Parallelize script verification

* During block verification (when parallelism is requested), script
  check actions are stored instead of being executed immediately.
* After every processed transactions, its signature actions are
  pushed to a CScriptCheckQueue, which maintains a queue and some
  synchronization mechanism.
* Two or more threads (if enabled) start processing elements from
  this queue,
* When the block connection code is finished processing transactions,
  it joins the worker pool until the queue is empty.

As cs_main is held the entire time, and all verification must be
finished before the block continues processing, this does not reach
the best possible performance. It is a less drastic change than
some more advanced mechanisms (like doing verification out-of-band
entirely, and rolling back blocks when a failure is detected).

The -par=N flag controls the number of threads (1-16). 0 means auto,
and is the default.

src/checkqueue.h

@@ -0,0 +1,206 @@
+// Copyright (c) 2012 The Bitcoin developers
+// Distributed under the MIT/X11 software license, see the accompanying
+// file COPYING or http://www.opensource.org/licenses/mit-license.php.
+#ifndef CHECKQUEUE_H
+#define CHECKQUEUE_H
+
+#include <boost/thread/mutex.hpp>
+#include <boost/thread/locks.hpp>
+#include <boost/thread/condition_variable.hpp>
+
+#include <vector>
+#include <algorithm>
+
+template<typename T> class CCheckQueueControl;
+
+/** Queue for verifications that have to be performed.
+  * The verifications are represented by a type T, which must provide an
+  * operator(), returning a bool.
+  *
+  * One thread (the master) is assumed to push batches of verifications
+  * onto the queue, where they are processed by N-1 worker threads. When
+  * the master is done adding work, it temporarily joins the worker pool
+  * as an N'th worker, until all jobs are done.
+  */
+template<typename T> class CCheckQueue {
+private:
+    // Mutex to protect the inner state
+    boost::mutex mutex;
+
+    // Worker threads block on this when out of work
+    boost::condition_variable condWorker;
+
+    // Master thread blocks on this when out of work
+    boost::condition_variable condMaster;
+
+    // Quit method blocks on this until all workers are gone
+    boost::condition_variable condQuit;
+
+    // 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;
+
+    // The number of workers (including the master) that are idle.
+    int nIdle;
+
+    // The total number of workers (including the master).
+    int nTotal;
+
+    // The temporary evaluation result.
+    bool fAllOk;
+
+    // Number of verifications that haven't completed yet.
+    // This includes elements that are not anymore in queue, but still in
+    // worker's own batches.
+    unsigned int nTodo;
+
+    // Whether we're shutting down.
+    bool fQuit;
+
+    // The maximum number of elements to be processed in one batch
+    unsigned int nBatchSize;
+
+    // Internal function that does bulk of the verification work.
+    bool Loop(bool fMaster = false) {
+        boost::condition_variable &cond = fMaster ? condMaster : condWorker;
+        std::vector<T> vChecks;
+        vChecks.reserve(nBatchSize);
+        unsigned int nNow = 0;
+        bool fOk = true;
+        do {
+            {
+                boost::unique_lock<boost::mutex> lock(mutex);
+                // 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 he can exit and return the result
+                        condMaster.notify_one();
+                } else {
+                    // first iteration
+                    nTotal++;
+                }
+                // logically, the do loop starts here
+                while (queue.empty()) {
+                    if ((fMaster || fQuit) && nTodo == 0) {
+                        nTotal--;
+                        if (nTotal==0)
+                            condQuit.notify_one();
+                        bool fRet = fAllOk;
+                        // reset the status for new work later
+                        if (fMaster)
+                            fAllOk = true;
+                        // return the current status
+                        return fRet;
+                    }
+                    nIdle++;
+                    cond.wait(lock); // wait
+                    nIdle--;
+                }
+                // 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.
+                // * Try to account for idle jobs which will instantly start helping.
+                // * Don't do batches smaller than 1 (duh), or larger than nBatchSize.
+                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
+                     // queue to the local batch vector instead of copying.
+                     vChecks[i].swap(queue.back());
+                     queue.pop_back();
+                }
+                // Check whether we need to do work at all
+                fOk = fAllOk;
+            }
+            // execute work
+            BOOST_FOREACH(T &check, vChecks)
+                if (fOk)
+                    fOk = check();
+            vChecks.clear();
+        } while(true);
+    }
+
+public:
+    // Create a new check queue
+    CCheckQueue(unsigned int nBatchSizeIn) :
+        nIdle(0), nTotal(0), fAllOk(true), nTodo(0), fQuit(false), nBatchSize(nBatchSizeIn) {}
+
+    // Worker thread
+    void Thread() {
+        Loop();
+    }
+
+    // Wait until execution finishes, and return whether all evaluations where succesful.
+    bool Wait() {
+        return Loop(true);
+    }
+
+    // Add a batch of checks to the queue
+    void Add(std::vector<T> &vChecks) {
+        boost::unique_lock<boost::mutex> lock(mutex);
+        BOOST_FOREACH(T &check, vChecks) {
+            queue.push_back(T());
+            check.swap(queue.back());
+        }
+        nTodo += vChecks.size();
+        if (vChecks.size() == 1)
+            condWorker.notify_one();
+        else if (vChecks.size() > 1)
+            condWorker.notify_all();
+    }
+
+    // Shut the queue down
+    void Quit() {
+        boost::unique_lock<boost::mutex> lock(mutex);
+        fQuit = true;
+        // No need to wake the master, as he will quit automatically when all jobs are
+        // done.
+        condWorker.notify_all(); 
+
+        while (nTotal > 0)
+            condQuit.wait(lock);
+    }
+
+    friend class CCheckQueueControl<T>;
+};
+
+/** RAII-style controller object for a CCheckQueue that guarantees the passed
+ *  queue is finished before continuing.
+ */
+template<typename T> class CCheckQueueControl {
+private:
+    CCheckQueue<T> *pqueue;
+    bool fDone;
+
+public:
+    CCheckQueueControl(CCheckQueue<T> *pqueueIn) : pqueue(pqueueIn), fDone(false) {
+        // passed queue is supposed to be unused, or NULL
+        if (pqueue != NULL) {
+            assert(pqueue->nTotal == pqueue->nIdle);
+            assert(pqueue->nTodo == 0);
+            assert(pqueue->fAllOk == true);
+        }
+    }
+
+    bool Wait() {
+        if (pqueue == NULL)
+            return true;
+        bool fRet = pqueue->Wait();
+        fDone = true;
+        return fRet;
+    }
+
+    void Add(std::vector<T> &vChecks) {
+        if (pqueue != NULL)
+            pqueue->Add(vChecks);
+    }
+
+    ~CCheckQueueControl() {
+        if (!fDone)
+            Wait();
+    }
+};
+
+#endif