      
      Unnamed semaphores are similar to named semaphores, but it's easier to guarantee that an unnamed semaphore can be used only by threads belonging to a certain process.
      
      However, it is possible for unnamed semaphores to be shared by threads from different processes.
      
      
  • 7.3.3 POSIX Condition Variables
    
    POSIX condition variables work much like the condition variables that were presented in chapter 6 as part of the structure of monitors. However, POSIX does not support monitors, so POSIX condition variables are implemented differently.
    
    It is necessary to use a POSIX mutex lock with a POSIX condition variable. Using the two together, it is possible for a process to wait on a condition variable for some Boolean value to become true (like waiting until x == y, where x and y are two variables).
  
  
• 7.4 Synchronization in Java
  
  Java has a kind of monitor, reentrant locks, semaphores, condition variables, atomic variables, and a compare_and_swap instruction.
  
  
  • 7.4.1 Java Monitors
    
    Every Java object has an associated lock. Programmers can declare a method to be synchronized, and that means a thread calling the method is required to hold the lock. Threads can block waiting for the lock in a structure called the entry set.
    
    When a thread begins executing in the monitor, it may need to wait for a condition to be met. (One example: a consumer that waits when the buffer is empty.) There is a wait() function for that purpose. There is also a notify() function that another thread can use to signal that the condition may have changed. (For example, to signal to a waiting consumer, a producer can call notify() after it puts an item in the buffer.) A thread that calls wait() is placed in a structure called the wait set, and it blocks.
    
    
  • 7.4.2 Java Reentrant Locks
    
    The reentrant lock is an object used for enforcing mutual exclusion. There are lock() and unlock() methods for acquiring and releasing access. If a thread calls lock() when the reentrant lock is available, or if the thread already owns the lock, then the thread returns with possession of the lock. Otherwise, the thread blocks.
    
    
  • 7.4.3 Java Semaphores
    
    Java has counting semaphores. The wait and signal operations are called acquire() and release(). The semantics of these semaphores are different from those discussed in chapter 6. It is possible to initialize a Java semaphore with a negative value. If so, threads must perform releases before acquires can succeed.
    
    
  • 7.4.4 Java Condition Variables
    
    Java condition variables are similar to those described in section 6.7. Similar to the case with POSIX condition variables, it is necessary to use a Java condition variable with a reentrant lock. There are await() and signal() methods for waiting and signaling on Java condition variables.
  
  
• 7.5 Alternative Approaches
  
  This section explores hardware and program language support for designing thread-safe concurrent applications. Such support is increasingly important because of the widespread emphasis nowadays on multicore and multithreaded computing.
  
  
  • 7.5.1 Transactional Memory
    
    
    • A transactional memory system is capable of insuring that a sequence of memory read-write operations are atomic. Such a sequence is called a memory transaction.
      
      
    • Given a transactional memory system to rely upon, a programmer can solve a mutual exclusion problem simply by specifying that a critical section is to be treated as a memory transaction.
      
      
    • With software transactional memory (STM), the compiler generates code that insures the atomicity of the memory transaction, resorting to the use of "low-level" locking, where necessary.
      
      
    • Hardware transactional memory (HTM) uses "hardware cache hierarchies and cache coherency protocols."
      
      
  • 7.5.2 OpenMP
    
    
    • OpenMP has a #pragma omp critical compiler directive that marks a code region as a critical section.
      
      
    • The compiler generates code that protects the region. It works about the same as protecting the region with a mutex.
      
      
  • 7.5.3 Functional Programming Languages
    
    
    • There has been a greater emphasis on functional programming languages because, unlike imperative languages, they do not maintain state. Consequently most of the synchronization problems discussed in this chapter are not at issue for functional language programmers.