How should I deal with mutexes in movable types in C++?
By design, std::mutex is not movable nor copyable. This means that a class A holding a mutex won't receive a default move constructor.
How would I make this type A movable in a thread-safe way?
Solution 1:
Let's start with a bit of code:
class A
{
using MutexType = std::mutex;
using ReadLock = std::unique_lock<MutexType>;
using WriteLock = std::unique_lock<MutexType>;
mutable MutexType mut_;
std::string field1_;
std::string field2_;
public:
...
I've put some rather suggestive type aliases in there that we won't really take advantage of in C++11, but become much more useful in C++14. Be patient, we'll get there.
Your question boils down to:
How do I write the move constructor and move assignment operator for this class?
We'll start with the move constructor.
Move Constructor
Note that the member mutex has been made mutable. Strictly speaking this isn't necessary for the move members, but I'm assuming you also want copy members. If that is not the case, there is no need to make the mutex mutable.
When constructing A, you do not need to lock this->mut_. But you do need to lock the mut_ of the object you're constructing from (move or copy). This can be done like so:
A(A&& a)
{
WriteLock rhs_lk(a.mut_);
field1_ = std::move(a.field1_);
field2_ = std::move(a.field2_);
}
Note that we had to default construct the members of this first, and then assign them values only after a.mut_ is locked.
Move Assignment
The move assignment operator is substantially more complicated because you do not know if some other thread is accessing either the lhs or rhs of the assignment expression. And in general, you need to guard against the following scenario:
// Thread 1
x = std::move(y);
// Thread 2
y = std::move(x);
Here is the move assignment operator that correctly guards the above scenario:
A& operator=(A&& a)
{
if (this != &a)
{
WriteLock lhs_lk(mut_, std::defer_lock);
WriteLock rhs_lk(a.mut_, std::defer_lock);
std::lock(lhs_lk, rhs_lk);
field1_ = std::move(a.field1_);
field2_ = std::move(a.field2_);
}
return *this;
}
Note that one must use std::lock(m1, m2) to lock the two mutexes, instead of just locking them one after the other. If you lock them one after the other, then when two threads assign two objects in opposite order as shown above, you can get a deadlock. The point of std::lock is to avoid that deadlock.
Copy Constructor
You didn't ask about the copy members, but we might as well talk about them now (if not you, somebody will need them).
A(const A& a)
{
ReadLock rhs_lk(a.mut_);
field1_ = a.field1_;
field2_ = a.field2_;
}
The copy constructor looks much like the move constructor except the ReadLock alias is used instead of the WriteLock. Currently these both alias std::unique_lock<std::mutex> and so it doesn't really make any difference.
But in C++14, you will have the option of saying this:
using MutexType = std::shared_timed_mutex;
using ReadLock = std::shared_lock<MutexType>;
using WriteLock = std::unique_lock<MutexType>;
This may be an optimization, but not definitely. You will have to measure to determine if it is. But with this change, one can copy construct from the same rhs in multiple threads simultaneously. The C++11 solution forces you to make such threads sequential, even though the rhs isn't being modified.
Copy Assignment
For completeness, here is the copy assignment operator, which should be fairly self explanatory after reading about everything else:
A& operator=(const A& a)
{
if (this != &a)
{
WriteLock lhs_lk(mut_, std::defer_lock);
ReadLock rhs_lk(a.mut_, std::defer_lock);
std::lock(lhs_lk, rhs_lk);
field1_ = a.field1_;
field2_ = a.field2_;
}
return *this;
}
And etc.
Any other members or free functions that access A's state will also need to be protected if you expect multiple threads to be able to call them at once. For example, here's swap:
friend void swap(A& x, A& y)
{
if (&x != &y)
{
WriteLock lhs_lk(x.mut_, std::defer_lock);
WriteLock rhs_lk(y.mut_, std::defer_lock);
std::lock(lhs_lk, rhs_lk);
using std::swap;
swap(x.field1_, y.field1_);
swap(x.field2_, y.field2_);
}
}
Note that if you just depend on std::swap doing the job, the locking will be at the wrong granularity, locking and unlocking between the three moves that std::swap would internally perform.
Indeed, thinking about swap can give you insight into the API you might need to provide for a "thread-safe" A, which in general will be different from a "non-thread-safe" API, because of the "locking granularity" issue.
Also note the need to protect against "self-swap". "self-swap" should be a no-op. Without the self-check one would recursively lock the same mutex. This could also be solved without the self-check by using std::recursive_mutex for MutexType.
Update
In the comments below Yakk is pretty unhappy about having to default construct things in the copy and move constructors (and he has a point). Should you feel strongly enough about this issue, so much so that you are willing to spend memory on it, you can avoid it like so:
-
Add whatever lock types you need as data members. These members must come before the data that is being protected:
mutable MutexType mut_; ReadLock read_lock_; WriteLock write_lock_; // ... other data members ... -
And then in the constructors (e.g. the copy constructor) do this:
A(const A& a) : read_lock_(a.mut_) , field1_(a.field1_) , field2_(a.field2_) { read_lock_.unlock(); }
Oops, Yakk erased his comment before I had the chance to complete this update. But he deserves credit for pushing this issue, and getting a solution into this answer.
Update 2
And dyp came up with this good suggestion:
A(const A& a)
: A(a, ReadLock(a.mut_))
{}
private:
A(const A& a, ReadLock rhs_lk)
: field1_(a.field1_)
, field2_(a.field2_)
{}
Solution 2:
Given there doesn't seem to be a nice, clean, easy way to answer this - Anton's solution I think is correct but its definately debatable, unless a better answer comes up I would recommend putting such a class on the heap and looking after it via a std::unique_ptr:
auto a = std::make_unique<A>();
Its now a fully movable type and anyone who has a lock on the internal mutex whilst a move happens is still safe, even if its debateable whether this is a good thing to do
If you need copy semantics just use
auto a2 = std::make_shared<A>();
Solution 3:
This is an upside-down answer. Instead of embedding "this objects needs to be synchronized" as a base of the type, instead inject it under any type.
You deal with a synchronized object very differently. One big issue is you have to worry about deadlocks (locking multiple objects). It also should basically never be your "default version of an object": synchronized objects are for objects that will be in contention, and your goal should be to minimize contention between threads, not sweep it under the rug.
But synchronizing objects is still useful. Instead of inheriting from a synchronizer, we can write a class that wraps an arbitrary type in synchronization. Users have to jump through a few hoops to do operations on the object now that it is synchronized, but they are not limited to some hand-coded limited set of operations on the object. They can compose multiple operations on the object into one, or have an operation on multiple objects.
Here is a synchronized wrapper around an arbitrary type T:
template<class T>
struct synchronized {
template<class F>
auto read(F&& f) const&->std::result_of_t<F(T const&)> {
return access(std::forward<F>(f), *this);
}
template<class F>
auto read(F&& f) &&->std::result_of_t<F(T&&)> {
return access(std::forward<F>(f), std::move(*this));
}
template<class F>
auto write(F&& f)->std::result_of_t<F(T&)> {
return access(std::forward<F>(f), *this);
}
// uses `const` ness of Syncs to determine access:
template<class F, class... Syncs>
friend auto access( F&& f, Syncs&&... syncs )->
std::result_of_t< F(decltype(std::forward<Syncs>(syncs).t)...) >
{
return access2( std::index_sequence_for<Syncs...>{}, std::forward<F>(f), std::forward<Syncs>(syncs)... );
};
synchronized(synchronized const& o):t(o.read([](T const&o){return o;})){}
synchronized(synchronized && o):t(std::move(o).read([](T&&o){return std::move(o);})){}
// special member functions:
synchronized( T & o ):t(o) {}
synchronized( T const& o ):t(o) {}
synchronized( T && o ):t(std::move(o)) {}
synchronized( T const&& o ):t(std::move(o)) {}
synchronized& operator=(T const& o) {
write([&](T& t){
t=o;
});
return *this;
}
synchronized& operator=(T && o) {
write([&](T& t){
t=std::move(o);
});
return *this;
}
private:
template<class X, class S>
static auto smart_lock(S const& s) {
return std::shared_lock< std::shared_timed_mutex >(s.m, X{});
}
template<class X, class S>
static auto smart_lock(S& s) {
return std::unique_lock< std::shared_timed_mutex >(s.m, X{});
}
template<class L>
static void lock(L& lockable) {
lockable.lock();
}
template<class...Ls>
static void lock(Ls&... lockable) {
std::lock( lockable... );
}
template<size_t...Is, class F, class...Syncs>
friend auto access2( std::index_sequence<Is...>, F&&f, Syncs&&...syncs)->
std::result_of_t< F(decltype(std::forward<Syncs>(syncs).t)...) >
{
auto locks = std::make_tuple( smart_lock<std::defer_lock_t>(syncs)... );
lock( std::get<Is>(locks)... );
return std::forward<F>(f)(std::forward<Syncs>(syncs).t ...);
}
mutable std::shared_timed_mutex m;
T t;
};
template<class T>
synchronized< T > sync( T&& t ) {
return {std::forward<T>(t)};
}
C++14 and C++1z features included.
this assumes that const operations are multiple-reader safe (which is what std containers assume).
Use looks like:
synchronized<int> x = 7;
x.read([&](auto&& v){
std::cout << v << '\n';
});
for an int with synchronized access.
I'd advise against having synchronized(synchronized const&). It is rarely needed.
If you need synchronized(synchronized const&), I'd be tempted to replace T t; with std::aligned_storage, allowing manual placement construction, and do manual destruction. That allows proper lifetime management.
Barring that, we could copy the source T, then read from it:
synchronized(synchronized const& o):
t(o.read(
[](T const&o){return o;})
)
{}
synchronized(synchronized && o):
t(std::move(o).read(
[](T&&o){return std::move(o);})
)
{}
for assignment:
synchronized& operator=(synchronized const& o) {
access([](T& lhs, T const& rhs){
lhs = rhs;
}, *this, o);
return *this;
}
synchronized& operator=(synchronized && o) {
access([](T& lhs, T&& rhs){
lhs = std::move(rhs);
}, *this, std::move(o));
return *this;
}
friend void swap(synchronized& lhs, synchronized& rhs) {
access([](T& lhs, T& rhs){
using std::swap;
swap(lhs, rhs);
}, *this, o);
}
the placement and aligned storage versions are a bit messier. Most access to t would be replaced by a member function T&t() and T const&t()const, except at construction where you'd have to jump through some hoops.
By making synchronized a wrapper instead of part of the class, all we have to ensure is that the class internally respects const as being multiple-reader, and write it in a single-threaded manner.
In the rare cases we need a synchronized instance, we jump through hoops like the above.
Apologies for any typos in the above. There are probably some.
A side benefit to the above is that n-ary arbitrary operations on synchronized objects (of the same type) work together, without having to hard-code it before hand. Add in a friend declaration and n-ary synchronized objects of multiple types might work together. I might have to move access out of being an inline friend to deal with overload conficts in that case.
live example
Solution 4:
First of all, there must be something wrong with your design if you want to move an object containing a mutex.
But if you decide to do it anyway, you have to create a new mutex in move constructor, that is e.g:
// movable
struct B{};
class A {
B b;
std::mutex m;
public:
A(A&& a)
: b(std::move(a.b))
// m is default-initialized.
{
}
};
This is thread-safe, because the move constructor can safely assume that its argument isn't used anywhere else, so the locking of the argument isn't required.