CSE 120 Nachos Project 1: Threads

Fall 2001

Due: October 28 at Midnight

Submission Instructions

In this assignment, we will give you part of a working thread system. Your job is to complete it and then use it to solve several synchronization problems.

The first step is to read and understand the partial thread system we've provided. This thread system implements thread fork, thread completion, and semaphores for synchronization.

Properly synchronized code should work no matter what order the scheduler chooses to run the threads on the ready list. In other words, we should be able to put a call to Thread::Yield() (which causes the scheduler to pick another thread to run) anywhere in your code where interrupts are enabled without changing the correctness of your code. You will be asked to write properly synchronized code as part of later assignments, so understanding how to do this is crucial to being able to do the project.

To aid you in this task, code linked in with Nachos will cause Thread::Yield() to be called in a repeatable but unpredictable way. Nachos code is repeatable in that if you call it repeatedly with the same arguments, it will do exactly the same thing each time. However, if you invoke Nachos with "nachos -rs #" with a different number each time, calls to Thread::Yield() will be inserted in different places in the code.

Warning: In our implementation of threads, each thread is assigned a small, fixed-size execution stack. This may cause bizarre problems (such as segmentation faults at strange lines of code) if you declare large data structures to be automatic variables (e.g., ``int buf[1000];''). You will probably not notice this during the term, but, if you do, you may change the size of the stack by modifying the StackSize #define in switch.h.

Although the solutions can be written as normal C routines, you will find organizing your code to be easier if you structure your code as C++ classes. Also, there should be no busy-waiting in any of your solutions to this assignment.

Don't worry if you don't have to write much code for each of these: the assignment is largely conceptual and not a programming chore. For some hints on getting started, here are some suggestions.

  1. [15 pts] Implement condition variables using interrupt enable and disable to provide atomicity. The file code/threads/synch.h defines the classes "Lock" and "Condition", and it is your task to implement the functions defined by those classes in synch.cc. The file badtest.cc is an example program that demonstrates how race conditions can happen inside of Nachos. It is similar in spirit to the withdraw() example from lecture. [Sample Output]

    [10 pts] You should write your code "defensively" in the sense that you should make an attempt to detect error conditions and react appropriately. For error conditions that could result in a race condition or deadlock, your library routines should exit -- there is no way to recover from these errors, so they should be fatal to the program. There is a convenient macro ASSERT() that you can use to check for error conditions and abort if necessary (grep through the Nachos source code files for examples of how to use it).

    To help motivate you to get into the habit of testing for error conditions, write test programs that test for the following errors: (1) acquiring the same Lock twice, (2) releasing a Lock that isn't held, (3) deleting a Lock that is held, (4) waiting on a condition variable without holding a Lock, (5) signalling a condition variable wakes only one thread and broadcasting wakes up all threads, (6) signalling and broadcasting to a condition variable with no waiters is a no-op, and future threads that wait will block (i.e., the signal/broadcast is "lost"). These are the minimal set of error conditions for which we'll test your Lock and Condition implementations.

  2. [15 pts] Implement synchronous send and receive of one word messages using condition variables. Create a "Mailbox" class with the operations Mailbox::Send(int message) and Mailbox::Receive(int * Message). Send atomically waits until Receive is called on the same mailbox, and then copies the message into the receive buffer. Once the copy is made, both can return. Similarly, Receive waits until Send is called, at which point the copy is made and both calls return. Your solution should work even if there are multiple senders and receivers for the same mailbox. (Hint: this is equivalent to a zero-length bounded buffer).

    You can implement the "Mailbox" class in synch.h and sync.cc, or in new files. If you create new files, be sure to update the dependency information in the Makefile; see Installing and Building Nachos from the Duke equivalent of this course for directions on how to do this.

    [5 pts] Write test cases that demonstrate that your implementation of the Mailbox class is faithful to the semantics described above: a receiver will only return when a sender sends, and blocks otherwise (and vice-versa); only one receiver and sender synchronize at a time, even when there are multiple senders and receivers.

  3. [15 pts] Implement Thread::Join(). Add a parameter to the thread constructor to indicate whether or not Join will be called on this thread. Your solution should (1) properly delete the thread control block whether or not Join is to be called, and (2) whether or not the forked thread finishes before the Join is called. Note that you do not need to implement Join so that it returns the value returned from the forked thread. The implemention of the Thread class is in code/threads/thread.{h,cc}; add the Join() method declaration in the header file, and implement the method in the source file. Use the following signatures for the updated Thread constructor and Join method:
    Thread(char* debugName, int join = 0);
    
    void Join();
    
    [5 pts] Write test cases that test for the requirements outlined above for the semantics of the Join method. Be sure to test for both possibilites of each requirement.

  4. [15 pts] Implement preemptive priority scheduling in Nachos. Priority scheduling is a key building block for real-time systems. Add a call to the Thread class to set the priority of the thread. When a thread is added to the ready list that is a higher priority than the currently running thread, the current thread gives up the processor to the new thread (a thread with the same priority as the current thread does not preempt it). Similarly, when threads are waiting for a lock, semaphore, or condition variable, the highest priority waiting thread should be woken up first. Use the following signature for the setPriority() method:
    void setPriority(int newPriority);
    

    The range of valid priorities is the entire range of an "int". Assume that all threads are created with priority 0. Roughly speaking, threads set to have a negative priority have "less priority", and threads set to have a positive priority have "more priority". Compare thread priorities directly to determine higher priority (e.g., a priority of 1 is lower than a priority of 2).

    An issue with priority scheduling is "priority inversion". If a high priority thread needs to wait for a low priority thread, such as for a lock held by a low priority thread or for a Join to complete, and a middle priority thread is on the ready list, then the high priority thread will never get the CPU because the low priority thread will not get any CPU time. A partial fix for this problem is to have the waiting thread "donate" its priority to the low priority thread while it is holding the lock. Implement this fix separately for both situations: (1) the Lock class and (2) the Join method.

    [10 pts] Write test programs that (1) demonstrate that threads with higher priority get service first in the cases outlined above (both when added to the ready list, and when woken up when waiting on a synchronization variable), and (2) demonstrate that you solve the priority inversion problem for Locks and Join().

  5. [10 pts] You have been hired by Greenpeace to help the environment. Because unscrupulous commercial interests have dangerously lowered the whale population, whales are having synchronization problems in finding a mate. The trick is that in order to have children, whales are needed, one male, one female, and one to play matchmaker -- literally, to push the other two whales together (I'm not making this up!). Your job is to write the three procedures Male(), Female(), and Matchmaker(). Each whale is represented by a separate thread. A male whale calls Male(), which waits until there is a waiting female and matchmaker; similarly, a female whale must wait until a male whale and a matchmaker are present. Once all three are present, all three return.

Submitting The Project

All the code you write should be well commented so we can understand what you changed. However, your grade on the project fundamentally depends upon how well your solutions will pass the test cases. As a result, it is important that (1) your code compiles cleanly, (2) the nachos executable will run, and (3) you write test cases to test your solutions to the problems.

You should hand in the entire threads directory, including your test cases, as well as a file named "README" containing your group name and the members of the team. In addition, we would like a short paper write-up describing what changes you made, how well they worked, how you tested your changes, and how each group member contributed to the project. An example of writing up a problem is shown in this skeleton.

To turn in your project, create a tar file of your threads directory and Mail it to Rishi, Kuo-Wen, and me. If you were group 57, use the following sequence of commands to minimize the size of the file you send:

% cd nachos-3.4/code
% gmake clean
% tar cvf group57.tar threads
% gzip group57.tar
% [mail group57.tar.gz using your favorite method of email attachments]

Look on the project groups list if you do not remember your group number.




voelker@cs.ucsd.edu