Logo AND Algorithmique Numérique Distribuée

Public GIT Repository
transparent cleanups around simcalls mechanism
[simgrid.git] / src / kernel / activity / ExecImpl.cpp
index 4b66dd2..cfdf73e 100644 (file)
@@ -13,6 +13,8 @@
 
 #include "simgrid/s4u/Host.hpp"
 
+#include <boost/range/algorithm.hpp>
+
 XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(simix_process);
 
 void simcall_HANDLER_execution_wait(smx_simcall_t simcall, simgrid::kernel::activity::ExecImpl* synchro)
@@ -20,8 +22,7 @@ void simcall_HANDLER_execution_wait(smx_simcall_t simcall, simgrid::kernel::acti
   XBT_DEBUG("Wait for execution of synchro %p, state %d", synchro, (int)synchro->state_);
 
   /* Associate this simcall to the synchro */
-  synchro->simcalls_.push_back(simcall);
-  simcall->issuer->waiting_synchro = synchro;
+  synchro->register_simcall(simcall);
 
   /* set surf's synchro */
   if (MC_is_active() || MC_record_replay_is_active()) {
@@ -37,91 +38,123 @@ void simcall_HANDLER_execution_wait(smx_simcall_t simcall, simgrid::kernel::acti
 
 void simcall_HANDLER_execution_test(smx_simcall_t simcall, simgrid::kernel::activity::ExecImpl* synchro)
 {
-  int res = (synchro->state_ != SIMIX_WAITING && synchro->state_ != SIMIX_RUNNING);
+  bool res = (synchro->state_ != SIMIX_WAITING && synchro->state_ != SIMIX_RUNNING);
   if (res) {
     synchro->simcalls_.push_back(simcall);
     synchro->finish();
   } else {
-    SIMIX_simcall_answer(simcall);
+    simcall->issuer->simcall_answer();
   }
   simcall_execution_test__set__result(simcall, res);
 }
 
-namespace simgrid {
-namespace kernel {
-namespace activity {
-
-ExecImpl::ExecImpl(const std::string& name, const std::string& tracing_category) : ActivityImpl(name)
+void simcall_HANDLER_execution_waitany_for(smx_simcall_t simcall, simgrid::kernel::activity::ExecImpl* execs[],
+                                           size_t count, double timeout)
 {
-  this->state_ = SIMIX_RUNNING;
-  this->set_category(tracing_category);
+  if (timeout < 0.0) {
+    simcall->timeout_cb = nullptr;
+  } else {
+    simcall->timeout_cb = simgrid::simix::Timer::set(SIMIX_get_clock() + timeout, [simcall, execs, count]() {
+      for (size_t i = 0; i < count; i++) {
+        // Remove the first occurence of simcall:
+        auto* exec = execs[i];
+        auto j     = boost::range::find(exec->simcalls_, simcall);
+        if (j != exec->simcalls_.end())
+          exec->simcalls_.erase(j);
+      }
+      simcall_execution_waitany_for__set__result(simcall, -1);
+      simcall->issuer->simcall_answer();
+    });
+  }
+
+  for (size_t i = 0; i < count; i++) {
+    /* associate this simcall to the the synchro */
+    auto* exec = execs[i];
+    exec->simcalls_.push_back(simcall);
 
-  XBT_DEBUG("Create exec %p", this);
+    /* see if the synchro is already finished */
+    if (exec->state_ != SIMIX_WAITING && exec->state_ != SIMIX_RUNNING) {
+      exec->finish();
+      break;
+    }
+  }
 }
 
+namespace simgrid {
+namespace kernel {
+namespace activity {
+
 ExecImpl::~ExecImpl()
 {
-  if (surf_action_)
-    surf_action_->unref();
   if (timeout_detector_)
     timeout_detector_->unref();
   XBT_DEBUG("Destroy exec %p", this);
 }
 
-ExecImpl* ExecImpl::set_host(s4u::Host* host)
+ExecImpl& ExecImpl::set_host(s4u::Host* host)
 {
-  host_ = host;
-  return this;
+  if (not hosts_.empty())
+    hosts_.clear();
+  hosts_.push_back(host);
+  return *this;
 }
 
-ExecImpl* ExecImpl::set_timeout(double timeout)
+ExecImpl& ExecImpl::set_hosts(const std::vector<s4u::Host*>& hosts)
 {
-  if (timeout > 0 && not MC_is_active() && not MC_record_replay_is_active()) {
-    timeout_detector_ = host_->pimpl_cpu->sleep(timeout);
-    timeout_detector_->set_data(this);
-  }
-  return this;
+  hosts_ = hosts;
+  return *this;
 }
 
-ExecImpl* ExecImpl::start(double flops_amount, double priority, double bound)
+ExecImpl& ExecImpl::set_timeout(double timeout)
 {
-  if (not MC_is_active() && not MC_record_replay_is_active()) {
-    surf_action_ = host_->pimpl_cpu->execution_start(flops_amount);
-    surf_action_->set_data(this);
-    surf_action_->set_priority(priority);
-    if (bound > 0)
-      surf_action_->set_bound(bound);
+  if (timeout > 0 && not MC_is_active() && not MC_record_replay_is_active()) {
+    timeout_detector_ = hosts_.front()->pimpl_cpu->sleep(timeout);
+    timeout_detector_->set_activity(this);
   }
+  return *this;
+}
 
-  XBT_DEBUG("Create execute synchro %p: %s", this, get_cname());
-  ExecImpl::on_creation(this);
-  return this;
+ExecImpl& ExecImpl::set_flops_amount(double flops_amount)
+{
+  if (not flops_amounts_.empty())
+    flops_amounts_.clear();
+  flops_amounts_.push_back(flops_amount);
+  return *this;
 }
 
-ExecImpl* ExecImpl::start(const std::vector<s4u::Host*>& hosts, const std::vector<double>& flops_amounts,
-                          const std::vector<double>& bytes_amounts)
+ExecImpl& ExecImpl::set_flops_amounts(const std::vector<double>& flops_amounts)
 {
-  /* set surf's synchro */
-  if (not MC_is_active() && not MC_record_replay_is_active()) {
-    surf_action_ = surf_host_model->execute_parallel(hosts, flops_amounts.data(), bytes_amounts.data(), -1);
-    if (surf_action_ != nullptr) {
-      surf_action_->set_data(this);
-    }
-  }
-  XBT_DEBUG("Create parallel execute synchro %p", this);
-  ExecImpl::on_creation(this);
-  return this;
+  flops_amounts_ = flops_amounts;
+  return *this;
 }
-void ExecImpl::cancel()
+
+ExecImpl& ExecImpl::set_bytes_amounts(const std::vector<double>& bytes_amounts)
 {
-  XBT_VERB("This exec %p is canceled", this);
-  if (surf_action_ != nullptr)
-    surf_action_->cancel();
+  bytes_amounts_ = bytes_amounts;
+
+  return *this;
 }
 
-double ExecImpl::get_remaining()
+ExecImpl* ExecImpl::start()
 {
-  return surf_action_ ? surf_action_->get_remains() : 0;
+  state_ = SIMIX_RUNNING;
+  if (not MC_is_active() && not MC_record_replay_is_active()) {
+    if (hosts_.size() == 1) {
+      surf_action_ = hosts_.front()->pimpl_cpu->execution_start(flops_amounts_.front());
+      surf_action_->set_sharing_penalty(sharing_penalty_);
+      surf_action_->set_category(get_tracing_category());
+
+      if (bound_ > 0)
+        surf_action_->set_bound(bound_);
+    } else {
+      surf_action_ = surf_host_model->execute_parallel(hosts_, flops_amounts_.data(), bytes_amounts_.data(), -1);
+    }
+    surf_action_->set_activity(this);
+  }
+
+  XBT_DEBUG("Create execute synchro %p: %s", this, get_cname());
+  ExecImpl::on_creation(*this);
+  return this;
 }
 
 double ExecImpl::get_seq_remaining_ratio()
@@ -135,20 +168,21 @@ double ExecImpl::get_par_remaining_ratio()
   return (surf_action_ == nullptr) ? 0 : surf_action_->get_remains();
 }
 
-void ExecImpl::set_bound(double bound)
+ExecImpl& ExecImpl::set_bound(double bound)
 {
-  if (surf_action_)
-    surf_action_->set_bound(bound);
+  bound_ = bound;
+  return *this;
 }
-void ExecImpl::set_priority(double priority)
+
+ExecImpl& ExecImpl::set_sharing_penalty(double sharing_penalty)
 {
-  if (surf_action_)
-    surf_action_->set_priority(priority);
+  sharing_penalty_ = sharing_penalty;
+  return *this;
 }
 
 void ExecImpl::post()
 {
-  if (host_ && not host_->is_on()) { /* FIXME: handle resource failure for parallel tasks too */
+  if (hosts_.size() == 1 && not hosts_.front()->is_on()) { /* FIXME: handle resource failure for parallel tasks too */
     /* If the host running the synchro failed, notice it. This way, the asking
      * process can be killed if it runs on that host itself */
     state_ = SIMIX_FAILED;
@@ -161,12 +195,10 @@ void ExecImpl::post()
     state_ = SIMIX_DONE;
   }
 
-  on_completion(this);
+  on_completion(*this);
+
+  clean_action();
 
-  if (surf_action_) {
-    surf_action_->unref();
-    surf_action_ = nullptr;
-  }
   if (timeout_detector_) {
     timeout_detector_->unref();
     timeout_detector_ = nullptr;
@@ -182,6 +214,37 @@ void ExecImpl::finish()
   while (not simcalls_.empty()) {
     smx_simcall_t simcall = simcalls_.front();
     simcalls_.pop_front();
+
+    /* If a waitany simcall is waiting for this synchro to finish, then remove it from the other synchros in the waitany
+     * list. Afterwards, get the position of the actual synchro in the waitany list and return it as the result of the
+     * simcall */
+
+    if (simcall->call == SIMCALL_NONE) // FIXME: maybe a better way to handle this case
+      continue;                        // if process handling comm is killed
+    if (simcall->call == SIMCALL_EXECUTION_WAITANY_FOR) {
+      simgrid::kernel::activity::ExecImpl** execs = simcall_execution_waitany_for__get__execs(simcall);
+      size_t count                                = simcall_execution_waitany_for__get__count(simcall);
+
+      for (size_t i = 0; i < count; i++) {
+        // Remove the first occurence of simcall:
+        auto* exec = execs[i];
+        auto j     = boost::range::find(exec->simcalls_, simcall);
+        if (j != exec->simcalls_.end())
+          exec->simcalls_.erase(j);
+
+        if (simcall->timeout_cb) {
+          simcall->timeout_cb->remove();
+          simcall->timeout_cb = nullptr;
+        }
+      }
+
+      if (not MC_is_active() && not MC_record_replay_is_active()) {
+        ExecImpl** element = std::find(execs, execs + count, this);
+        int rank           = (element != execs + count) ? element - execs : -1;
+        simcall_execution_waitany_for__set__result(simcall, rank);
+      }
+    }
+
     switch (state_) {
 
       case SIMIX_DONE:
@@ -206,7 +269,7 @@ void ExecImpl::finish()
 
       case SIMIX_TIMEOUT:
         XBT_DEBUG("ExecImpl::finish(): execution timeouted");
-        simcall->issuer->exception_ = std::make_exception_ptr(simgrid::TimeoutError(XBT_THROW_POINT, "Timeouted"));
+        simcall->issuer->exception_ = std::make_exception_ptr(simgrid::TimeoutException(XBT_THROW_POINT, "Timeouted"));
         break;
 
       default:
@@ -214,11 +277,9 @@ void ExecImpl::finish()
     }
 
     simcall->issuer->waiting_synchro = nullptr;
-    simcall_execution_wait__set__result(simcall, state_);
-
     /* Fail the process if the host is down */
     if (simcall->issuer->get_host()->is_on())
-      SIMIX_simcall_answer(simcall);
+      simcall->issuer->simcall_answer();
     else
       simcall->issuer->context_->iwannadie = true;
   }
@@ -230,28 +291,28 @@ ActivityImpl* ExecImpl::migrate(s4u::Host* to)
     resource::Action* old_action = this->surf_action_;
     resource::Action* new_action = to->pimpl_cpu->execution_start(old_action->get_cost());
     new_action->set_remains(old_action->get_remains());
-    new_action->set_data(this);
-    new_action->set_priority(old_action->get_priority());
+    new_action->set_activity(this);
+    new_action->set_sharing_penalty(old_action->get_sharing_penalty());
 
     // FIXME: the user-defined bound seem to not be kept by LMM, that seem to overwrite it for the multi-core modeling.
     // I hope that the user did not provide any.
 
-    old_action->set_data(nullptr);
+    old_action->set_activity(nullptr);
     old_action->cancel();
     old_action->unref();
     this->surf_action_ = new_action;
   }
 
-  on_migration(this, to);
+  on_migration(*this, to);
   return this;
 }
 
 /*************
  * Callbacks *
  *************/
-xbt::signal<void(ExecImplPtr)> ExecImpl::on_creation;
-xbt::signal<void(ExecImplPtr)> ExecImpl::on_completion;
-xbt::signal<void(ExecImplPtr, s4u::Host*)> ExecImpl::on_migration;
+xbt::signal<void(ExecImpl&)> ExecImpl::on_creation;
+xbt::signal<void(ExecImpl const&)> ExecImpl::on_completion;
+xbt::signal<void(ExecImpl const&, s4u::Host*)> ExecImpl::on_migration;
 
 } // namespace activity
 } // namespace kernel