uthread is a lightweight fiber/user threading library I wrote for fun as a learning experience. It's far from "complete" but includes the following core features...
- Cooperative scheduling with fast context switching (~30M thread context switches/second on a single core)
- Synchronization primitives like mutexes, condition variables, etc. for use between user threads
- Asynchronous IO notifications for timers, sockets, etc. courtesy of libevent
Some missing features which may come in the future...
- Guard pages for stack overflow protection
uthread is designed around an N:1 (user-thread:kernel-thread) mapping model. An executor is responsible for multiplexing a set of user threads on top of a kernel thread. See executor.hpp for more detailed info. The Executor::sleep(...) and Executor::ready(...) functions are especially important for being the basis of the provided synchronization primitives and IO notification system.
Once you have a high level understanding of the executor take a look at the Mutex as an example of using the Executor::sleep(...) and Executor::ready(...) functions. The IO system works in a similar fassion and relies on libevent as a backend. You can read the libevent online book to learn more.
- tests/executor.cpp for examples of spawning, executing, and switching between user threads
- tests/mutex.cpp and tests/condition_variable.cpp for examples of using the provided synchronization primitives
- tests/io.cpp for an example of IO notifications
- examples/echo.cpp for an example of a TCP echo server
- examples/nc.cpp for a bare-bones nc clone
- examples/chat.cpp for a basic TCP chat server
You can run example <X> via GLOG_logtostderr=1 ./<X> after building.
A Vagrant VM is provided with a complete development environment. You should first bootup the VM from the vagrant directory...
vagrant up && vagrant ssh
Once the VM is up you can perform a build...
cd /uthread && \
./cmake.sh -DCMAKE_BUILD_TYPE=Release && \
cd build && \
make -j8 && \
make tests && \
make bench
The following are some good links regarding memory models, compiler/hardware instruction reordering, etc that are important to understand in concurrent environments. uthread multiplexes user threads on top of a single kernel thread so most of this is not immediately relevant but useful to understanding what kind of compiler optimizations might occur around context switches. Despite thinking I understand this stuff, I wouldn't be surprised if uthreads has some bugs with regards to these ideas :)
- http://preshing.com/20120625/memory-ordering-at-compile-time/
- http://preshing.com/20120515/memory-reordering-caught-in-the-act/
- http://preshing.com/20120913/acquire-and-release-semantics/
- https://gcc.gnu.org/wiki/Atomic/GCCMM/AtomicSync
- https://stackoverflow.com/questions/14182066/is-memory-reordering-visible-to-other-threads-on-a-uniprocessor