X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/3580c1f2e5172b721449d0e79c72241bd012152a..4a4884ead9a3d55f9247a3facc7b310f1a0fe942:/src/include/xbt/parmap.hpp

diff --git a/src/include/xbt/parmap.hpp b/src/include/xbt/parmap.hpp
index 2e998ea071..18bf58d496 100644
--- a/src/include/xbt/parmap.hpp
+++ b/src/include/xbt/parmap.hpp
@@ -1,6 +1,6 @@
 /* A thread pool (C++ version).                                             */
 
-/* Copyright (c) 2004-2018 The SimGrid Team. All rights reserved.           */
+/* Copyright (c) 2004-2019 The SimGrid Team. All rights reserved.           */
 
 /* This program is free software; you can redistribute it and/or modify it
  * under the terms of the license (GNU LGPL) which comes with this package. */
@@ -10,10 +10,11 @@
 
 #include "src/internal_config.h" // HAVE_FUTEX_H
 #include "src/kernel/context/Context.hpp"
-#include "xbt/xbt_os_thread.h"
+#include "src/simix/smx_private.hpp" /* simix_global */
 
 #include <boost/optional.hpp>
 #include <condition_variable>
+#include <functional>
 #include <mutex>
 #include <thread>
 
@@ -22,6 +23,10 @@
 #include <sys/syscall.h>
 #endif
 
+#if HAVE_PTHREAD_NP_H
+#include <pthread_np.h>
+#endif
+
 XBT_LOG_EXTERNAL_CATEGORY(xbt_parmap);
 
 namespace simgrid {
@@ -38,12 +43,10 @@ public:
   Parmap(const Parmap&) = delete;
   Parmap& operator=(const Parmap&) = delete;
   ~Parmap();
-  void apply(void (*fun)(T), const std::vector<T>& data);
+  void apply(std::function<void(T)>&& fun, const std::vector<T>& data);
   boost::optional<T> next();
 
 private:
-  enum Flag { PARMAP_WORK, PARMAP_DESTROY };
-
   /**
    * @brief Thread data transmission structure
    */
@@ -95,10 +98,10 @@ private:
   public:
     explicit PosixSynchro(Parmap<T>& parmap);
     ~PosixSynchro();
-    void master_signal();
-    void master_wait();
-    void worker_signal();
-    void worker_wait(unsigned round);
+    void master_signal() override;
+    void master_wait() override;
+    void worker_signal() override;
+    void worker_wait(unsigned round) override;
 
   private:
     std::condition_variable ready_cond;
@@ -111,40 +114,40 @@ private:
   class FutexSynchro : public Synchro {
   public:
     explicit FutexSynchro(Parmap<T>& parmap) : Synchro(parmap) {}
-    void master_signal();
-    void master_wait();
-    void worker_signal();
-    void worker_wait(unsigned);
+    void master_signal() override;
+    void master_wait() override;
+    void worker_signal() override;
+    void worker_wait(unsigned) override;
 
   private:
-    static void futex_wait(unsigned* uaddr, unsigned val);
-    static void futex_wake(unsigned* uaddr, unsigned val);
+    static void futex_wait(std::atomic_uint* uaddr, unsigned val);
+    static void futex_wake(std::atomic_uint* uaddr, unsigned val);
   };
 #endif
 
   class BusyWaitSynchro : public Synchro {
   public:
     explicit BusyWaitSynchro(Parmap<T>& parmap) : Synchro(parmap) {}
-    void master_signal();
-    void master_wait();
-    void worker_signal();
-    void worker_wait(unsigned);
+    void master_signal() override;
+    void master_wait() override;
+    void worker_signal() override;
+    void worker_wait(unsigned) override;
   };
 
-  static void* worker_main(void* arg);
+  static void worker_main(ThreadData* data);
   Synchro* new_synchro(e_xbt_parmap_mode_t mode);
   void work();
 
-  Flag status;              /**< is the parmap active or being destroyed? */
-  unsigned work_round;      /**< index of the current round */
-  xbt_os_thread_t* workers; /**< worker thread handlers */
+  bool destroying;                   /**< is the parmap being destroyed? */
+  std::atomic_uint work_round;       /**< index of the current round */
+  std::vector<std::thread*> workers; /**< worker thread handlers */
   unsigned num_workers;     /**< total number of worker threads including the controller */
   Synchro* synchro;         /**< synchronization object */
 
-  unsigned thread_counter    = 0;       /**< number of workers that have done the work */
-  void (*fun)(const T)       = nullptr; /**< function to run in parallel on each element of data */
+  std::atomic_uint thread_counter{0};   /**< number of workers that have done the work */
+  std::function<void(T)> fun;           /**< function to run in parallel on each element of data */
   const std::vector<T>* data = nullptr; /**< parameters to pass to fun in parallel */
-  std::atomic<unsigned> index;          /**< index of the next element of data to pick */
+  std::atomic_uint index;               /**< index of the next element of data to pick */
 };
 
 /**
@@ -157,23 +160,38 @@ template <typename T> Parmap<T>::Parmap(unsigned num_workers, e_xbt_parmap_mode_
   XBT_CDEBUG(xbt_parmap, "Create new parmap (%u workers)", num_workers);
 
   /* Initialize the thread pool data structure */
-  this->status      = PARMAP_WORK;
+  this->destroying  = false;
   this->work_round  = 0;
-  this->workers     = new xbt_os_thread_t[num_workers];
+  this->workers.resize(num_workers);
   this->num_workers = num_workers;
   this->synchro     = new_synchro(mode);
 
   /* Create the pool of worker threads (the caller of apply() will be worker[0]) */
   this->workers[0] = nullptr;
-  unsigned int core_bind = 0;
+  XBT_ATTRIB_UNUSED unsigned int core_bind = 0;
+
   for (unsigned i = 1; i < num_workers; i++) {
     ThreadData* data = new ThreadData(*this, i);
-    this->workers[i] = xbt_os_thread_create(worker_main, data);
-    xbt_os_thread_bind(this->workers[i], core_bind);
+    this->workers[i] = new std::thread(worker_main, data);
+
+    /* Bind the worker to a core if possible */
+#if HAVE_PTHREAD_SETAFFINITY
+#if HAVE_PTHREAD_NP_H /* FreeBSD ? */
+    cpuset_t cpuset;
+    size_t size = sizeof(cpuset_t);
+#else /* Linux ? */
+    cpu_set_t cpuset;
+    size_t size = sizeof(cpu_set_t);
+#endif
+    pthread_t pthread = this->workers[i]->native_handle();
+    CPU_ZERO(&cpuset);
+    CPU_SET(core_bind, &cpuset);
+    pthread_setaffinity_np(pthread, size, &cpuset);
     if (core_bind != std::thread::hardware_concurrency() - 1)
       core_bind++;
     else
       core_bind = 0;
+#endif
   }
 }
 
@@ -182,13 +200,13 @@ template <typename T> Parmap<T>::Parmap(unsigned num_workers, e_xbt_parmap_mode_
  */
 template <typename T> Parmap<T>::~Parmap()
 {
-  status = PARMAP_DESTROY;
+  destroying = true;
   synchro->master_signal();
 
-  for (unsigned i = 1; i < num_workers; i++)
-    xbt_os_thread_join(workers[i], nullptr);
-
-  delete[] workers;
+  for (unsigned i = 1; i < num_workers; i++) {
+    workers[i]->join();
+    delete workers[i];
+  }
   delete synchro;
 }
 
@@ -197,10 +215,10 @@ template <typename T> Parmap<T>::~Parmap()
  * @param fun the function to call in parallel
  * @param data each element of this vector will be passed as an argument to fun
  */
-template <typename T> void Parmap<T>::apply(void (*fun)(T), const std::vector<T>& data)
+template <typename T> void Parmap<T>::apply(std::function<void(T)>&& fun, const std::vector<T>& data)
 {
   /* Assign resources to worker threads (we are maestro here)*/
-  this->fun   = fun;
+  this->fun   = std::move(fun);
   this->data  = &data;
   this->index = 0;
   this->synchro->master_signal(); // maestro runs futex_wake to wake all the minions (the working threads)
@@ -273,12 +291,11 @@ template <typename T> typename Parmap<T>::Synchro* Parmap<T>::new_synchro(e_xbt_
 }
 
 /** @brief Main function of a worker thread */
-template <typename T> void* Parmap<T>::worker_main(void* arg)
+template <typename T> void Parmap<T>::worker_main(ThreadData* data)
 {
-  ThreadData* data      = static_cast<ThreadData*>(arg);
   Parmap<T>& parmap     = data->parmap;
   unsigned round        = 0;
-  smx_context_t context = SIMIX_context_new(std::function<void()>(), nullptr, nullptr);
+  kernel::context::Context* context = simix_global->context_factory->create_context(std::function<void()>(), nullptr);
   kernel::context::Context::set_current(context);
 
   XBT_CDEBUG(xbt_parmap, "New worker thread created");
@@ -287,7 +304,7 @@ template <typename T> void* Parmap<T>::worker_main(void* arg)
   while (1) {
     round++; // New scheduling round
     parmap.synchro->worker_wait(round);
-    if (parmap.status == PARMAP_DESTROY)
+    if (parmap.destroying)
       break;
 
     XBT_CDEBUG(xbt_parmap, "Worker %d got a job", data->worker_id);
@@ -298,7 +315,6 @@ template <typename T> void* Parmap<T>::worker_main(void* arg)
   /* We are destroying the parmap */
   delete context;
   delete data;
-  return nullptr;
 }
 
 template <typename T> Parmap<T>::PosixSynchro::PosixSynchro(Parmap<T>& parmap) : Synchro(parmap)
@@ -347,13 +363,13 @@ template <typename T> void Parmap<T>::PosixSynchro::worker_wait(unsigned round)
 }
 
 #if HAVE_FUTEX_H
-template <typename T> inline void Parmap<T>::FutexSynchro::futex_wait(unsigned* uaddr, unsigned val)
+template <typename T> inline void Parmap<T>::FutexSynchro::futex_wait(std::atomic_uint* uaddr, unsigned val)
 {
   XBT_CVERB(xbt_parmap, "Waiting on futex %p", uaddr);
   syscall(SYS_futex, uaddr, FUTEX_WAIT_PRIVATE, val, nullptr, nullptr, 0);
 }
 
-template <typename T> inline void Parmap<T>::FutexSynchro::futex_wake(unsigned* uaddr, unsigned val)
+template <typename T> inline void Parmap<T>::FutexSynchro::futex_wake(std::atomic_uint* uaddr, unsigned val)
 {
   XBT_CVERB(xbt_parmap, "Waking futex %p", uaddr);
   syscall(SYS_futex, uaddr, FUTEX_WAKE_PRIVATE, val, nullptr, nullptr, 0);
@@ -361,25 +377,25 @@ template <typename T> inline void Parmap<T>::FutexSynchro::futex_wake(unsigned*
 
 template <typename T> void Parmap<T>::FutexSynchro::master_signal()
 {
-  __atomic_store_n(&this->parmap.thread_counter, 1, __ATOMIC_SEQ_CST);
-  __atomic_add_fetch(&this->parmap.work_round, 1, __ATOMIC_SEQ_CST);
+  this->parmap.thread_counter.store(1);
+  this->parmap.work_round.fetch_add(1);
   /* wake all workers */
   futex_wake(&this->parmap.work_round, std::numeric_limits<int>::max());
 }
 
 template <typename T> void Parmap<T>::FutexSynchro::master_wait()
 {
-  unsigned count = __atomic_load_n(&this->parmap.thread_counter, __ATOMIC_SEQ_CST);
+  unsigned count = this->parmap.thread_counter.load();
   while (count < this->parmap.num_workers) {
     /* wait for all workers to be ready */
     futex_wait(&this->parmap.thread_counter, count);
-    count = __atomic_load_n(&this->parmap.thread_counter, __ATOMIC_SEQ_CST);
+    count = this->parmap.thread_counter.load();
   }
 }
 
 template <typename T> void Parmap<T>::FutexSynchro::worker_signal()
 {
-  unsigned count = __atomic_add_fetch(&this->parmap.thread_counter, 1, __ATOMIC_SEQ_CST);
+  unsigned count = this->parmap.thread_counter.fetch_add(1) + 1;
   if (count == this->parmap.num_workers) {
     /* all workers have finished, wake the controller */
     futex_wake(&this->parmap.thread_counter, std::numeric_limits<int>::max());
@@ -388,37 +404,37 @@ template <typename T> void Parmap<T>::FutexSynchro::worker_signal()
 
 template <typename T> void Parmap<T>::FutexSynchro::worker_wait(unsigned round)
 {
-  unsigned work_round = __atomic_load_n(&this->parmap.work_round, __ATOMIC_SEQ_CST);
+  unsigned work_round = this->parmap.work_round.load();
   /* wait for more work */
   while (work_round != round) {
     futex_wait(&this->parmap.work_round, work_round);
-    work_round = __atomic_load_n(&this->parmap.work_round, __ATOMIC_SEQ_CST);
+    work_round = this->parmap.work_round.load();
   }
 }
 #endif
 
 template <typename T> void Parmap<T>::BusyWaitSynchro::master_signal()
 {
-  __atomic_store_n(&this->parmap.thread_counter, 1, __ATOMIC_SEQ_CST);
-  __atomic_add_fetch(&this->parmap.work_round, 1, __ATOMIC_SEQ_CST);
+  this->parmap.thread_counter.store(1);
+  this->parmap.work_round.fetch_add(1);
 }
 
 template <typename T> void Parmap<T>::BusyWaitSynchro::master_wait()
 {
-  while (__atomic_load_n(&this->parmap.thread_counter, __ATOMIC_SEQ_CST) < this->parmap.num_workers) {
+  while (this->parmap.thread_counter.load() < this->parmap.num_workers) {
     std::this_thread::yield();
   }
 }
 
 template <typename T> void Parmap<T>::BusyWaitSynchro::worker_signal()
 {
-  __atomic_add_fetch(&this->parmap.thread_counter, 1, __ATOMIC_SEQ_CST);
+  this->parmap.thread_counter.fetch_add(1);
 }
 
 template <typename T> void Parmap<T>::BusyWaitSynchro::worker_wait(unsigned round)
 {
   /* wait for more work */
-  while (__atomic_load_n(&this->parmap.work_round, __ATOMIC_SEQ_CST) != round) {
+  while (this->parmap.work_round.load() != round) {
     std::this_thread::yield();
   }
 }