Mutual Exclusion
In this homework, we implement several types of mutual exclusion operators.
For concurrency homeworks such as this, qemu is not a good choice of execution environment: qemu uses a single thread to emulate one or more cores, meaning race conditions, if they happen at all, are going to be very rare indeed.
To properly test synchronization operations, we need to actually run our code on multiple cores. To do this, we run xv6 under VMWare instead of under qemu. VMWare isn't as convenient, however, so I would suggest using qemu for initial development and debugging, and then testing with VMWare.
To create a VMWare image use the qemu-img tool, like this:
qemu-img convert -O vmdk xv6.img xv6.vmdk
and the same for fs.img. Then create a VMWare VM, and replace any disks (including CD-ROM's) with these two images. xv6.vmdk first, then fs.img. Start the VM, and xv6 should boot up.
To try a new version, first shut down the VMWare VM. Then
make
your updated sources, convert the images as per above, and start the VM back up.
The lock implementations will be tested using programs starting with
race
. The first program that actually uses locks is race2.c. It includes the header file "multilock.h". By changing
#define spinlock
to
#define waitlock
etc, you can switch the lock type used.
(I tried to do this more elegantly with the ## operator and defines passed in from the Makefile, but failed. Suggestions welcome!)
Basic user space spinlock
For this part, use the atomic
xchg
operator to implement a basic spinlock. The functions spinlock_init, spinlock_acquire and spinlock_release have empty definitions in uspinlock.h. Implement real versions of these functions in uspinlock.c.
Basic user space ticket lock
In a ticket lock, an atomic increment operation is used instead of a compare and exchange. When entering acquire, take a ticket by incrementing a queue count. Then wait until your number comes up on a dequeue count. Upon release, increment the dequeue count. Fill in ticketlock.c and ticketlock.h.
Kernel-support for wait locks in userspace
Here, the thread should go to sleep until the lock becomes available. Rather than do any sort of check in user space, have the kernel decide. We'll do the joint userland/kernel locks in the next homework.
For this part, fill in waitlock.h and waitlock.c, and add the necessary system calls to support this.
Note that while only the race2 test exists right now, you
will need to support more than one lock for subsequent tests to be added shortly.