Programming Project 2:

Threads and Interprocess Communication

 

 

Due Date: ______________________________

Duration:  One Week

 

 

Overview and Goal

 

In this project you will learn about threads and gain familiarity writing programs involving concurrency control.  You will begin by studying the thread package, which implements multi-threading.  You will make modifications and additions to the existing code.  Then you will use the threads package to solve some traditional concurrency problems.

 

In addition, you will gain familiarity programming in the KPL language.

 

If you have difficulties with this project or want to go into more depth, take a look at the document called ╥The Thread Scheduler and Concurrency Control Primitives╙:

 

            http://web.cecs.pdx.edu/~harry/Blitz/BlitzDoc/ThreadScheduler.htm

            http://web.cecs.pdx.edu/~harry/Blitz/BlitzDoc/ThreadScheduler.pdf

 

 

Step 1:  Download the Files

 

Start by creating a directory for the files you╒ll need for this project.  You might call it:

 

     ~YourUserName/cs333/p2

 

Then download all the files from:

 

            http://www.cs.pdx.edu/~harry/Blitz/OSProject/p2/

 

into your directory.  You should get the following files:

 

     makefile

     DISK

     System.h

     System.c

     Runtime.s

     Switch.s

     List.h

     List.c

     Thread.h

     Thread.c

     Main.h

     Main.c

     Synch.h

     Synch.c

 

In this project, you will modify and hand in the following files:

 

     Main.c

     Synch.h

     Synch.c

 

After getting the files, you should be able to compile all the code (as is) with the UNIX make command.  The program executable we are building will be called ╥os╙.  You can execute the program by typing:

 

     % make

     % blitz –g os

 

In the course of your experimentation, you may modify other files besides Synch.h, Synch.c and Main.c, but the code you are required to write and turn in doesn╒t require any changes to the other files.  For example, you may wish to uncomment some of the print statements, to see what happens.  However, your final versions of Synch.h, Synch.c and Main.c must work with the other files, exactly as they are distributed.

 

Be sure you copy all files.  Even though there are similarities with some of the files used for the previous project, there may be some subtle but important differences.

 

 

Step 2:  Study the Existing Code

 

The code provided in this project provides the ability to create and run multiple threads, and to control concurrency through several synchronization methods.

 

Start by looking over the System package.  Focus on the material toward the beginning of file System.c, namely the functions:

 

     print

     printInt

     printHex

     printChar

     printBool

     nl

     MemoryEqual

     StrEqual

     StrCopy

     StrCmp

     Min

     Max

     printIntVar

     printHexVar

     printBoolVar

     printCharVar

     printPtr

 

Get familiar with the printing functions; you╒ll be calling them a lot.  Some are implemented in assembly code and some are implemented in KPL in the System package.

 

The following functions are used to implement the heap in KPL:

 

     KPLSystemInitialize

     KPLMemoryAlloc

     KPLMemoryFree

 

Objects can be allocated on the heap and freed with the alloc and free statements.  The HEAP implementation is very rudimentary in this implementation.  In your kernel, you may allocate objects during start-up but after that, YOU MUST NOT ALLOCATE OBJECTS ON THE HEAP!  Why?  Because the heap might fill up and then what is a kernel supposed to do?  Crash.

 

In this project, you should not allocate anything on the heap.

 

The following functions can be ignored since they concern aspects of the KPL language that we will not be using:

 

     KPLUncaughtThrow

     UncaughtThrowError

     KPLIsKindOf

     KPLSystemError

 

The Runtime.s file contains a number of routines coded in assembly language.  It contains the program entry point and the interrupt vector in low memory.  Take a look at it.  Follow what happens when program execution begins at location 0x00000000 (the label ╥_entry╙).  The code labeled ╥_mainEntry╙ is included in code the compiler produces.  The ╥_mainEntry╙ code will call the main function, which appears in the file Main.c.

 

In Runtime.s, follow what happens when a timer interrupt occurs.  It makes an ╥up-call╙ to a routine called _P_Thread_TimerInterruptHandler.  This name refers to ╥a function called TimerInterruptHandler in a package called Thread.╙  (_P_Thread_TimerInterruptHandler is the name the compiler will give this function.)

 

All the code in this project assumes that no other interrupt types (such as a DiskInterrupt) occur.  In Runtime.s, follow what would happen if another sort of interrupt should ever occur.

 

The KPL language will check for lots of error conditions, such as the use of a null pointer.  Try changing the program to make this error.  Follow in Runtime.s what happens when this occurs.

 

Next take a look at the List package.  Read the header file carefully.  This package provides code that implements a linked list.  We╒ll use linked lists in this project.  For example, the threads that are ready to run (and waiting for time on the CPU) will be kept in a linked list called the ╥ready list╙.  Threads that become BLOCKED will sit on other linked lists.  Also look over the List.c code file.

 

The key class in this project is named Thread, and it is located in the Thread package along with other stuff (see Thread.h, Thread.c).  For each thread, there will be a single Thread object.  Thread is a subclass of Listable, which means that each Thread object contains a next pointer and can be added to a linked list.

 

The Thread package is central and you should study this code thoroughly.  This package contains one class (called Thread) and several functions related to thread scheduling and time-slicing:

 

     InitializeScheduler ()

     IdleFunction (arg: int)

     Run (nextThread: ptr to Thread)

     PrintReadyList ()

     ThreadStartMain ()

     ThreadFinish ()

     FatalError (errorMessage: ptr to array of char)

     SetInterruptsTo (newStatus: int) returns int

     TimerInterruptHandler ()

 

FatalError is the simplest function.  We will call FatalError whenever we wish to print an error message and abort the program.  Typically, we╒ll call FatalError after making some check and finding that things are not as expected.  FatalError will print the name of the thread invoking it, print the message, and then shut down.  It will throw us into the BLITZ emulator command line mode.  Normally, the next thing to do might be to type the ╥st╙ command (short for ╥stack╙), to see which functions and methods were active.

 

(Of course the information printed out by the emulator will pertain to only the thread that invoked FatalError.  The emulator does not know about threads, and it is pretty much impossible to extract information about other threads by examining bytes in memory.)

 

The next function to look at is SetInterruptsTo, which is used to change the ╥I╙ interrupt bit in the CPU.  We can use it to disable interrupts with code like this:

 

     ... = SetInterruptsTo (DISABLED)

 

and we can use it to enable interrupts:

 

     ... = SetInterruptsTo (ENABLED)

 

This function returns the previous status.  This is very useful because we often want to DISABLE interrupts (regardless of what they were before) and then later we want to return the interrupt status to whatever it was before.  In our kernel, we╒ll often see code like:

 

     var oldIntStat: int

     ...

     oldIntStat = SetInterruptsTo (DISABLED)

     ...

     oldIntStat = SetInterruptsTo (oldIntStat)

 

Next take a look at the Thread class.  Here are the fields of Thread:

 

     name: ptr to array of char

     status: int

     systemStack: array [SYSTEM_STACK_SIZE] of int

     regs: array [13] of int

     stackTop: ptr to void

initialFunction: ptr to function (int)

initialArgument: int

 

Here are the operations (i.e., methods) you can do on a Thread:

 

     Init (n: ptr to array of char)

     Fork (fun: ptr to function (int), arg: int)

     Yield ()

     Sleep ()

     CheckOverflow ()

     Print ()

 

Each thread is in one of the following states:  JUST_CREATED, READY, RUNNING, BLOCKED, and UNUSED, and this is given in the status field.  (The UNUSED status is given to a Thread after it has terminated.  We╒ll need this in later projects.)

 

Each thread has a name.  To create a thread, you╒ll need a Thread variable.  First, use Init to initialize it, providing a name.

 

Each thread needs its own stack and space for this stack is placed directly in the Thread object in the field called systemStack.  Currently, this is an array of 1000 words, which should be enough.  (It is conceivable our code could overflow this limit; there is a check to make sure we don╒t overflow this limited area.)

 

All threads in this project will run in System mode.  Therefore the stack is called the ╥system stack╙.  In later projects, we╒ll see that this stack is used only for kernel routines.  User programs will have their own stacks in their virtual address spaces, but we are getting ahead of ourselves.

 

The Thread object also has space to store the state of the CPU, namely the registers.  Whenever a thread switch occurs, the registers will be saved in the Thread object.  These fields (regs and stackTop) are used by the assembly code routine named Switch.

 

The Thread object also has space to store a pointer to a function (the initialFunction field) and an argument for this function (the initialArgument field).  This pointer will point to the ╥main╙ function of this thread; this is the function that will get executed when this thread begins execution.  We are storing a pointer to the function because this is a variable: different threads may execute different functions.

 

We are also able to supply an initial argument to this thread, through the initialArgument field.  This argument must be an integer.  Often there will be several threads executing the same ╥main╙ function.  The argument is a handy way to let each thread know what its role should be.  For example, we might create 10 threads each using the same ╥main╙ function, but passing each thread a different integer (say, between 1 and 10) to let it know which thread it is.

 

After initializing a new Thread, we can start it running with the Fork method.  This doesn╒t immediately begin the thread execution; instead it makes the thread READY to run and places it on the readyList.  The readyList is a linked list of Threads.  All Threads on the readyList have status READY.  ReadyList is a global variable.  There is another global variable named currentThread, which points to the currently executing Thread object; i.e., the Thread whose status is RUNNING.

 

The Yield method should only be invoked on the currently running thread.  It will cause a switch to some other thread.

 

Follow the code in Yield closely to see what happens when a thread switch occurs.  First, interrupts are disabled; we don╒t want any interference during a thread switch.  The readyList and currentThread are shared variables and, while switching threads, we want to be able to access and update them safely.  Then Yield will find the next thread from the readyList.  (If there is no other thread, then Yield is effectively a nop.)  Then Yield will make the currently running process READY (i.e., no longer RUNNING) and it will add the current thread to the tail end of the readyList.  Finally, it will call the Run function to do the thread switch.

 

The Run method will check for stack overflow on the current thread.  It will then call Switch to do the actual Switch.

 

Switch may be the most fascinating function you ever encounter!  It is located in the assembly code file Switch.s, which you should look at carefully.  Switch does not return to the function that called it.  Instead, it switches to another thread.  Then it returns.  Therefore, the return happens to another function in another thread!

 

The only place Switch is called is from the Run function, so Switch returns to some invocation of the Run function in some other thread.  That copy (i.e., invocation) of Run will then return to whoever called it.  This could have been some other call to Yield, so we╒ll return to another Yield which will return to whoever called it.

 

And this is exactly the desired functionality of Yield!  A call to Yield should give up the processor for a while, and eventually return after other threads have had a chance to execute.

 

Run is also called from Sleep, so we might be returning from a call to Sleep after a thread switch.

 

How is everything set up when a thread is first created?  How can we ╥return to a function╙ when we have not ever called it?  Take a look at function ThreadStartMain in file Thread.c and look at function ThreadStartUp in file Switch.s.

 

What happens when a thread is terminated?  Take a look at ThreadFinish in file Thread.c.  Essentially, the thread is put to sleep with no hope of ever being awakened.  (No wonder they call it ╥Thread Death!╙)

 

Next, take a look at what happens when a Timer interrupt occurs while some thread is executing.  This is an interrupt, so the CPU begins by interrupting the current routine╒s execution and pushing some state onto its (system) stack.  Then it disables interrupts and jumps to the assembly code routine called TimerInterruptHandler in Runtime.s, which just calls the TimerInterruptHandler function in Thread.c.

 

In TimerInterruptHandler, we call Yield, which then switches to another thread.  Later, we╒ll come back here, when this thread gets another chance to run.  Then, we╒ll return to the assembly language routine which will execute a ╥reti╙ instruction.  This will restore the state to exactly what it was before and the interrupted routine (whatever it was) will get to continue.

 

Note that this code maintains a variable called currentInterruptStatus.  This is because it is rather difficult to query the ╥I╙ bit of the CPU.  It is easier to just change the variable whenever a change to the interrupt status changes.  We see this occurring in the TimerInterruptHandler function.  Clearly interrupts will be disabled immediately after the interrupt occurs.  And the Yield function will preserve the interrupt status.  So when we return from Yield, interrupts will once again be disabled.  Before returning to the interrupted thread, we set the currentInterruptStatus to ENABLED.  (They must have been enabled before the interrupt occurred—or else it could not have occurred—so after we execute the ╥reti╙ instruction, the status will revert to what it was before, namely ENABLED.)

 

Now you are ready to start playing with and modifying the code!  Please experiment with the code we have just discussed, as necessary to understand it.

 

 

 

Step 3:  Run the ╥SimpleThreadExample╙ Code

 

Execute and trace through the output of SimpleThreadExample in file Main.c.

 

In TimerInterruptHandler there is a statement

 

     printChar ('_')

 

which is commented out.  Try uncommenting it.  Make sure you understand the output.

 

In TimerInterruptHandler, there is a call to Yield.  Why is this there?  Try commenting this statement?  What happens.  Make sure you understand how Yield works here.

 

 

 

 

Step 4:  Run the ╥MoreThreadExamples╙ Code

 

Trace through the output.  Try changing this code to see what happens.

 

 

 

 

Step 5:  Implement the ╥Mutex╙ Class

 

In this part, you must implement the class Mutex.  The class specification for Mutex is given to you in Synch.h:

 

  class Mutex

    superclass Object

    methods

      Init ()

      Lock ()

      Unlock ()

      IsHeldByCurrentThread () returns bool

  endClass

 

You will need to provide code for each of these methods.  In Synch.c you╒ll see a behavior construct for Mutex.  There are methods for Init, Lock, Unlock, and IsHeldByCurrentThread, but these have dummy bodies.  You╒ll need to write the code for these four methods.

 

You will also need to add a couple of fields to the class specification of Mutex to implement the desired functionality.

 

How can you implement the Mutex class?  Take a close look at the Semaphore class; your implementation of Mutex will be quite similar.

 

First consider the IsHeldByCurrentThread method, which may be invoked by any thread.  The code of this method will need to know which thread is holding a lock on the mutex; then it can compare that to the currentThread to see if they are the same.  So, you might consider adding a field (perhaps called heldBy) to the Mutex class, which will be a pointer to the thread holding the mutex.  Of course, you╒ll need to set it to the current thread whenever the mutex is locked.  You might use a null value in this field to indicate that no thread is holding a lock on the mutex.

 

When a lock is requested on the mutex, you╒ll need to see if any thread already has a lock on this mutex.  If so, you╒ll need to put the current process to sleep.  For putting a thread to sleep, take a look at the method Semaphore.Down.  At any one time, there may be zero, one, or many threads waiting to acquire a lock on the mutex; you╒ll need to keep a list of these threads so that when an Unlock is executed, you can wake up one of them.  As in the case of Semaphores, you should use a FIFO queue, waking up the thread that has been waiting longest.

 

When a mutex lock is released (in the Unlock method), you╒ll need to see if there are any threads waiting to acquire a lock on the mutex.  You can choose one and move it back onto the readyList.  Now the waiting thread will begin running when it gets a turn.  The code in Semaphore.Up does something similar.

 

It is also a good idea to add an error check in the Lock method to make sure that the current thread asking to lock the mutex doesn╒t already hold a lock on the mutex.  If it does, you can simply invoke FatalError.  (This would probably indicate a logic error in the code using the mutex.  It would lead to a deadlock, with a thread frozen forever, waiting for itself to release the lock.)  Likewise, you should also add a check in Unlock to make sure the current thread really does hold the lock and call FatalError if not.  You╒ll be using your Mutex class later, so these checks will help your debugging in later projects.

 

The function TestMutex in Main.c is provided to exercise your implementation of Mutex.  It creates 7 threads that compete vigorously for a single mutex lock.

 

 

 

Step 6:  Implement the Producer-Consumer Solution

 

Both the Tanenbaum and Silberschatz, et al. textbooks contain a discussion of the Producer-Consumer problem, including a solution.  Implement this in KPL using the classes Mutex and Semaphore.  Deal with multiple producers and multiple consumers, all sharing a single bounded buffer.

 

The Main package contains some code that will serve as a framework.  The buffer is called buffer and contains up to BUFFER_SIZE (e.g., 5) characters.  There are 5 producer processes, each modeled by a thread, and 3 consumer processes, each modeled by a thread.  Thus, there are 8 threads in addition to the main thread that creates the others.

 

Each producer will loop, adding 5 characters to the buffer.  The first producer will add five ╘A╒ characters, the second producer will add five ╘B╒s, etc.  However, since the execution of these threads will be interleaved, the characters will be added in a somewhat random order.

 

 

 

Step 7:  Implement the Dining Philosopher╒s Solution Using a Monitor

 

A starting framework for your solution is provided in Main.c.  Each philosopher is modeled with a thread and the code we╒ve provided sets up these threads.  The synchronization will be controlled by a ╥monitor╙ called ForkMonitor.

 

The code for each thread/philosopher is provided for you.  Look over the PhilosophizeAndEat method; you should not need to change this code.

 

The monitor to control synchronization between the threads is implemented with a class called ForkMonitor.  The following class specification of ForkMonitor is provided:

 

  class ForkMonitor

    superclass Object

    fields

      status: array [5] of int      -- For each philosopher: HUNGRY,

                                    -- EATING, or THINKING

    methods

      Init ()

      PickupForks (p: int)

      PutDownForks (p: int)

      PrintAllStatus ()

  endClass

 

You╒ll need to provide the code for the Init, PickupForks and PutDownForks methods.  You╒ll also need to add additional fields and perhaps even add another method.

 

The code for PrintAllStatus is provided.  You should call this method whenever you change the status of any philosopher.  This method will print a line of output, so you can see what is happening.

 

How can you proceed?  You╒ll need a mutex to protect the monitor itself.  There are two main methods (PickupForks and PutDownForks) which are called by the philosopher threads.  Upon beginning each of these methods, the first thing is to lock the monitor mutex.  This will ensure that only one thread at a time is executing within the monitor.  Just before each of these methods returns, it must unlock the monitor (by unlocking the monitor╒s mutex) so that other threads can enter the monitor code.

 

You╒ll also need to use the Condition class, which is provided in the Synch package.  (The Condition class uses the class Mutex, so it is assumed that you╒ve finished and tested the Mutex class.)

 

The BLITZ emulator has a number of parameters and one of these is how often a timer interrupt occurs.  The default value is every 5000 instructions.  You might try changing this parameter to see how it affects your programs behavior.  To change the simulation parameters, type the sim command into the emulator.  This command will give you the option to create a file called

 

     .blitzrc

 

After creating this file, you can edit it by hand.  The next time you run the emulator, it will use this new value.  Also note that too small a value—like 1000—will cause the program to hang.  What do you suppose causes this effect?

 

QUESTION:  Both textbooks discuss the semantics of signaling a condition variable.  They mention ╥Hoare semantics.╙  The comments in the code in Synch.c say that this version implements ╥Mesa-style╙ semantics.  Is this the same or different from Hoare semantics?

 

 

 

An Example of Correct Output

 

The following files contain an example of what correct output should look like:

 

     DesiredOutput1.pdf

     DesiredOutput2.pdf

     DesiredOutput3.pdf

 

 

 

What to Hand In

 

Complete all the above steps.

 

Please submit hardcopy of the following files:

 

     Synch.h

     Synch.c

     Main.c

 

Also include hardcopy showing the output for steps 5, 6, and 7.

 

Please print both your code and the output using a fixed-width font like this, in this and all future assignments.  Much of the output is designed to look nice when printed with a fixed-width font, but is more difficult to read in a standard variable-width font.

 

In LARGE BLOCK LETTERS, write your full name on the first page.

 

PLEASE STAPLE ALL PAGES!

 

 

Basis for Grading

 

In this course, the code you submit will be graded on both correctness and style.  Correctness means that the code must work right and not contain bugs.  Style means that the code must be neat, well commented, well organized, and easy to read.

 

In style, your code should match the code we are distributing to you.  The indentation should be the same and the degree of commenting should be the same.