Define (and use) a callback for RawImpl::finish.

[simgrid.git] / src / kernel / activity / ExecImpl.cpp

diff --git a/src/kernel/activity/ExecImpl.cpp b/src/kernel/activity/ExecImpl.cpp
index 74129f3..0f925c9 100644 (file)
--- a/src/kernel/activity/ExecImpl.cpp
+++ b/src/kernel/activity/ExecImpl.cpp
@@ -1,11 +1,13 @@
-/* Copyright (c) 2007-2020. The SimGrid Team. All rights reserved.          */
+/* Copyright (c) 2007-2021. 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. */
 
 #include "src/kernel/activity/ExecImpl.hpp"
 #include "simgrid/Exception.hpp"
+#include "simgrid/kernel/routing/NetPoint.hpp"
 #include "simgrid/modelchecker.h"
+#include "simgrid/s4u/Exec.hpp"
 #include "src/mc/mc_replay.hpp"
 #include "src/surf/HostImpl.hpp"
 #include "src/surf/cpu_interface.hpp"
@@ -57,6 +59,7 @@ namespace activity {
 
 ExecImpl::ExecImpl()
 {
+  piface_                = new s4u::Exec(this);
   actor::ActorImpl* self = actor::ActorImpl::self();
   if (self) {
     actor_ = self;
@@ -78,7 +81,7 @@ ExecImpl& ExecImpl::set_hosts(const std::vector<s4u::Host*>& hosts)
 
 ExecImpl& ExecImpl::set_timeout(double timeout)
 {
-  if (timeout > 0 && not MC_is_active() && not MC_record_replay_is_active()) {
+  if (timeout >= 0 && not MC_is_active() && not MC_record_replay_is_active()) {
     timeout_detector_.reset(hosts_.front()->pimpl_cpu->sleep(timeout));
     timeout_detector_->set_activity(this);
   }
@@ -113,10 +116,14 @@ ExecImpl* ExecImpl::start()
       surf_action_->set_sharing_penalty(sharing_penalty_);
       surf_action_->set_category(get_tracing_category());
 
-      if (bound_ > 0)
+      if (bound_ > 0) {
         surf_action_->set_bound(bound_);
+        surf_action_->set_user_bound(bound_);
+      }
     } else {
-      surf_action_ = surf_host_model->execute_parallel(hosts_, flops_amounts_.data(), bytes_amounts_.data(), -1);
+      // FIXME[donassolo]: verify if all hosts belongs to the same netZone?
+      auto host_model = hosts_.front()->get_netpoint()->get_englobing_zone()->get_host_model();
+      surf_action_    = host_model->execute_parallel(hosts_, flops_amounts_.data(), bytes_amounts_.data(), -1);
     }
     surf_action_->set_activity(this);
   }
@@ -150,19 +157,27 @@ ExecImpl& ExecImpl::set_sharing_penalty(double sharing_penalty)
 
 void ExecImpl::post()
 {
-  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
+  xbt_assert(surf_action_ != nullptr);
+  if (std::any_of(hosts_.begin(), hosts_.end(), [](const s4u::Host* host) { return not host->is_on(); })) {
+    /* If one of the hosts running the synchro failed, notice it. This way, the asking
      * process can be killed if it runs on that host itself */
     state_ = State::FAILED;
-  } else if (surf_action_ && surf_action_->get_state() == resource::Action::State::FAILED) {
-    /* If the host running the synchro didn't fail, then the synchro was canceled */
+  } else if (surf_action_->get_state() == resource::Action::State::FAILED) {
+    /* If all the hosts are running the synchro didn't fail, then the synchro was canceled */
     state_ = State::CANCELED;
   } else if (timeout_detector_ && timeout_detector_->get_state() == resource::Action::State::FINISHED) {
-    state_ = State::TIMEOUT;
+    if (surf_action_->get_remains() > 0.0) {
+      surf_action_->set_state(resource::Action::State::FAILED);
+      state_ = State::TIMEOUT;
+    } else {
+      state_ = State::DONE;
+    }
   } else {
     state_ = State::DONE;
   }
 
+  get_iface()->set_finish_time(surf_action_->get_finish_time());
+
   clean_action();
   timeout_detector_.reset();
   if (actor_) {
@@ -183,9 +198,9 @@ void ExecImpl::finish()
      * 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) {
+    if (simcall->call_ == simix::Simcall::NONE) // FIXME: maybe a better way to handle this case
+      continue;                                 // if process handling comm is killed
+    if (simcall->call_ == simix::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);
 
@@ -208,7 +223,6 @@ void ExecImpl::finish()
         simcall_execution_waitany_for__set__result(simcall, rank);
       }
     }
-
     switch (state_) {
       case State::DONE:
         /* do nothing, synchro done */
@@ -256,9 +270,7 @@ ActivityImpl* ExecImpl::migrate(s4u::Host* to)
     new_action->set_remains(old_action->get_remains());
     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.
+    new_action->set_user_bound(old_action->get_user_bound());
 
     old_action->set_activity(nullptr);
     old_action->cancel();