Problem set 5: Threads and a Student-chosen Project

A. Threads

Implement multithreaded processes: processes that can have multiple independent threads of control, but that share the same virtual memory space and file descriptor table. Use kernel threads, the design where every user-visible “thread” corresponds to a struct proc.

You’ll add three system calls:

• The new sys_clone system call starts a new thread in the current process. See below for more on sys_clone.
• The new sys_gettid system call returns the current thread’s thread ID. Different threads in a process receive different values upon calling sys_gettid, but receive the same value upon calling sys_getpid. Note that the process ID returned by sys_getpid need not equal any of the sys_gettids for that process’s threads!
• The new sys_texit system call exits the current thread. When the last thread in a process exits, sys_texit behaves like sys_exit; otherwise, the process just continues with one fewer thread.

You'll need to update some existing system calls:

• sys_exit exits all threads in the current process.
• sys_waitpid waits for processes, not threads (in other words, sys_waitpid(p) waits for all the threads of p to exit).
• sys_fork should clone only the calling thread.
• You may assume that sys_execv is called only by single-threaded processes.

Run make run-testthread to check your work.

Process IDs and thread IDs

Kernel threading introduces a distinction between process and thread IDs: every user task has one of each, and they may not be equal. fork must allocate one of each.

We recommend storing the thread ID in proc::id_. This may feel backwards but it’s just easier to keep ptable indexed by id_. Add a new member, such as proc::pid_, to store the true process ID; and add a table, such as proc* pidtable[NPROC], to store processes by true PID.

[For what it’s worth, Linux does the same thing. Its task_struct::pid member is a thread ID, while the true process ID is called a “thread group ID” and stored in task_struct::tgid; Linux’s getpid system call returns tgid and its gettid returns pid. ¯\_(ツ)_/¯]

sys_clone interface

Your goal is for sys_clone(function, arg, stack_top) to set up a new thread running function(arg) on the given stack, where when that function returns, the thread exits via sys_texit. This means the initial thread stack frame should contain a return address pointing to sys_texit. Multithreading packages typically divide responsibility for this between the kernel’s system call and user-level code in the system call wrapper; most typically the stack setup is left to wrapper code.

Some examples from real operating systems:

• In Linux, the clone system call takes several arguments, but none of those arguments correspond to function or arg, and the stack_top argument doesn’t actually change the active stack pointer (it is recorded in the kernel for later use). Linux clone’s return value resembles that of fork: it is 0 in the new thread and the new thread’s ID in the original thread. User-level wrapper code examines the return value and modifies the stack as appropriate. In a Linux-like Chickadee design, the sys_clone system call would take no arguments.

  References: clone manual page; uClibc’s clone implementation (documented); musl libc’s clone implementation (undocumented, but more compact).

• In FreeBSD, the thr_new system call does take arguments resembling function, arg, and stack_top. However, when used in a threading library, the wrapper code still adds a layer of indirection so that the sys_texit equivalent is called when the thread function returns.

In-depth suggestions

The key hurdles with implementing the sys_clone wrapper function are saving the arguments until they can be used—which is generally after the clone system call returns—and wrangling assembly.

One way forward is to write the sys_clone wrapper function entirely in assembly. Create a new u-clone.S file and add u-clone.uo to PROCESS_LIB_OBJS. You’ll want to use the C++ mangled name for sys_clone, which is likely _Z9sys_clonePFiPvES_Pc (“function named sys_clone, taking arguments with types int (*)(void*), void*, and char*”). Your assembly file might look like this:

#include "obj/u-asm.h"

.text
.globl _Z9sys_clonePFiPvES_Pc
_Z9sys_clonePFiPvES_Pc:
    # ... your code here ...

Another way is to use GCC inline assembly (reference, useful overview), the GCC extension that lets you integrate assembly with C++ code. This is how the existing wrappers are written. Some tips on wrangling inline assembly:

• Use syntax like callq *%%rax to call a function whose address is stored in %rax. (In inline assembly, as in printf, the double % means a literal percent character.)
• An input register constraint like "i" (SYSCALL_TEXIT) may be required to pass a numeric constant as a parameter, since the normal constraints like "r" do not always work for constants.

Either method will need to store arguments across the syscall instruction. It’s possible to store the arguments in places including the calling thread’s stack and the system’s callee-saved registers (%rbx and %r12-%r15). If you choose registers and inline assembly, you’ll likely need this style of declaration to force the compiler to use specific registers:

register uintptr_t x asm("r12") = reinterpret_cast<uintptr_t>(arg);
    // i.e., C++ variable `x` is stored in register `%r12`

Synchronization

In psets/pset5answers.md, explain your synchronization plan for the parts of a process's state that can now be shared by multiple tasks. Pay special attention to pagetable_ and the file descriptor table.

Exiting

Process exit is a special issue for multithreaded processes. The sys_exit system call should exit all of the process’s threads, not just the calling thread. But the process’s other threads might be blocked in the kernel or running on other CPUs! The threads’ memory and struct procs cannot be deleted until they have all stopped running. You’ll want to mark the threads as “exiting,” wait for them to stop running, and only then tear down the process as a whole. This will use some of the same techniques you used in problem set 2, part G (system call interruption). The later parts of p-testthread.cc check interactions between sys_exit and sys_texit.

B. Student-chosen Project

For this part, you should design and implement a project of your choice which adds a new feature to your operating system! Once you have an initial idea, check with the course staff to make sure that your idea is in scope and isn't too hard or too easy.

Describe your project in a file psets/pset5project.md that you add to your repository. Your writeup should answer these questions:

1. What was your goal?
2. What’s your design?
3. What code did you write (what files and functions)?
4. What challenges did you encounter?
5. How can we test your work?

The writeup need not be very long; 300 words can do it if you use the words well.

The course staff believe that the following project ideas are very doable by the due date for the project:

• More visualizations: Can you create console visualizations for other aspects of the problem sets than virtual memory? For example, can you create a visualization for the disk or VFS?

• Performance optimization and stress tests: Create a thorough and comprehensive test suite, a stress test plan, and/or a performance benchmark, then improve your operating system until it is amazing.

• Support for kernel introspection via a /proc-like file system.

• Futex support.

• File system symbolic link support, device support (e.g., /dev/null, /dev/zero), mount support, etc.

More ambitious ideas include:

• Networking support. Implement device support for networking, then link that device to a networking stack, such as lwIP or picoTCP. [You may be interested in a networking project from MIT’s 6.828 class; though that class uses 32-bit x86 rather than 64-bit x86-64, network device handling support should be similar.]

• Graphics support. Implement a fancier graphics mode. Maybe port a game!

• Debugging support. Can you make DWARF debugging information, or some other kind of debugging information, available in the kernel? This might let you print line numbers in backtraces!

• Dynamic linking support.

Turnin

Fill out psets/pset5collab.md and psets/pset5answers.md. Also place an overview of your student-chosen project in psets/pset5project.md. Push to GitHub and then submit on the grading server.

Checkin: All work is due by 11:59pm Wednesday 5/4.