dom/media/systemservices/MediaUtils.h
author Jan-Ivar Bruaroey <jib@mozilla.com>
Sat, 03 Oct 2015 20:42:26 -0400
changeset 296737 1ffe42de58bd7449cc05d1a5a2eb2db07f566169
parent 296676 4e86a0d1261a10ce8c1a81a43a5329a4340411a2
permissions -rw-r--r--
Bug 1210852 - do SelectSettings of device capabilities on media thread. r=jib a=lizzard

/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
/* vim: set sw=2 ts=8 et ft=cpp : */
/* This Source Code Form is subject to the terms of the Mozilla Public
 * License, v. 2.0. If a copy of the MPL was not distributed with this file,
 * You can obtain one at http://mozilla.org/MPL/2.0/. */

#ifndef mozilla_MediaUtils_h
#define mozilla_MediaUtils_h

#include "nsAutoPtr.h"
#include "nsThreadUtils.h"

namespace mozilla {
namespace media {

/*
 * media::Pledge - A promise-like pattern for c++ that takes lambda functions.
 *
 * Asynchronous APIs that proxy to another thread or to the chrome process and
 * back may find it useful to return a pledge to callers who then use
 * pledge.Then(func) to specify a lambda function to be invoked with the result
 * later back on this same thread.
 *
 * Callers will enjoy that lambdas allow "capturing" of local variables, much
 * like closures in JavaScript (safely by-copy by default).
 *
 * Callers will also enjoy that they do not need to be thread-safe (their code
 * runs on the same thread after all).
 *
 * Advantageously, pledges are non-threadsafe by design (because locking and
 * event queues are redundant). This means none of the lambdas you pass in,
 * or variables you lambda-capture into them, need be threasafe or support
 * threadsafe refcounting. After all, they'll run later on the same thread.
 *
 *   nsRefPtr<media::Pledge<Foo>> p = GetFooAsynchronously(); // returns a pledge
 *   p->Then([](const Foo& foo) {
 *     // use foo here (same thread. Need not be thread-safe!)
 *   });
 *
 * See media::CoatCheck below for an example of GetFooAsynchronously().
 */

class PledgeBase
{
public:
  NS_INLINE_DECL_REFCOUNTING(PledgeBase);
protected:
  virtual ~PledgeBase() {};
};

template<typename ValueType, typename ErrorType = nsresult>
class Pledge : public PledgeBase
{
  // TODO: Remove workaround once mozilla allows std::function from <functional>
  // wo/std::function support, do template + virtual trick to accept lambdas
  class FunctorsBase
  {
  public:
    FunctorsBase() {}
    virtual void Succeed(ValueType& result) = 0;
    virtual void Fail(ErrorType& error) = 0;
    virtual ~FunctorsBase() {};
  };

public:
  explicit Pledge() : mDone(false), mRejected(false) {}
  Pledge(const Pledge& aOther) = delete;
  Pledge& operator = (const Pledge&) = delete;

  template<typename OnSuccessType>
  void Then(OnSuccessType aOnSuccess)
  {
    Then(aOnSuccess, [](ErrorType&){});
  }

  template<typename OnSuccessType, typename OnFailureType>
  void Then(OnSuccessType aOnSuccess, OnFailureType aOnFailure)
  {
    class Functors : public FunctorsBase
    {
    public:
      Functors(OnSuccessType& aOnSuccess, OnFailureType& aOnFailure)
        : mOnSuccess(aOnSuccess), mOnFailure(aOnFailure) {}

      void Succeed(ValueType& result)
      {
        mOnSuccess(result);
      }
      void Fail(ErrorType& error)
      {
        mOnFailure(error);
      };

      OnSuccessType mOnSuccess;
      OnFailureType mOnFailure;
    };
    mFunctors = new Functors(aOnSuccess, aOnFailure);

    if (mDone) {
      if (!mRejected) {
        mFunctors->Succeed(mValue);
      } else {
        mFunctors->Fail(mError);
      }
    }
  }

  void Resolve(const ValueType& aValue)
  {
    mValue = aValue;
    Resolve();
  }

  void Reject(ErrorType rv)
  {
    if (!mDone) {
      mDone = mRejected = true;
      mError = rv;
      if (mFunctors) {
        mFunctors->Fail(mError);
      }
    }
  }

protected:
  void Resolve()
  {
    if (!mDone) {
      mDone = true;
      MOZ_ASSERT(!mRejected);
      if (mFunctors) {
        mFunctors->Succeed(mValue);
      }
    }
  }

  ValueType mValue;
private:
  ~Pledge() {};
  bool mDone;
  bool mRejected;
  ErrorType mError;
  ScopedDeletePtr<FunctorsBase> mFunctors;
};

/* media::NewRunnableFrom() - Create an nsRunnable from a lambda.
 * media::NewTaskFrom()     - Create a Task from a lambda.
 *
 * Passing variables (closures) to an async function is clunky with nsRunnable:
 *
 *   void Foo()
 *   {
 *     class FooRunnable : public nsRunnable
 *     {
 *     public:
 *       FooRunnable(const Bar &aBar) : mBar(aBar) {}
 *       NS_IMETHOD Run()
 *       {
 *         // Use mBar
 *       }
 *     private:
 *       nsRefPtr<Bar> mBar;
 *     };
 *
 *     nsRefPtr<Bar> bar = new Bar();
 *     NS_DispatchToMainThread(new FooRunnable(bar);
 *   }
 *
 * It's worse with more variables. Lambdas have a leg up with variable capture:
 *
 *   void Foo()
 *   {
 *     nsRefPtr<Bar> bar = new Bar();
 *     NS_DispatchToMainThread(media::NewRunnableFrom([bar]() mutable {
 *       // use bar
 *     });
 *   }
 *
 * Capture is by-copy by default, so the nsRefPtr 'bar' is safely copied for
 * access on the other thread (threadsafe refcounting in bar is assumed).
 *
 * The 'mutable' keyword is only needed for non-const access to bar.
 */

template<typename OnRunType>
class LambdaRunnable : public nsRunnable
{
public:
  explicit LambdaRunnable(OnRunType& aOnRun) : mOnRun(aOnRun) {}
private:
  NS_IMETHODIMP
  Run()
  {
    return mOnRun();
  }
  OnRunType mOnRun;
};

template<typename OnRunType>
LambdaRunnable<OnRunType>*
NewRunnableFrom(OnRunType aOnRun)
{
  return new LambdaRunnable<OnRunType>(aOnRun);
}

template<typename OnRunType>
class LambdaTask : public Task
{
public:
  explicit LambdaTask(OnRunType& aOnRun) : mOnRun(aOnRun) {}
private:
  void
  Run()
  {
    return mOnRun();
  }
  OnRunType mOnRun;
};

template<typename OnRunType>
LambdaTask<OnRunType>*
NewTaskFrom(OnRunType aOnRun)
{
  return new LambdaTask<OnRunType>(aOnRun);
}

/* media::CoatCheck - There and back again. Park an object in exchange for an id.
 *
 * A common problem with calling asynchronous functions that do work on other
 * threads or processes is how to pass in a heap object for use once the
 * function completes, without requiring that object to have threadsafe
 * refcounting, contain mutexes, be marshaled, or leak if things fail
 * (or worse, intermittent use-after-free because of lifetime issues).
 *
 * One solution is to set up a coat-check on the caller side, park your object
 * in exchange for an id, and send the id. Common in IPC, but equally useful
 * for same-process thread-hops, because by never leaving the thread there's
 * no need for objects to be threadsafe or use threadsafe refcounting. E.g.
 *
 *   class FooDoer
 *   {
 *     CoatCheck<Foo> mOutstandingFoos;
 *
 *   public:
 *     void DoFoo()
 *     {
 *       nsRefPtr<Foo> foo = new Foo();
 *       uint32_t requestId = mOutstandingFoos.Append(*foo);
 *       sChild->SendFoo(requestId);
 *     }
 *
 *     void RecvFooResponse(uint32_t requestId)
 *     {
 *       nsRefPtr<Foo> foo = mOutstandingFoos.Remove(requestId);
 *       if (foo) {
 *         // use foo
 *       }
 *     }
 *   };
 *
 * If you read media::Pledge earlier, here's how this is useful for pledges:
 *
 *   class FooGetter
 *   {
 *     CoatCheck<Pledge<Foo>> mOutstandingPledges;
 *
 *   public:
 *     already_addRefed<Pledge<Foo>> GetFooAsynchronously()
 *     {
 *       nsRefPtr<Pledge<Foo>> p = new Pledge<Foo>();
 *       uint32_t requestId = mOutstandingPledges.Append(*p);
 *       sChild->SendFoo(requestId);
 *       return p.forget();
 *     }
 *
 *     void RecvFooResponse(uint32_t requestId, const Foo& fooResult)
 *     {
 *       nsRefPtr<Foo> p = mOutstandingPledges.Remove(requestId);
 *       if (p) {
 *         p->Resolve(fooResult);
 *       }
 *     }
 *   };
 *
 * This helper is currently optimized for very small sets (i.e. not optimized).
 * It is also not thread-safe as the whole point is to stay on the same thread.
 */

template<class T>
class CoatCheck
{
public:
  typedef std::pair<uint32_t, nsRefPtr<T>> Element;

  uint32_t Append(T& t)
  {
    uint32_t id = GetNextId();
    mElements.AppendElement(Element(id, nsRefPtr<T>(&t)));
    return id;
  }

  already_AddRefed<T> Remove(uint32_t aId)
  {
    for (auto& element : mElements) {
      if (element.first == aId) {
        nsRefPtr<T> ref;
        ref.swap(element.second);
        mElements.RemoveElement(element);
        return ref.forget();
      }
    }
    MOZ_ASSERT_UNREACHABLE("Received id with no matching parked object!");
    return nullptr;
  }

private:
  static uint32_t GetNextId()
  {
    static uint32_t counter = 0;
    return ++counter;
  };
  nsAutoTArray<Element, 3> mElements;
};

/* media::Refcountable - Add threadsafe ref-counting to something that isn't.
 *
 * Often, reference counting is the most practical way to share an object with
 * another thread without imposing lifetime restrictions, even if there's
 * otherwise no concurrent access happening on the object.  For instance, an
 * algorithm on another thread may find it more expedient to modify a passed-in
 * object, rather than pass expensive copies back and forth.
 *
 * Lists in particular often aren't ref-countable, yet are expensive to copy,
 * e.g. nsTArray<nsRefPtr<Foo>>. Refcountable can be used to make such objects
 * (or owning smart-pointers to such objects) refcountable.
 *
 * Technical limitation: A template specialization is needed for types that take
 * a constructor. Please add below (ScopedDeletePtr covers a lot of ground though).
 */

template<typename T>
class Refcountable : public T
{
public:
  NS_INLINE_DECL_THREADSAFE_REFCOUNTING(Refcountable<T>)
private:
  ~Refcountable<T>() {}
};

template<typename T>
class Refcountable<ScopedDeletePtr<T>> : public ScopedDeletePtr<T>
{
public:
  explicit Refcountable<ScopedDeletePtr<T>>(T* aPtr) : ScopedDeletePtr<T>(aPtr) {}
  NS_INLINE_DECL_THREADSAFE_REFCOUNTING(Refcountable<T>)
private:
  ~Refcountable<ScopedDeletePtr<T>>() {}
};

} // namespace media
} // namespace mozilla

#endif // mozilla_MediaUtils_h