Skip to content

Latest commit

 

History

History
281 lines (221 loc) · 8.52 KB

README.md

File metadata and controls

281 lines (221 loc) · 8.52 KB
Raw

Eventually Persistent Engine

Threads

Code in ep-engine is executing in a multithreaded environment, two classes of thread exist.

  1. memcached's threads, for servicing a client and calling in via the [engine API] (https://github.com/couchbase/memcached/blob/master/include/memcached/engine.h)
  2. ep-engine's threads, for running tasks such as the document expiry pager (see subclasses of GlobalTasks).

Synchronisation Primitives

There are two mutual-exclusion primitives available in ep-engine (in addition to those provided by the C++ standard library):

  1. RWLock shared, reader/writer lock - rwlock.h
  2. SpinLock 1-byte exclusive lock - atomix.h

A condition-variable is also available called SyncObject syncobject.h. SyncObject glues a std::mutex and std::condition_variable together in one object.

These primitives are managed via RAII wrappers - locks.h.

  1. LockHolder - a deprecated alias for std::lock_guard
  2. MultiLockHolder - for acquiring an array of std::mutex or SyncObject.

Mutex

The general style is to create a std::lock_guard when you need to acquire a std::mutex, the constructor will acquire and when the lock_guard goes out of scope, the destructor will release the std::mutex. For certain use-cases the caller can explicitly lock/unlock a std::mutex via the std::unique_lock class.

std::mutex mutex;
void example1() {
    std::lock_guard<std::mutex> lockHolder(mutex);
    ...
    return;
}

void example2() {
    std::unique_lock<std::mutex> lockHolder(mutex);
    ...
    lockHolder.unlock();
    ...
    lockHolder.lock();
    ...
    return;
}

A MultiLockHolder allows an array of locks to be conveniently acquired and released, and similarly to LockHolder the caller can choose to manually lock/unlock at any time (with all locks locked/unlocked via one call).

std::mutex mutexes[10];
Object objects[10];
void foo() {
    MultiLockHolder lockHolder(&mutexes, 10);
    for (int ii = 0; ii < 10; ii++) {
        objects[ii].doStuff();
    }
    return;
}

RWLock

RWLock allows many readers to acquire it and exclusive access for a writer. Like a std::mutex RWLock can be used with a std::lock_guard. The RWLock can either be explicitly casted to a ReaderLock / WriterLock through its reader() and writer() member functions or you can rely on the implicit conversions used by the lock_guard constructor.

RWLock rwLock;
Object thing;

void foo1() {
    std::lock_guard<ReaderLock> rlh(rwLock);
    if (thing.getData()) {
    ...
    }
}

void foo2() {
    std::lock_guard<WriterLock> wlh(rwLock);
    thing.setData(...);
}

SyncObject

SyncObject inherits from std::mutex and is thus managed via a LockHolder or MultiLockHolder. The SyncObject provides the conditional-variable synchronisation primitive enabling threads to block and be woken.

The wait/wakeOne/wake method is provided by the SyncObject.

Note that wake will wake up a single blocking thread, wakeOne will wake up every thread that is blocking on the SyncObject.

SyncObject syncObject;
bool sleeping = false;
void foo1() {
    LockHolder lockHolder(&syncObject);
    sleeping = true;
    syncObject.wait(); // the mutex is released and the thread put to sleep
    // when wait returns the mutex is reacquired
    sleeping = false;
}

void foo2() {
    LockHolder lockHolder(&syncObject);
    if (sleeping) {
        syncObject.notifyOne();
    }
}

SpinLock

A SpinLock uses a single byte for the lock and our own code to spin until the lock is acquired. The intention for this lock is for low contention locks.

The RAII pattern is just like for a mutex.

SpinLock spinLock;
void example1() {
    std::lock_guard<SpinLock> lockHolder(&spinLock);
    ...
    return;
}

_UNLOCKED convention

ep-engine has a function naming convention that indicates the function should be called with a lock acquired.

For example the following doStuff_UNLOCKED method indicates that it expect a lock to be held before the function is called. What lock should be acquired before calling is not defined by the convention.

void Object::doStuff_UNLOCKED() {
}

void Object::run() {
    LockHolder lockHolder(&mutex);
    doStuff_UNLOCKED();
    return;
}

Atomic / thread-safe data structures

In addition to the basic synchronization primitives described above, there are also the following higher-level data structures which support atomic / thread-safe access from multiple threads:

  1. AtomicQueue: thread-safe, approximate-FIFO queue, optimized for multiple-writers, one reader - atomicqueue.h
  2. AtomicUnorderedMap : thread-safe unordered map - atomic_unordered_map.h

Thread Local Storage (ObjectRegistry).

Threads in ep-engine are servicing buckets and when a thread is dispatched to serve a bucket, the pointer to the EventuallyPersistentEngine representing the bucket is placed into thread local storage, this avoids the need for the pointer to be passed along the chain of execution as a formal parameter.

Both threads servicing frontend operations (memcached's threads) and ep-engine's own task threads will save the bucket's engine pointer before calling down into engine code.

Calling ObjectRegistry::onSwitchThread(enginePtr) will save the enginePtr in thread-local-storage so that subsequent task code can retrieve the pointer with ObjectRegistry::getCurrentEngine().

Tasks

A task is created by creating a sub-class (the run() method is the entry point of the task) of the GlobalTask class and it is scheduled onto one of 4 task queue types. Each task should be declared in src/tasks.defs.h using the TASK macro. Using this macro ensures correct generation of a task-type ID, priority, task name and ultimately ensures each task gets its own scheduling statistics.

The recipe is simple.

Add your task's class name with its priority into src/tasks.defs.h

  • A lower value priority is 'higher'.
TASK(MyNewTask, 1) // MyNewTask has priority 1.

Create your class and set its ID using MY_TASK_ID.

class MyNewTask : public GlobalTask {
public:
    MyNewTask(EventuallyPersistentEngine* e)
        : GlobalTask(e/*engine/,
                     MY_TASK_ID(MyNewTask),
                     0.0/*snooze*/){}
...

Define pure-virtual methods in MyNewTask

  • run method

The run method is invoked when the task is executed. The method should return true if it should be scheduled again. If false is returned, the instance of the task is never re-scheduled and will deleted once all references to the instance are gone.

bool run() {
   // Task code here
   return schedule again?;
}
  • Define the getDescription method to aid debugging and statistics.
std::string getDescription() {
    return "A brief description of what MyNewTask does";
}

Schedule your task to the desired queue.

ExTask myNewTask = new MyNewTask(&engine);
myNewTaskId = ExecutorPool::get()->schedule(myNewTask, NONIO_TASK_IDX);

The 4 task queue types are:

  • Readers - READER_TASK_IDX
  • Tasks that should primarily only read from 'disk'. They generally read from the vbucket database files, for example background fetch of a non-resident document.
  • Writers (they are allowed to read too) WRITER_TASK_IDX
  • Tasks that should primarily only write to 'disk'. They generally write to the vbucket database files, for example when flushing the write queue.
  • Auxilliary IO AUXIO_TASK_IDX
  • Tasks that read and write 'disk', but not necessarily the vbucket data files.
  • Non IO NONIO_TASK_IDX
  • Tasks that do not perform 'disk' I/O.

Utilise snooze

The snooze value of the task sets when the task should be executed. The initial snooze value is set when constructing GlobalTask. A value of 0.0 means attempt to execute the task as soon as scheduled and 5.0 would be 5 seconds from being scheduled (scheduled meaning when ExecutorPool::get()->schedule(...) is called).

The run() function can also call snooze(double snoozeAmount) to set how long before the task is rescheduled.

It is best practice for most tasks to actually do a sleep forever from their run function:

  snooze(INT_MAX);

Using INT_MAX means sleep forever and tasks should always sleep until they have real work todo. Tasks should not periodically poll for work with a snooze of n seconds.

Utilise wake()

When a task has work todo, some other function should be waking the task using the wake method.

ExecutorPool::get()->wake(myNewTaskId)`