[boost] [thread] synchronized_value: value and move semantics

Subject: [boost] [thread] synchronized_value: value and move semantics
From: Vicente J. Botet Escriba (vicente.botet_at_[hidden])
Date: 2013-02-20 13:02:52

Hi,

boost::synchronized_value (not released yet [1][2][3]) is based on [4]. 
See below part of this paper adapted to the Boost.Thread interface.

Currently boost::synchronized_value is in addition Copyable and 
Swappable. I was wondering if it is worth adding value and move 
semantics to synchronized_value. Making it EqualityComparable, 
LessThanComparable and Movable if the underlying type satisfy these 
requirements will allow to store them on a Boost.Container/C++11 container.

Do you see something completely wrong with this addition?
Has some of you had a need for this? Could you share the context?

Best,
Vicente

[1] 
https://svn.boost.org/svn/boost/trunk/boost/thread/synchronized_value.hpp
[2] 
https://svn.boost.org/svn/boost/trunk/libs/thread/example/synchronized_value.cpp
[3] 
https://svn.boost.org/svn/boost/trunk/libs/thread/example/synchronized_person.cpp
[4] "Enforcing Correct Mutex Usage with Synchronized Values" by Anthony 
Williams 
http://www.drdobbs.com/cpp/enforcing-correct-mutex-usage-with-synch/225200269

-------------

[section The Problem with Mutexes]

The key problem with protecting shared data with a mutex is that there 
is no easy way to associate the mutex with the data. It is thus 
relatively easy to accidentally write code that fails to lock the right 
mutex - or even locks the wrong mutex - and the compiler will not help you.

   std::mutex m1;
   int value1;
   std::mutex m2;
   int value2;

   int readValue1()
   {
     boost::lock_guard<boost::mutex> lk(m1);
     return value1;
   }
   int readValue2()
   {
     boost::lock_guard<boost::mutex> lk(m1); // oops: wrong mutex
     return value2;
   }

Moreover, managing the mutex lock also clutters the source code, making 
it harder to see what is really going on.

The use of synchronized_value solves both these problems - the mutex is 
intimately tied to the value, so you cannot access it without a lock, 
and yet access semantics are still straightforward. For simple accesses, 
synchronized_value behaves like a pointer-to-T; for example:

   boost::synchronized_value<std::string> value3;
   std::string readValue3()
   {
     return *value3;
   }
   void setValue3(std::string const& newVal)
   {
     *value3=newVal;
   }
   void appendToValue3(std::string const& extra)
   {
     value3->append(extra);
   }

Both forms of pointer dereference return a proxy object rather than a 
real reference, to ensure that the lock on the mutex is held across the 
assignment or method call, but this is transparent to the user.

[endsect] [/The Problem with Mutexes]

[section Beyond Simple Accesses]

The pointer-like semantics work very well for simple accesses such as 
assignment and calls to member functions. However, sometimes you need to 
perform an operation that requires multiple accesses under protection of 
the same lock, and that's what the synchronize() method provides.

By calling synchronize() you obtain an strict_lock_ptr object that holds 
a lock on the mutex protecting the data, and which can be used to access 
the protected data. The lock is held until the strict_lock_ptr object is 
destroyed, so you can safely perform multi-part operations. The 
strict_lock_ptr object also acts as a pointer-to-T, just like 
synchronized_value does, but this time the lock is already held. For 
example, the following function adds a trailing slash to a path held in 
a synchronized_value. The use of the strict_lock_ptr object ensures that 
the string hasn't changed in between the query and the update.

   void addTrailingSlashIfMissing(boost::synchronized_value<std::string> 
& path)
   {
     boost::strict_lock_ptr<std::string> u=path.synchronize();

     if(u->empty() || (*u->rbegin()!='/'))
     {
       *u+='/';
     }
   }

[endsect] [/Beyond Simple Accesses]

[section Operations Across Multiple Objects]

Though synchronized_value works very well for protecting a single object 
of type T, nothing that we've seen so far solves the problem of 
operations that require atomic access to multiple objects unless those 
objects can be combined within a single structure protected by a single 
mutex.

One way to protect access to two synchronized_value objects is to 
construct a strict_lock_ptr for each object and use those to access the 
respective protected values; for instance:

   synchronized_value<std::queue<MessageType> > q1,q2;
   void transferMessage()
   {
     strict_lock_ptr<std::queue<MessageType> > u1 = q1.synchronize();
     strict_lock_ptr<std::queue<MessageType> > u2 = q2.synchronize();

     if(!u1->empty())
     {
       u2->push_back(u1->front());
       u1->pop_front();
     }
   }

This works well in some scenarios, but not all - if the same two objects 
are updated together in different sections of code then you need to take 
care to ensure that the strict_lock_ptr objects are constructed in the 
same sequence in all cases, otherwise you have the potential for 
deadlock. This is just the same as when acquiring any two mutexes.

In order to be able to use the dead-lock free lock algorithms we need to 
use instead unique_lock_ptr, which is Lockable.

   synchronized_value<std::queue<MessageType> > q1,q2;
   void transferMessage()
   {
     unique_lock_ptr<std::queue<MessageType> > u1 = 
q1.unique_synchronize(boost::defer_lock);
     unique_lock_ptr<std::queue<MessageType> > u2 = 
q2.unique_synchronize(boost::defer_lock);
     boost::lock(u1,u2); // dead-lock free algorithm

     if(!u1->empty())
     {
       u2->push_back(u1->front());
       u1->pop_front();
     }
   }

[endsect] [/Operations Across Multiple Objects]