Merge #18710: Add local thread pool to CCheckQueue

bb6fcc7 refactor: Drop boost::thread stuff in CCheckQueue (Hennadii Stepanov)
6784ac4 bench: Use CCheckQueue local thread pool (Hennadii Stepanov)
dba3069 test: Use CCheckQueue local thread pool (Hennadii Stepanov)
0151177 Add local thread pool to CCheckQueue (Hennadii Stepanov)
0ef9386 refactor: Use member initializers in CCheckQueue (Hennadii Stepanov)

Pull request description:

  This PR:
  - gets rid of `boost::thread_group` in the `CCheckQueue` class
  - allows thread safety annotation usage in the `CCheckQueue` class
  - is alternative to #14464 (bitcoin/bitcoin#18710 (comment), bitcoin/bitcoin#18710 (comment))

  Also, with this PR (I hope) it could be easier to resurrect a bunch of brilliant ideas from #9938.

  Related: #17307

ACKs for top commit:
  laanwj:
    Code review ACK bb6fcc7
  LarryRuane:
    ACK bb6fcc7
  jonatack:
    Code review ACK bb6fcc7 and verified rebase to master builds cleanly with unit/functional tests green

Tree-SHA512: fddeb720d5a391b48bb4c6fa58ed34ccc3f57862fdb8e641745c021841c8340e35c5126338271446cbd98f40bd5484f27926aa6c3e76fa478ba1efafe72e73c1

Author: laanwj

src/bench/checkqueue.cpp

@@ -10,8 +10,6 @@
 #include <random.h>
 #include <util/system.h>
 
-#include <boost/thread/thread.hpp>
-
 #include <vector>
 
 static const size_t BATCHES = 101;
@@ -44,12 +42,9 @@ static void CCheckQueueSpeedPrevectorJob(benchmark::Bench& bench)
         void swap(PrevectorJob& x){p.swap(x.p);};
     };
     CCheckQueue<PrevectorJob> queue {QUEUE_BATCH_SIZE};
-    boost::thread_group tg;
     // The main thread should be counted to prevent thread oversubscription, and
     // to decrease the variance of benchmark results.
-    for (auto x = 0; x < GetNumCores() - 1; ++x) {
-       tg.create_thread([&]{queue.Thread();});
-    }
+    queue.StartWorkerThreads(GetNumCores() - 1);
 
     // create all the data once, then submit copies in the benchmark.
     FastRandomContext insecure_rand(true);
@@ -70,8 +65,7 @@ static void CCheckQueueSpeedPrevectorJob(benchmark::Bench& bench)
         // it is done explicitly here for clarity
         control.Wait();
     });
-    tg.interrupt_all();
-    tg.join_all();
+    queue.StopWorkerThreads();
     ECC_Stop();
 }
 BENCHMARK(CCheckQueueSpeedPrevectorJob);

src/checkqueue.h

@@ -6,13 +6,12 @@
 #define BITCOIN_CHECKQUEUE_H
 
 #include <sync.h>
+#include <tinyformat.h>
+#include <util/threadnames.h>
 
 #include <algorithm>
 #include <vector>
 
-#include <boost/thread/condition_variable.hpp>
-#include <boost/thread/mutex.hpp>
-
 template <typename T>
 class CCheckQueueControl;
 
@@ -31,61 +30,64 @@ class CCheckQueue
 {
 private:
     //! Mutex to protect the inner state
-    boost::mutex mutex;
+    Mutex m_mutex;
 
     //! Worker threads block on this when out of work
-    boost::condition_variable condWorker;
+    std::condition_variable m_worker_cv;
 
     //! Master thread blocks on this when out of work
-    boost::condition_variable condMaster;
+    std::condition_variable m_master_cv;
 
     //! The queue of elements to be processed.
     //! As the order of booleans doesn't matter, it is used as a LIFO (stack)
-    std::vector<T> queue;
+    std::vector<T> queue GUARDED_BY(m_mutex);
 
     //! The number of workers (including the master) that are idle.
-    int nIdle;
+    int nIdle GUARDED_BY(m_mutex){0};
 
     //! The total number of workers (including the master).
-    int nTotal;
+    int nTotal GUARDED_BY(m_mutex){0};
 
     //! The temporary evaluation result.
-    bool fAllOk;
+    bool fAllOk GUARDED_BY(m_mutex){true};
 
     /**
      * Number of verifications that haven't completed yet.
      * This includes elements that are no longer queued, but still in the
      * worker's own batches.
      */
-    unsigned int nTodo;
+    unsigned int nTodo GUARDED_BY(m_mutex){0};
 
     //! The maximum number of elements to be processed in one batch
-    unsigned int nBatchSize;
+    const unsigned int nBatchSize;
+
+    std::vector<std::thread> m_worker_threads;
+    bool m_request_stop GUARDED_BY(m_mutex){false};
 
     /** Internal function that does bulk of the verification work. */
-    bool Loop(bool fMaster = false)
+    bool Loop(bool fMaster)
     {
-        boost::condition_variable& cond = fMaster ? condMaster : condWorker;
+        std::condition_variable& cond = fMaster ? m_master_cv : m_worker_cv;
         std::vector<T> vChecks;
         vChecks.reserve(nBatchSize);
         unsigned int nNow = 0;
         bool fOk = true;
         do {
             {
-                boost::unique_lock<boost::mutex> lock(mutex);
+                WAIT_LOCK(m_mutex, lock);
                 // first do the clean-up of the previous loop run (allowing us to do it in the same critsect)
                 if (nNow) {
                     fAllOk &= fOk;
                     nTodo -= nNow;
                     if (nTodo == 0 && !fMaster)
                         // We processed the last element; inform the master it can exit and return the result
-                        condMaster.notify_one();
+                        m_master_cv.notify_one();
                 } else {
                     // first iteration
                     nTotal++;
                 }
                 // logically, the do loop starts here
-                while (queue.empty()) {
+                while (queue.empty() && !m_request_stop) {
                     if (fMaster && nTodo == 0) {
                         nTotal--;
                         bool fRet = fAllOk;
@@ -98,6 +100,10 @@ class CCheckQueue
                     cond.wait(lock); // wait
                     nIdle--;
                 }
+                if (m_request_stop) {
+                    return false;
+                }
+
                 // Decide how many work units to process now.
                 // * Do not try to do everything at once, but aim for increasingly smaller batches so
                 //   all workers finish approximately simultaneously.
@@ -106,7 +112,7 @@ class CCheckQueue
                 nNow = std::max(1U, std::min(nBatchSize, (unsigned int)queue.size() / (nTotal + nIdle + 1)));
                 vChecks.resize(nNow);
                 for (unsigned int i = 0; i < nNow; i++) {
-                    // We want the lock on the mutex to be as short as possible, so swap jobs from the global
+                    // We want the lock on the m_mutex to be as short as possible, so swap jobs from the global
                     // queue to the local batch vector instead of copying.
                     vChecks[i].swap(queue.back());
                     queue.pop_back();
@@ -124,40 +130,68 @@ class CCheckQueue
 
 public:
     //! Mutex to ensure only one concurrent CCheckQueueControl
-    boost::mutex ControlMutex;
+    Mutex m_control_mutex;
 
     //! Create a new check queue
-    explicit CCheckQueue(unsigned int nBatchSizeIn) : nIdle(0), nTotal(0), fAllOk(true), nTodo(0), nBatchSize(nBatchSizeIn) {}
+    explicit CCheckQueue(unsigned int nBatchSizeIn)
+        : nBatchSize(nBatchSizeIn)
+    {
+    }
 
-    //! Worker thread
-    void Thread()
+    //! Create a pool of new worker threads.
+    void StartWorkerThreads(const int threads_num)
     {
-        Loop();
+        {
+            LOCK(m_mutex);
+            nIdle = 0;
+            nTotal = 0;
+            fAllOk = true;
+        }
+        assert(m_worker_threads.empty());
+        for (int n = 0; n < threads_num; ++n) {
+            m_worker_threads.emplace_back([this, n]() {
+                util::ThreadRename(strprintf("scriptch.%i", n));
+                Loop(false /* worker thread */);
+            });
+        }
     }
 
     //! Wait until execution finishes, and return whether all evaluations were successful.
     bool Wait()
     {
-        return Loop(true);
+        return Loop(true /* master thread */);
     }
 
     //! Add a batch of checks to the queue
     void Add(std::vector<T>& vChecks)
     {
-        boost::unique_lock<boost::mutex> lock(mutex);
+        LOCK(m_mutex);
         for (T& check : vChecks) {
             queue.push_back(T());
             check.swap(queue.back());
         }
         nTodo += vChecks.size();
         if (vChecks.size() == 1)
-            condWorker.notify_one();
+            m_worker_cv.notify_one();
         else if (vChecks.size() > 1)
-            condWorker.notify_all();
+            m_worker_cv.notify_all();
+    }
+
+    //! Stop all of the worker threads.
+    void StopWorkerThreads()
+    {
+        WITH_LOCK(m_mutex, m_request_stop = true);
+        m_worker_cv.notify_all();
+        for (std::thread& t : m_worker_threads) {
+            t.join();
+        }
+        m_worker_threads.clear();
+        WITH_LOCK(m_mutex, m_request_stop = false);
     }
 
     ~CCheckQueue()
     {
+        assert(m_worker_threads.empty());
     }
 
 };
@@ -181,7 +215,7 @@ class CCheckQueueControl
     {
         // passed queue is supposed to be unused, or nullptr
         if (pqueue != nullptr) {
-            ENTER_CRITICAL_SECTION(pqueue->ControlMutex);
+            ENTER_CRITICAL_SECTION(pqueue->m_control_mutex);
         }
     }
 
@@ -205,7 +239,7 @@ class CCheckQueueControl
         if (!fDone)
             Wait();
         if (pqueue != nullptr) {
-            LEAVE_CRITICAL_SECTION(pqueue->ControlMutex);
+            LEAVE_CRITICAL_SECTION(pqueue->m_control_mutex);
         }
     }
 };

src/init.cpp

@@ -223,6 +223,7 @@ void Shutdown(NodeContext& node)
     if (g_load_block.joinable()) g_load_block.join();
     threadGroup.interrupt_all();
     threadGroup.join_all();
+    StopScriptCheckWorkerThreads();
 
     // After the threads that potentially access these pointers have been stopped,
     // destruct and reset all to nullptr.
@@ -1334,9 +1335,7 @@ bool AppInitMain(const util::Ref& context, NodeContext& node, interfaces::BlockA
     LogPrintf("Script verification uses %d additional threads\n", script_threads);
     if (script_threads >= 1) {
         g_parallel_script_checks = true;
-        for (int i = 0; i < script_threads; ++i) {
-            threadGroup.create_thread([i]() { return ThreadScriptCheck(i); });
-        }
+        StartScriptCheckWorkerThreads(script_threads);
     }
 
     assert(!node.scheduler);