thread.h

Go to the documentation of this file.

/*
  Copyright 2008 Larry Gritz and the other authors and contributors.
  All Rights Reserved.

  Redistribution and use in source and binary forms, with or without
  modification, are permitted provided that the following conditions are
  met:
  * Redistributions of source code must retain the above copyright
    notice, this list of conditions and the following disclaimer.
  * Redistributions in binary form must reproduce the above copyright
    notice, this list of conditions and the following disclaimer in the
    documentation and/or other materials provided with the distribution.
  * Neither the name of the software's owners nor the names of its
    contributors may be used to endorse or promote products derived from
    this software without specific prior written permission.
  THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

  (This is the Modified BSD License)
*/

// clang-format off

/////////////////////////////////////////////////////////////////////////
/// @file   thread.h
///
/// @brief  Wrappers and utilities for multithreading.
/////////////////////////////////////////////////////////////////////////


#pragma once

#include <algorithm>
#include <atomic>
#include <chrono>
#include <functional>
#include <future>
#include <iostream>
#include <memory>
#include <mutex>
#include <thread>
#include <vector>

#include <OpenImageIO/atomic.h>
#include <OpenImageIO/dassert.h>
#include <OpenImageIO/export.h>
#include <OpenImageIO/oiioversion.h>
#include <OpenImageIO/platform.h>



// OIIO_THREAD_ALLOW_DCLP, if set to 0, prevents us from using a dodgy
// "double checked lock pattern" (DCLP).  We are very careful to construct
// it safely and correctly, and these uses improve thread performance for
// us.  But it confuses Thread Sanitizer, so this switch allows you to turn
// it off. Also set to 0 if you don't believe that we are correct in
// allowing this construct on all platforms.
#ifndef OIIO_THREAD_ALLOW_DCLP
#    define OIIO_THREAD_ALLOW_DCLP 1
#endif



// Some helpful links:
//
// Descriptions of the "new" gcc atomic intrinsics:
//    https://gcc.gnu.org/onlinedocs/gcc/_005f_005fatomic-Builtins.html
// Old gcc atomic intrinsics:
//    https://gcc.gnu.org/onlinedocs/gcc-4.4.2/gcc/Atomic-Builtins.html
// C++11 and beyond std::atomic:
//    http://en.cppreference.com/w/cpp/atomic



OIIO_NAMESPACE_BEGIN

/// Null mutex that can be substituted for a real one to test how much
/// overhead is associated with a particular mutex.
class null_mutex {
public:
    null_mutex() {}
    ~null_mutex() {}
    void lock() {}
    void unlock() {}
    void lock_shared() {}
    void unlock_shared() {}
    bool try_lock() { return true; }
};

/// Null lock that can be substituted for a real one to test how much
/// overhead is associated with a particular lock.
template<typename T> class null_lock {
public:
    null_lock(T& m) {}
};



using std::mutex;
using std::recursive_mutex;
using std::thread;
typedef std::lock_guard<mutex> lock_guard;
typedef std::lock_guard<recursive_mutex> recursive_lock_guard;



/// Yield the processor for the rest of the timeslice.
///
inline void
yield()
{
#if defined(__GNUC__)
    sched_yield();
#elif defined(_MSC_VER)
    SwitchToThread();
#else
#    error No yield on this platform.
#endif
}



// Slight pause
inline void
pause(int delay)
{
#if defined(__GNUC__) && (defined(__x86_64__) || defined(__i386__))
    for (int i = 0; i < delay; ++i)
        __asm__ __volatile__("pause;");

#elif defined(__GNUC__) && (defined(__arm__) || defined(__s390__))
    for (int i = 0; i < delay; ++i)
        __asm__ __volatile__("NOP;");

#elif defined(_MSC_VER)
    for (int i = 0; i < delay; ++i) {
#    if defined(_WIN64)
        YieldProcessor();
#    else
        _asm pause
#    endif /* _WIN64 */
    }

#else
    // No pause on this platform, just punt
    for (int i = 0; i < delay; ++i)
        ;
#endif
}



// Helper class to deliver ever longer pauses until we yield our timeslice.
class atomic_backoff {
public:
    atomic_backoff(int pausemax = 16)
        : m_count(1)
        , m_pausemax(pausemax)
    {
    }

    void operator()()
    {
        if (m_count <= m_pausemax) {
            pause(m_count);
            m_count *= 2;
        } else {
            yield();
        }
    }

private:
    int m_count;
    int m_pausemax;
};




/// A spin_mutex is semantically equivalent to a regular mutex, except
/// for the following:
///  - A spin_mutex is just 1 byte, whereas a regular mutex is quite
///    large (44 bytes for pthread).
///  - A spin_mutex is extremely fast to lock and unlock, whereas a regular
///    mutex is surprisingly expensive just to acquire a lock.
///  - A spin_mutex takes CPU while it waits, so this can be very
///    wasteful compared to a regular mutex that blocks (gives up its
///    CPU slices until it acquires the lock).
///
/// The bottom line is that mutex is the usual choice, but in cases where
/// you need to acquire locks very frequently, but only need to hold the
/// lock for a very short period of time, you may save runtime by using
/// a spin_mutex, even though it's non-blocking.
///
/// N.B. A spin_mutex is only the size of a bool.  To avoid "false
/// sharing", be careful not to put two spin_mutex objects on the same
/// cache line (within 128 bytes of each other), or the two mutexes may
/// effectively (and wastefully) lock against each other.
///
class spin_mutex {
public:
    spin_mutex(void) {}
    ~spin_mutex(void) {}

    /// Copy constructor -- initialize to unlocked.
    ///
    spin_mutex(const spin_mutex&) {}

    /// Assignment does not do anything, since lockedness should not
    /// transfer.
    const spin_mutex& operator=(const spin_mutex&) { return *this; }

    /// Acquire the lock, spin until we have it.
    ///
    void lock()
    {
        // To avoid spinning too tightly, we use the atomic_backoff to
        // provide increasingly longer pauses, and if the lock is under
        // lots of contention, eventually yield the timeslice.
        atomic_backoff backoff;

        // Try to get ownership of the lock. Though experimentation, we
        // found that OIIO_UNLIKELY makes this just a bit faster on gcc
        // x86/x86_64 systems.
        while (!OIIO_UNLIKELY(try_lock())) {
#if OIIO_THREAD_ALLOW_DCLP
            // The full try_lock() involves a test_and_set, which
            // writes memory, and that will lock the bus.  But a normal
            // read of m_locked will let us spin until the value
            // changes, without locking the bus. So it's faster to
            // check in this manner until the mutex appears to be free.
            // HOWEVER... Thread Sanitizer things this is an instance of
            // an unsafe "double checked lock pattern" (DCLP) and flags it
            // as an error. I think it's a false negative, because the
            // outer loop is still an atomic check, the inner non-atomic
            // loop only serves to delay, and can't lead to a true data
            // race. But we provide this build-time switch to, at least,
            // give a way to use tsan for other checks.
            do {
                backoff();
            } while (*(volatile bool*)&m_locked);
#else
            backoff();
#endif
        }
    }

    /// Release the lock that we hold.
    ///
    void unlock()
    {
        // Fastest way to do it is with a clear with "release" semantics
        m_locked.clear(std::memory_order_release);
    }

    /// Try to acquire the lock.  Return true if we have it, false if
    /// somebody else is holding the lock.
    bool try_lock()
    {
        return !m_locked.test_and_set(std::memory_order_acquire);
    }

    /// Helper class: scoped lock for a spin_mutex -- grabs the lock upon
    /// construction, releases the lock when it exits scope.
    class lock_guard {
    public:
        lock_guard(spin_mutex& fm)
            : m_fm(fm)
        {
            m_fm.lock();
        }
        ~lock_guard() { m_fm.unlock(); }

    private:
        lock_guard() = delete;
        lock_guard(const lock_guard& other) = delete;
        lock_guard& operator=(const lock_guard& other) = delete;
        spin_mutex& m_fm;
    };

private:
    std::atomic_flag m_locked = ATOMIC_FLAG_INIT;  // initialize to unlocked
};


typedef spin_mutex::lock_guard spin_lock;



#if 0

// OLD CODE vvvvvvvv


/// Spinning reader/writer mutex.  This is just like spin_mutex, except
/// that there are separate locking mechanisms for "writers" (exclusive
/// holders of the lock, presumably because they are modifying whatever
/// the lock is protecting) and "readers" (non-exclusive, non-modifying
/// tasks that may access the protectee simultaneously).
class spin_rw_mutex {
public:
    /// Default constructor -- initialize to unlocked.
    ///
    spin_rw_mutex (void) { m_readers = 0; }

    ~spin_rw_mutex (void) { }

    /// Copy constructor -- initialize to unlocked.
    ///
    spin_rw_mutex (const spin_rw_mutex &) { m_readers = 0; }

    /// Assignment does not do anything, since lockedness should not
    /// transfer.
    const spin_rw_mutex& operator= (const spin_rw_mutex&) { return *this; }

    /// Acquire the reader lock.
    ///
    void read_lock () {
        // Spin until there are no writers active
        m_locked.lock();
        // Register ourself as a reader
        ++m_readers;
        // Release the lock, to let other readers work
        m_locked.unlock();
    }

    /// Release the reader lock.
    ///
    void read_unlock () {
        --m_readers;  // it's atomic, no need to lock to release
    }

    /// Acquire the writer lock.
    ///
    void write_lock () {
        // Make sure no new readers (or writers) can start
        m_locked.lock();
        // Spin until the last reader is done, at which point we will be
        // the sole owners and nobody else (reader or writer) can acquire
        // the resource until we release it.
#if OIIO_THREAD_ALLOW_DCLP
        while (*(volatile int *)&m_readers > 0)
                ;
#else
        while (m_readers > 0)
                ;
#endif
    }

    /// Release the writer lock.
    ///
    void write_unlock () {
        // Let other readers or writers get the lock
        m_locked.unlock ();
    }

    /// Acquire an exclusive ("writer") lock.
    void lock () { write_lock(); }

    /// Release an exclusive ("writer") lock.
    void unlock () { write_unlock(); }

    /// Acquire a shared ("reader") lock.
    void lock_shared () { read_lock(); }

    /// Release a shared ("reader") lock.
    void unlock_shared () { read_unlock(); }

    /// Helper class: scoped read lock for a spin_rw_mutex -- grabs the
    /// read lock upon construction, releases the lock when it exits scope.
    class read_lock_guard {
    public:
        read_lock_guard (spin_rw_mutex &fm) : m_fm(fm) { m_fm.read_lock(); }
        ~read_lock_guard () { m_fm.read_unlock(); }
    private:
        read_lock_guard(); // Do not implement
        read_lock_guard(const read_lock_guard& other); // Do not implement
        read_lock_guard& operator = (const read_lock_guard& other); // Do not implement
        spin_rw_mutex & m_fm;
    };

    /// Helper class: scoped write lock for a spin_rw_mutex -- grabs the
    /// read lock upon construction, releases the lock when it exits scope.
    class write_lock_guard {
    public:
        write_lock_guard (spin_rw_mutex &fm) : m_fm(fm) { m_fm.write_lock(); }
        ~write_lock_guard () { m_fm.write_unlock(); }
    private:
        write_lock_guard(); // Do not implement
        write_lock_guard(const write_lock_guard& other); // Do not implement
        write_lock_guard& operator = (const write_lock_guard& other); // Do not implement
        spin_rw_mutex & m_fm;
    };

private:
    OIIO_CACHE_ALIGN
    spin_mutex m_locked;   // write lock
    char pad1_[OIIO_CACHE_LINE_SIZE-sizeof(spin_mutex)];
    OIIO_CACHE_ALIGN
    atomic_int m_readers;  // number of readers
    char pad2_[OIIO_CACHE_LINE_SIZE-sizeof(atomic_int)];
};


#else

// vvv New spin rw lock Oct 2017

/// Spinning reader/writer mutex.  This is just like spin_mutex, except
/// that there are separate locking mechanisms for "writers" (exclusive
/// holders of the lock, presumably because they are modifying whatever
/// the lock is protecting) and "readers" (non-exclusive, non-modifying
/// tasks that may access the protectee simultaneously).
class spin_rw_mutex {
public:
    /// Default constructor -- initialize to unlocked.
    ///
    spin_rw_mutex() {}

    ~spin_rw_mutex() {}

    // Do not allow copy or assignment.
    spin_rw_mutex(const spin_rw_mutex&) = delete;
    const spin_rw_mutex& operator=(const spin_rw_mutex&) = delete;

    /// Acquire the reader lock.
    ///
    void read_lock()
    {
        // first increase the readers, and if it turned out nobody was
        // writing, we're done. This means that acquiring a read when nobody
        // is writing is a single atomic operation.
        int oldval = m_bits.fetch_add(1, std::memory_order_acquire);
        if (!(oldval & WRITER))
            return;
        // Oops, we incremented readers but somebody was writing. Backtrack
        // by subtracting, and do things the hard way.
        int expected = (--m_bits) & NOTWRITER;

        // Do compare-and-exchange until we can increase the number of
        // readers by one and have no writers.
        if (m_bits.compare_exchange_weak(expected, expected + 1,
                                         std::memory_order_acquire))
            return;
        atomic_backoff backoff;
        do {
            backoff();
            expected = m_bits.load() & NOTWRITER;
        } while (!m_bits.compare_exchange_weak(expected, expected + 1,
                                               std::memory_order_acquire));
    }

    /// Release the reader lock.
    ///
    void read_unlock()
    {
        // Atomically reduce the number of readers.  It's at least 1,
        // and the WRITER bit should definitely not be set, so this just
        // boils down to an atomic decrement of m_bits.
        m_bits.fetch_sub(1, std::memory_order_release);
    }

    /// Acquire the writer lock.
    ///
    void write_lock()
    {
        // Do compare-and-exchange until we have just ourselves as writer
        int expected = 0;
        if (m_bits.compare_exchange_weak(expected, WRITER,
                                         std::memory_order_acquire))
            return;
        atomic_backoff backoff;
        do {
            backoff();
            expected = 0;
        } while (!m_bits.compare_exchange_weak(expected, WRITER,
                                               std::memory_order_acquire));
    }

    /// Release the writer lock.
    ///
    void write_unlock()
    {
        // Remove the writer bit
        m_bits.fetch_sub(WRITER, std::memory_order_release);
    }

    /// Helper class: scoped read lock for a spin_rw_mutex -- grabs the
    /// read lock upon construction, releases the lock when it exits scope.
    class read_lock_guard {
    public:
        read_lock_guard (spin_rw_mutex &fm) : m_fm(fm) { m_fm.read_lock(); }
        ~read_lock_guard () { m_fm.read_unlock(); }
    private:
        read_lock_guard(const read_lock_guard& other) = delete;
        read_lock_guard& operator = (const read_lock_guard& other) = delete;
        spin_rw_mutex & m_fm;
    };

    /// Helper class: scoped write lock for a spin_rw_mutex -- grabs the
    /// read lock upon construction, releases the lock when it exits scope.
    class write_lock_guard {
    public:
        write_lock_guard (spin_rw_mutex &fm) : m_fm(fm) { m_fm.write_lock(); }
        ~write_lock_guard () { m_fm.write_unlock(); }
    private:
        write_lock_guard(const write_lock_guard& other) = delete;
        write_lock_guard& operator = (const write_lock_guard& other) = delete;
        spin_rw_mutex & m_fm;
    };

private:
    // Use one word to hold the reader count, with a high bit indicating
    // that it's locked for writing.  This will only work if we have
    // fewer than 2^30 simultaneous readers.  I think that should hold
    // us for some time.
    enum { WRITER = 1<<30, NOTWRITER = WRITER-1 };
    std::atomic<int> m_bits { 0 };
};

#endif


typedef spin_rw_mutex::read_lock_guard spin_rw_read_lock;
typedef spin_rw_mutex::write_lock_guard spin_rw_write_lock;



/// Mutex pool. Sometimes, we have lots of objects that need to be
/// individually locked for thread safety, but two separate objects don't
/// need to lock against each other. If there are many more objects than
/// threads, it's wasteful for each object to contain its own mutex. So a
/// solution is to make a mutex_pool -- a collection of several mutexes.
/// Each object uses a hash to choose a consistent mutex for itself, but
/// which will be unlikely to be locked simultaneously by different object.
/// Semantically, it looks rather like an associative array of mutexes. We
/// also ensure that the mutexes are all on different cache lines, to ensure
/// that they don't exhibit false sharing. Try to choose Bins larger than
/// the expected number of threads that will be simultaneously locking
/// mutexes.
template<class Mutex, class Key, class Hash, size_t Bins = 16>
class mutex_pool {
public:
    mutex_pool() {}
    Mutex& operator[](const Key& key) { return m_mutex[m_hash(key) % Bins].m; }

private:
    // Helper type -- force cache line alignment. This should make an array
    // of these also have padding so that each individual mutex is aligned
    // to its own cache line, thus eliminating any "false sharing."
    struct AlignedMutex {
        OIIO_CACHE_ALIGN Mutex m;
    };

    AlignedMutex m_mutex[Bins];
    Hash m_hash;
};



/// Simple thread group class: lets you spawn a group of new threads,
/// then wait for them to all complete.
class thread_group {
public:
    thread_group() {}
    ~thread_group() { join_all(); }

    void add_thread(thread* t)
    {
        if (t) {
            lock_guard lock(m_mutex);
            m_threads.emplace_back(t);
        }
    }

    template<typename FUNC, typename... Args>
    thread* create_thread(FUNC func, Args&&... args)
    {
        thread* t = new thread(func, std::forward<Args>(args)...);
        add_thread(t);
        return t;
    }

    void join_all()
    {
        lock_guard lock(m_mutex);
        for (auto& t : m_threads)
            if (t->joinable())
                t->join();
    }

    size_t size() const
    {
        lock_guard lock(m_mutex);
        return m_threads.size();
    }

private:
    mutable mutex m_mutex;
    std::vector<std::unique_ptr<thread>> m_threads;
};



/// thread_pool is a persistent set of threads watching a queue to which
/// tasks can be submitted.
///
/// Call default_thread_pool() to retrieve a pointer to a single shared
/// thread_pool that will be initialized the first time it's needed, running
/// a number of threads corresponding to the number of cores on the machine.
///
/// It's possible to create other pools, but it's not something that's
/// recommended unless you really know what you're doing and are careful
/// that the sum of threads across all pools doesn't cause you to be highly
/// over-threaded. An example of when this might be useful is if you want
/// one pool of 4 threads to handle I/O without interference from a separate
/// pool of 4 (other) threads handling computation.
///
/// Submitting an asynchronous task to the queue follows the following
/// pattern:
///    /* func that takes a thread ID followed possibly by more args */
///    result_t my_func (int thread_id, Arg1 arg1, ...) { }
///    pool->push (my_func, arg1, ...);
///
/// If you just want to "fire and forget", then:
///    pool->push (func, ...args...);
/// But if you need a result, or simply need to know when the task has
/// completed, note that the push() method will return a future<result_t>
/// that you can check, like this:
///     std::future<result_t> f = pool->push (my_task);
/// And then you can
///     find out if it's done:              if (f.valid()) ...
///     wait for it to get done:            f.wait();
///     get result (waiting if necessary):  result_t r = f.get();
///
/// A common idiom is to fire a bunch of sub-tasks at the queue, and then
/// wait for them to all complete. We provide a helper class, task_set,
/// to make this easy:
///     task_set tasks (pool);
///     for (int i = 0; i < n_subtasks; ++i)
///         tasks.push (pool->push (myfunc));
///     tasks.wait ();
/// Note that the tasks.wait() is optional -- it will be called
/// automatically when the task_set exits its scope.
///
/// The task function's first argument, the thread_id, is the thread number
/// for the pool, or -1 if it's being executed by a non-pool thread (this
/// can happen in cases where the whole pool is occupied and the calling
/// thread contributes to running the work load).
///
/// Thread pool. Have fun, be safe.
///
class OIIO_API thread_pool {
public:
    /// Initialize the pool.  This implicitly calls resize() to set the
    /// number of worker threads, defaulting to a number of workers that is
    /// one less than the number of hardware cores.
    thread_pool(int nthreads = -1);
    ~thread_pool();

    /// How many threads are in the pool?
    int size() const;

    /// Sets the number of worker threads in the pool. If the pool size is
    /// 0, any tasks added to the pool will be executed immediately by the
    /// calling thread. Requesting nthreads < 0 will cause it to resize to
    /// the number of hardware cores minus one (one less, to account for the
    /// fact that the calling thread will also contribute). BEWARE! Resizing
    /// the queue should not be done while jobs are running.
    void resize(int nthreads = -1);

    /// Return the number of currently idle threads in the queue. Zero
    /// means the queue is fully engaged.
    int idle() const;

    /// Run the user's function that accepts argument int - id of the
    /// running thread. The returned value is templatized std::future, where
    /// the user can get the result and rethrow any exceptions. If the queue
    /// has no worker threads, the task will be run immediately by the
    /// calling thread.
    template<typename F> auto push(F&& f) -> std::future<decltype(f(0))>
    {
        auto pck = std::make_shared<std::packaged_task<decltype(f(0))(int)>>(
            std::forward<F>(f));
        if (size() < 1) {
            (*pck)(-1);  // No worker threads, run it with the calling thread
        } else {
            auto _f = new std::function<void(int id)>(
                [pck](int id) { (*pck)(id); });
            push_queue_and_notify(_f);
        }
        return pck->get_future();
    }

    /// Run the user's function that accepts an arbitrary set of arguments
    /// (also passed). The returned value is templatized std::future, where
    /// the user can get the result and rethrow any exceptions. If the queue
    /// has no worker threads, the task will be run immediately by the
    /// calling thread.
    template<typename F, typename... Rest>
    auto push (F && f, Rest&&... rest) ->std::future<decltype(f(0, rest...))> {
        auto pck = std::make_shared<std::packaged_task<decltype(f(0, rest...))(int)>>(
            std::bind(std::forward<F>(f), std::placeholders::_1, std::forward<Rest>(rest)...)
        );
        if (size() < 1) {
            (*pck)(-1); // No worker threads, run it with the calling thread
        } else {
            auto _f = new std::function<void(int id)>([pck](int id) {
                (*pck)(id);
            });
            push_queue_and_notify (_f);
        }
        return pck->get_future();
    }

    /// If there are any tasks on the queue, pull one off and run it (on
    /// this calling thread) and return true. Otherwise (there are no
    /// pending jobs), return false immediately. This utility is what makes
    /// it possible for non-pool threads to also run tasks from the queue
    /// when they would ordinarily be idle. The thread id of the caller
    /// should be passed.
    bool run_one_task(std::thread::id id);

    /// Return true if the calling thread is part of the thread pool. This
    /// can be used to limit a pool thread from inadvisedly adding its own
    /// subtasks to clog up the pool.
    bool this_thread_is_in_pool() const;

    /// Register a thread (not already in the thread pool itself) as working
    /// on tasks in the pool. This is used to avoid recursion.
    void register_worker(std::thread::id id);
    /// De-register a thread, saying it is no longer in the process of
    /// taking work from the thread pool.
    void deregister_worker(std::thread::id id);
    /// Is the thread in the pool or currently engaged in taking tasks from
    /// the pool?
    bool is_worker(std::thread::id id);
    bool is_worker() { return is_worker(std::this_thread::get_id()); }

    /// How many jobs are waiting to run?  (Use with caution! Can be out of
    /// date by the time you look at it.)
    size_t jobs_in_queue() const;

    /// Is the pool very busy? Meaning that there are significantly more
    /// tasks in the queue waiting to run than there are threads in the
    /// pool. It may be wise for a caller to check this before submitting
    /// tasks -- if the queue is very busy, it's probably more expedient to
    /// execute the code directly rather than add it to an oversubscribed
    /// queue.
    bool very_busy() const;

private:
    // Disallow copy construction and assignment
    thread_pool(const thread_pool&) = delete;
    thread_pool(thread_pool&&) = delete;
    thread_pool& operator=(const thread_pool&) = delete;
    thread_pool& operator=(thread_pool&&) = delete;

    // PIMPL pattern hides all the guts far away from the public API
    class Impl;
    std::unique_ptr<Impl> m_impl;

    // Utility function that helps us hide the implementation
    void push_queue_and_notify(std::function<void(int id)>* f);
};



/// Return a reference to the "default" shared thread pool. In almost all
/// ordinary circumstances, you should use this exclusively to get a
/// single shared thread pool, since creating multiple thread pools
/// could result in hilariously over-threading your application.
OIIO_API thread_pool* default_thread_pool ();



/// task_set is a group of future<void>'s from a thread_queue that you can
/// add to, and when you either call wait() or just leave the task_set's
/// scope, it will wait for all the tasks in the set to be done before
/// proceeding.
///
/// A typical idiom for using this is:
///
///    void myfunc (int id) { ... do something ... }
///
///    thread_pool* pool (default_thread_pool());
///    {
///        task_set tasks (pool);
///        // Launch a bunch of tasks into the thread pool
///        for (int i = 0; i < ntasks; ++i)
///            tasks.push (pool->push (myfunc));
///        // The following brace, by ending the scope of 'tasks', will
///        // wait for all those queue tasks to finish.
///    }
///
class OIIO_API task_set {
public:
    task_set(thread_pool* pool = nullptr)
        : m_pool(pool ? pool : default_thread_pool())
        , m_submitter_thread(std::this_thread::get_id())
    {
    }
    ~task_set() { wait(); }

    task_set(const task_set&) = delete;
    const task_set& operator=(const task_set&) = delete;

    // Return the thread id of the thread that set up this task_set and
    // submitted its tasks to the thread pool.
    std::thread::id submitter() const { return m_submitter_thread; }

    // Save a future (presumably returned by a threadpool::push() as part
    // of this task set.
    void push(std::future<void>&& f)
    {
        DASSERT(
            std::this_thread::get_id() == submitter()
            && "All tasks in a tast_set should be added by the same thread");
        m_futures.emplace_back(std::move(f));
    }

    // Wait for the given taskindex (0..n-1, where n is the number of tasks
    // submitted as part of this task_set). If block == true, fully block
    // while waiting for that task to finish. If block is false, then busy
    // wait, and opportunistically run queue tasks yourself while you are
    // waiting for the task to finish.
    void wait_for_task(size_t taskindex, bool block = false);

    // Wait for all tasks in the set to finish. If block == true, fully
    // block while waiting for the pool threads to all finish. If block is
    // false, then busy wait, and opportunistically run queue tasks yourself
    // while you are waiting for other tasks to finish.
    void wait(bool block = false);

    // Debugging sanity check, called after wait(), to ensure that all the
    // tasks were completed.
    void check_done()
    {
        const std::chrono::milliseconds wait_time(0);
        for (auto&& f : m_futures)
            ASSERT(f.wait_for(wait_time) == std::future_status::ready);
    }

private:
    thread_pool* m_pool;
    std::thread::id m_submitter_thread;
    std::vector<std::future<void>> m_futures;
};


OIIO_NAMESPACE_END