Rseq

Rseq is a userspace take on the proposed kernel restartable sequences API, and provides and mechanism to perform efficient per-cpu operations.

This isn't intended to be a long-running project. Instead, its goal is to allow userland experiments with rseq (without having to recompile the kernel), to collect data as to how useful it would be.

Example

Here is a simple demonstration of how to use rseq and why it might be useful. A more thorough explanation (together with history and some implementation notes) can be found in Rseq.md.

rseq::Value<int> counterByCpu[kNumCpus];
void bumpCounter() {
  while (true) {
    int cpu = rseq::begin();
    // A plain atomic load. rseq::Value types are API-compatible with
    // std::atomic.
    int curValue = counterByCpu[cpu].load();
    // rseq::store takes no action and returns false if another thread might
    // have run on the same CPU after the call to rseq::begin(). Otherwise, the
    // store happens and the call returns true.
    bool success = rseq::store(&counterByCpu[cpu], curValue + 1);
    if (success) {
      break;
    }
  }
}

Requirements

Rseq only works on Linux and x86-64. Building requires CMake and a recent version of clang or g++. Building the tests requires a gtest installation.

Building Rseq

In this directory, run:

mkdir build
cd build
# Include the former option to produce an optimized build, and the latter to
# enable running tests.
cmake [-DCMAKE_BUILD_TYPE=Release] [-Dtest=ON] [-DCMAKE_INSTALL_PREFIX=</path/to/install/dir>] ../
make

# Now we can take some of our binaries for a test drive

# If you passed -Dtest=ON above, this will run all tests.
make test
# Run a benchmark of a variety of mechanisms for incrementing a set of
# counters.
./rseq_benchmark all 8 10000000

Installing Rseq

For the common case, you probably want:

mkdir build && cd build
cmake -DCMAKE_BUILD_TYPE=Release ../
sudo make install

You can then compile programs that #include "rseq/Rseq.h" with g++ myProgram.cpp -lrseq.

How Rseq works

See Rseq.md for a more thorough description. Essentially, each thread gets its own copy of the code that does an rseq operation. When one thread wants to evict another from ownership of a CPU, it patches that thread's copy of the function to jump to a failure path instead of doing the operation.

Full documentation

Rseq.md contains a more thorough description. Reading the comments in rseq/Rseq.h should give a working understanding of the API.

License

Rseq is BSD-licensed. We also provide an additional patent grant.

doronrk/Rseq