Memory leak in condition_variable

[zhnmju123@…]

Hi,

an internal structure within condition_variable keeps growing and so does CPU consumption, until it brings my application to a crawl.

My analysis:

The follwing change in the recent boost version is responsible: boost/thread/win32/condition_variable.hpp:

Now: (boost 1.52)

        ~entry_manager()
        {
            if (! entry->is_notified())
            {
                entry->remove_waiter();
            }
        }

Before: (boost 1.51)

      ~entry_manager()
                {
                    entry->remove_waiter();
                }

Result: basic_condition_variable::generations keeps growing endlessly; source:

class basic_condition_variable
{
    entry_ptr get_wait_entry()
    {
[...]
            entry_ptr new_entry(new list_entry(wake_sem));
[...]
};

My test code:

//producer:
    void setWorkload(const std::vector<Zstring>& newLoad) //context of main thread
    {
        {
            boost::unique_lock<boost::mutex> dummy(lockFiles);
            filesToLoad = newLoad;
        }

        conditionNewFiles.notify_one();
    }

//consumer
    Zstring extractNextFile() //context of worker thread, blocking
    {
        boost::unique_lock<boost::mutex> dummy(lockFiles);

        while (filesToLoad.empty())
            conditionNewFiles.timed_wait(dummy, boost::posix_time::milliseconds(50)); //interruption point!

        Zstring fileName = filesToLoad.back();
        filesToLoad.pop_back();
        return fileName;
    }

Expectation: conditionNewFiles.timed_wait() should not increase "basic_condition_variable::generations"!

Test to verify this bug:

Using "conditionNewFiles.notify_all()" instead of "conditionNewFiles.notify_one()" let's my application behave correctly without growing CPU consumption; internal table "basic_condition_variable::generations" has one entry.

Regards, Zenju

[viboes]

Duplicate #7657 Serious performance and memory consumption hit if condition_variable methods condition notify_one or notify_all is used repeatedly