Introduction
Concurrent programming is hard to get right. In this project, you will write a concurrent data structure and recover passwords from their hashes in a parallel program.
Setup
Like usual, you should register for the assignment by going to https://grinch.caltech.edu/register which will create you a repository on gitlab with the starter code.
The Scenario
The sysadmins lost all the password to the CMS clusters, and they need your help to recover them. Each student has been given a subset of these “CMS hashes” (which can be found at https://grinch.caltech.edu/etcshadow) to recover. Luckily, the passwords are all of the same form: a word with exactly one number inserted somewhere. Unfortunately, there are so many dictionary words, that it won’t be possible to just write a single-threaded password cracker. Instead, you’ll have to use your concurrency skills to recover these passwords! (More on this later.)
Warning 1
This project is actually not that much code, but the code you do have to write is very tricky. Code that works “most of the time” is NOT correct and will receive a low score. The point of this project is to get used to thinking about race conditions and writing code that is race-free.
Warning 2
Unlike most of the previous projects, we have not given you fully automated tests for this one. The reason for this is simple: concurrent code is very hard to test automatically. We’ll likely have to actually read your code to find concurrency bugs; so, it was not possible for us to give tests that completely exercise the functionality.
This means that passing the gitlab tests does not mean you got full credit. This will be determined later!!
Warning 3
There is virtually no skeleton code for this project. We expect you to design any structures or helper functions you need on your own. Most of the instructions for what the functionality of interfaces should be is in the header files. We strongly recommend you read those before attempting to code anything.
Google Compute Engine
Because this project will involve using many processors all at the same time, we can’t just have everyone use labradoodle
simultaneously. So, we’ve secured google compute credits
for you to use which will allow you to have your own server that is uncontested. You can redeem the credits using this link:
The link will ask you to enter your name and e-mail. Then, it will send you an e-mail verification. Finally, it will allow you to redeem the coupon.
Creating Your VM
When you follow these steps, your coupon will be charged unless you TURN OFF the GCE machine. The coupon only gets you around 48 hours of usage of a machine. So, either do this at the end, or shutdown the machine when you’re not using it.
Once you’ve redeemed the coupon, go to
https://console.cloud.google.com/projectselector/compute/instances
Then, choose “Create a Project” (which is all the way at the bottom). It will take a minute for GCE to create your project, but once it’s done, create a VM instance by clicking “Create”.
You will need to change two settings. First change the “Machine type” to e2-standard-16
which will give you 16 cores to work with.
Second, click on “Management, security, disks, networking, sole tenancy” and choose the “Security” Tab. You will need to enter your SSH key which we generated for you when you first connected
to labradoodle
. On labradoodle
, run the command cat ~/.ssh/id_rsa.pub
and copy the output into the big box labelled “SSH”. Take note of the username that appears on the left–that will
be the user you use to connect to your remote machine.
SSHing to Your VM
Once the VM is created, you should see an IP address on the right side of the screen under “External IP”. This is the IP address you will use to connect to your machine.
In a terminal on labradoodle
, run the command ssh USER@IP
where USER
is the username from the previous step and IP
is the external IP listed on the GCE website. If all
goes well, you should be connected to your remote machine. Once you’ve connected, run sudo apt install -y git clang make
to install the necessary software for the project.
Working with Two Machines
We recommend you WRITE your code on labradoodle
as usual, and only TEST on your GCE machine. You will probably find it easiest to just use scp -r <code directory> USER@IP:
(Note the trailing colon–this is important!) to get your code to the remote machine (from labradoodle
). This will avoid you having to generate yet another key to use with the new machine.
Part 0: Writing a Queue (50 points)
Ultimately, for our threadpool to work, it will need an underlying (thread-safe) queue which we can use to queue up tasks to be completed by the threads in the threadpool. In this part, you will write a linked-list-based FIFO queue. A queue implementation that doesn’t use a linked list will not receive any credit!
Task 1.
Write a FIFO queue in queue.c
conforming to the interface in queue.h
. You may add things to queue.h
, but you may not remove them. In this task, you should only write a single-threaded
queue (meaning, no pthreads
involved). For this version of the queue, queue_dequeue
should immediately return NULL
if there is nothing in the queue. (This is called a non-blocking
implementation.)
To test your queue, run make DEBUG=true test_squeue
. Note that these are very simple tests and will likely not find all bugs.
Task 2.
In this task, you will update your queue to support multiple threads adding and removing. Same rules as before, except, now, queue_dequeue
should “block” rather than returning
if there is nothing in queue. In other words, queue_dequeue
should wait until another thread enqueues something. If you implement a busy-wait solution, you will receive no
credit for any part that uses your multi-threaded queue.
To test your queue, run make DEBUG=true test_queue
. Note that these are very simple tests and will likely not find all bugs.
Part 1: Writing a Thread Pool (15 points)
Creating threads is relatively expensive, and only so many of them can run at once (i.e., one per processor core). So, to avoid wasting the time instantiating one thread per task, we can use a thread pool which allocates tasks to threads that are part of the “pool”. Then, these threads get “reused” when they finish their tasks.
You should create your pthread
s in thread_pool_init
. Then, these threads (called “workers”) should wait until a queue that is shared across all of them has work to be done. Once
there is work, one worker should wake up, complete that work, and wait for more work.
Task 3.
Write a thread pool conforming to the interface in thread_pool.h
. Your implementation should enqueue one NULL
task per thread in the pool, when the user is done providing tasks, to indicate
that the workers should stop. That is, thread_pool_finish
should enqueue a bunch of NULL
tasks before waiting on the threads, and if a worker encounters a NULL
task, it should return.
To test your thread pool, run make DEBUG=true test
. Note that these are very simple tests and will likely not find all bugs.
Part 2: Recovering Passwords (30 points)
Passwords are not stored as plaintext by your computer. Instead, when you log into a machine, it hashes the password and compares the hash to a stored hash in a file called /etc/shadow
.
If an attacker can get access to the /etc/shadow
file, they can brute force the password by figuring out what input hashes to the required result. That is exactly what you’ll do in this
step. You will write a client for your threadpool that brute forces passwords from their corresponding hashes. Each student has a different set of passwords to reverse which can be found here:
https://grinch.caltech.edu/etcshadow
All passwords are a word in the DICTIONARY
augmented with a single digit (0-9) somewhere in the word (which could be before or after it).
For testing purposes, here is a hash of the password q6uests
:
$6$sI0nXQzT/Pg.zJae$MCYdZmuv89Q6Oij.bXWArQOXcsDvOrN.PdiB8Za6lXtaajwdb6GkgXGy3i/fTYogJT6kiMWb9BHBfnB3HTme90
We have provided some starter code for this section in src/password_cracker.c
. Specifically, you will find hashes_match
to be useful to compare a word with a hash, and we have written
the annoying parsing code for the password hashes. Each “task” that your threadpool executes should consist of trying a single dictionary word (and its variants) on all possible hashes. You
should print any recovered passwords to standard out.
We have provided you with a simple “prime printing” client in tests/prime_printer.c
which you might consider referencing when writing your password cracker.
Task 4. Write a password cracker to the above specification. Note that this program will be very slow. To run the program, you should do the following steps:
- Put your hashes in
hashes.txt
-
scp
all your files to your GCE machine -
ssh
to your GCE machine - Run
make passwords.txt
These commands should run your password cracker in the background; so, you do not have to stay connected to your machine as you are running it. Please note that a correct version of the program might take several hours to run!!!