This is a nice concept that provides the behaviour of a condition variable's notify all; parlance of pthreads, a broadcast, while in windows land, that's a wake all or wake by all.
The idea:
- Some object of type Foo is worked on by a thread
- We have many Foo object types that are being worked on by indepedent threads
- Each thread performs the same validation based on the Foo object
- Some thread ends execution if the Foo object dictates it so
- The work of one thread ending based on the state of a Foo translates to all threads working on Foo to complete in the same fashion.
Using atomics provides a smooth way to improve interprocess communication for this style of notification. While underlying implementations for condition variables on particular platforms may indeed invoke atomics, more generally, their approach is to hold onto a mutex.
While holding onto a mutex is alright, notification to all threads is invariably a loop construct, and suffers with linear response, and can be subject to pre-emption by the OS scheduler.
The downside to using atomics, is that the entire scheduled time slice is being used in an effective spin-lock on the atomic condition variable, and not inherent thread yield occurs if the condition is not met. This is really raw, down to the wire type of stuff.
The mechanism for providing the object level notification system is through the C++ static keyword. Making a static class member variable std::atomic will provide the primary mechanism.
static std::atomic_[type] _conditional;
This is very raw and low level. As we will see. There are approaches for making this more palatable for users that are averse to dealing with such constructs.
This is broken into a number of different sections.
/// \file header.h
class MyClass
{
public:
/// .. code snip ..
static std::atomic_bool accessible_;
/// .. code snip ..
};
/// \file header.cpp
std::atomic_bool MyClass::accessible_ { false };
In a reasonable world, we want to remain sequentially consistent. Where we want to improve, we remove the reigns, and ask the horse to run free.
Queries
In the processing thread's work function, we must check the class static member variable, to ensure whether the object should be accessible or not:
do_thing( Foo &obj )
{
if( Foo::accessible_.load() )
{
obj.do_stuff();
}
}
As each object is running the same code, each thread will synchronise their loads with any stores across other threads.
In the form written above, we perform the atomic load() operation with sequentially consistent memory ordering.
Setting accessible_
In which ever is the controller thread, ie, the one that dictates that for all the Foo
object types its accessible_ static member variable should be flipped, we perform the
following:
bool finish_work = Foo::accessible_.load();
while(!Foo::accessible_.compare_exchange_weak(finish_work, true));
It is important to note that in the example provided, this is performed from the main thread, however, this could just as easily be applied from any thread.. hence the weakly ordered CAS operation. There are alternatives to this notation. One could perform a hard store,
Foo::accessible_.store(true);
An exchange,
Foo::accessible_.exchange(true);
Use a strongly typed CAS..
bool finish_work = Foo::accessible_.load();
while(!Foo::accessible_.compare_exchange_strong(finish_work, true));
When we perform our CAS loops, whether they be weak or strong, there is a nice optimisation we can make.
bool finish_work = Foo::accessible_.load();
while(!Foo::accessible_.compare_exchange_weak(finish_work, true)
&& !finish_work);
It is entirely possible that another thread may well have changed the value of finish_work to that which we desire, true. Without the !finish_work, we would performing another loop before trying to write true again to finish_work.
It's time to rock on :D
Synchronize your loads and your stores.
Enable vcvars.bat amd64
cl /EHsc main.cpp
g++ -std=c++14 -Wall -Werror -pedantic -O3 -lpthread main.cpp -o lnxwin_main
g++ -std=c++14 -Wall -Werror -pedantic -O3 -lpthread main.cpp -o lnxwin_main
g++ -std=c++14 -Wall -Werror -pedantic -O3 -lpthread main.cpp -o lnxwin_main
Out of source build:
cmake .. -G "Visual Studio 2017 15 Win64" -DCMAKE_INSTALL_PREFIX=./install
cmake --build . --target all_build --config Release
cmake --build . --target install --config Release
cmake .. -DCMAKE_INSTALL_PREFIX=./install -DCMAKE_BUILD_TYPE=Release
make
make install
Once the concept is firmed up and solidifed, it's nice to explore and test your approaches. This concept is situational, and it is important to see how it fits in with what you are writing and what you are writing for.
This may seem neat, but it may not totally apply. You should consult your architecture.
http://godbolt.org
http://godbolt.org
http://godbolt.org