X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/c587485ee2aae6b2c54ed80656cbdc4e3ba8d157..ea02d47c2ee462e84798356e4dda87b1467e22b8:/src/smpi/mpi/smpi_comm.cpp

diff --git a/src/smpi/mpi/smpi_comm.cpp b/src/smpi/mpi/smpi_comm.cpp
index 85267e8749..117d76f8b6 100644
--- a/src/smpi/mpi/smpi_comm.cpp
+++ b/src/smpi/mpi/smpi_comm.cpp
@@ -1,4 +1,4 @@
-/* Copyright (c) 2010-2019. The SimGrid Team. All rights reserved.          */
+/* Copyright (c) 2010-2021. 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. */
@@ -16,8 +16,18 @@
 
 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_comm, smpi, "Logging specific to SMPI (comm)");
 
-simgrid::smpi::Comm mpi_MPI_COMM_UNINITIALIZED;
-MPI_Comm MPI_COMM_UNINITIALIZED=&mpi_MPI_COMM_UNINITIALIZED;
+simgrid::smpi::Comm smpi_MPI_COMM_UNINITIALIZED;
+MPI_Comm MPI_COMM_UNINITIALIZED=&smpi_MPI_COMM_UNINITIALIZED;
+/**
+ * Setting MPI_COMM_WORLD to MPI_COMM_UNINITIALIZED (it's a variable)
+ * is important because the implementation of MPI_Comm checks
+ * "this == MPI_COMM_UNINITIALIZED"? If yes, it uses smpi_process()->comm_world()
+ * instead of "this".
+ * This is basically how we only have one global variable but all processes have
+ * different communicators (the one their SMPI instance uses).
+ *
+ */
+MPI_Comm MPI_COMM_WORLD = MPI_COMM_UNINITIALIZED;
 
 /* Support for cartesian topology was added, but there are 2 other types of topology, graph et dist graph. In order to
  * support them, we have to add a field SMPI_Topo_type, and replace the MPI_Topology field by an union. */
@@ -28,11 +38,14 @@ namespace smpi{
 std::unordered_map<int, smpi_key_elem> Comm::keyvals_;
 int Comm::keyval_id_=0;
 
-Comm::Comm(MPI_Group group, MPI_Topology topo, int smp, int in_id) : group_(group), topo_(topo),is_smp_comm_(smp), id_(in_id)
+Comm::Comm(MPI_Group group, MPI_Topology topo, bool smp, int in_id)
+    : group_(group), topo_(topo), is_smp_comm_(smp), id_(in_id)
 {
   errhandler_->ref();
   //First creation of comm is done before SIMIX_run, so only do comms for others
   if(in_id==MPI_UNDEFINED && smp==0 && this->rank()!=MPI_UNDEFINED ){
+    this->add_f();
+    group->c2f();
     int id;
     if(this->rank()==0){
       static int global_id_ = 0;
@@ -52,6 +65,8 @@ void Comm::destroy(Comm* comm)
     Comm::destroy(smpi_process()->comm_world());
     return;
   }
+  if(comm != MPI_COMM_WORLD)
+    comm->mark_as_deleted();
   Comm::unref(comm);
 }
 
@@ -60,7 +75,7 @@ int Comm::dup(MPI_Comm* newcomm){
     // we need to switch as the called function may silently touch global variables
     smpi_switch_data_segment(s4u::Actor::self());
   }
-  MPI_Group cp = new  Group(this->group());
+  auto* cp     = new Group(this->group());
   (*newcomm)   = new  Comm(cp, this->topo());
   int ret      = MPI_SUCCESS;
 
@@ -98,7 +113,10 @@ int Comm::dup(MPI_Comm* newcomm){
   if(info_!=MPI_INFO_NULL)
     (*newcomm)->info_ = new simgrid::smpi::Info(info_);
   //duplicate errhandler
-  (*newcomm)->set_errhandler(errhandler_);
+  if (errhandlers_ != nullptr)//MPI_COMM_WORLD, only grab our own
+    (*newcomm)->set_errhandler(errhandlers_[this->rank()]);
+  else
+    (*newcomm)->set_errhandler(errhandler_);
   return ret;
 }
 
@@ -124,26 +142,26 @@ MPI_Group Comm::group()
   return group_;
 }
 
-int Comm::size()
+int Comm::size() const
 {
   if (this == MPI_COMM_UNINITIALIZED)
     return smpi_process()->comm_world()->size();
   return group_->size();
 }
 
-int Comm::rank()
+int Comm::rank() const
 {
   if (this == MPI_COMM_UNINITIALIZED)
     return smpi_process()->comm_world()->rank();
   return group_->rank(s4u::Actor::self());
 }
 
-int Comm::id()
+int Comm::id() const
 {
   return id_;
 }
 
-void Comm::get_name (char* name, int* len)
+void Comm::get_name(char* name, int* len) const
 {
   if (this == MPI_COMM_UNINITIALIZED){
     smpi_process()->comm_world()->get_name(name, len);
@@ -178,43 +196,50 @@ void Comm::set_leaders_comm(MPI_Comm leaders){
   leaders_comm_=leaders;
 }
 
-int* Comm::get_non_uniform_map(){
+int* Comm::get_non_uniform_map() const
+{
   if (this == MPI_COMM_UNINITIALIZED)
     return smpi_process()->comm_world()->get_non_uniform_map();
   return non_uniform_map_;
 }
 
-int* Comm::get_leaders_map(){
+int* Comm::get_leaders_map() const
+{
   if (this == MPI_COMM_UNINITIALIZED)
     return smpi_process()->comm_world()->get_leaders_map();
   return leaders_map_;
 }
 
-MPI_Comm Comm::get_leaders_comm(){
+MPI_Comm Comm::get_leaders_comm() const
+{
   if (this == MPI_COMM_UNINITIALIZED)
     return smpi_process()->comm_world()->get_leaders_comm();
   return leaders_comm_;
 }
 
-MPI_Comm Comm::get_intra_comm(){
+MPI_Comm Comm::get_intra_comm() const
+{
   if (this == MPI_COMM_UNINITIALIZED || this==MPI_COMM_WORLD)
     return smpi_process()->comm_intra();
   else return intra_comm_;
 }
 
-int Comm::is_uniform(){
+bool Comm::is_uniform() const
+{
   if (this == MPI_COMM_UNINITIALIZED)
     return smpi_process()->comm_world()->is_uniform();
-  return is_uniform_;
+  return is_uniform_ != 0;
 }
 
-int Comm::is_blocked(){
+bool Comm::is_blocked() const
+{
   if (this == MPI_COMM_UNINITIALIZED)
     return smpi_process()->comm_world()->is_blocked();
-  return is_blocked_;
+  return is_blocked_ != 0;
 }
 
-int Comm::is_smp_comm(){
+bool Comm::is_smp_comm() const
+{
   if (this == MPI_COMM_UNINITIALIZED)
     return smpi_process()->comm_world()->is_smp_comm();
   return is_smp_comm_;
@@ -225,7 +250,6 @@ MPI_Comm Comm::split(int color, int key)
   if (this == MPI_COMM_UNINITIALIZED)
     return smpi_process()->comm_world()->split(color, key);
   int system_tag = -123;
-  int* recvbuf;
 
   MPI_Group group_root = nullptr;
   MPI_Group group_out  = nullptr;
@@ -233,19 +257,14 @@ MPI_Comm Comm::split(int color, int key)
   int myrank           = this->rank();
   int size             = this->size();
   /* Gather all colors and keys on rank 0 */
-  int* sendbuf = xbt_new(int, 2);
-  sendbuf[0] = color;
-  sendbuf[1] = key;
-  if (myrank == 0) {
-    recvbuf = xbt_new(int, 2 * size);
-  } else {
-    recvbuf = nullptr;
-  }
-  gather__default(sendbuf, 2, MPI_INT, recvbuf, 2, MPI_INT, 0, this);
-  xbt_free(sendbuf);
+  const std::array<int, 2> sendbuf = {{color, key}};
+  std::vector<int> recvbuf;
+  if (myrank == 0)
+    recvbuf.resize(2 * size);
+  gather__default(sendbuf.data(), 2, MPI_INT, recvbuf.data(), 2, MPI_INT, 0, this);
   /* Do the actual job */
   if (myrank == 0) {
-    MPI_Group* group_snd = xbt_new(MPI_Group, size);
+    std::vector<MPI_Group> group_snd(size);
     std::vector<std::pair<int, int>> rankmap;
     rankmap.reserve(size);
     for (int i = 0; i < size; i++) {
@@ -254,12 +273,12 @@ MPI_Comm Comm::split(int color, int key)
         for (int j = i + 1; j < size; j++) {
           if(recvbuf[2 * i] == recvbuf[2 * j]) {
             recvbuf[2 * j] = MPI_UNDEFINED;
-            rankmap.push_back({recvbuf[2 * j + 1], j});
+            rankmap.emplace_back(recvbuf[2 * j + 1], j);
           }
         }
         /* Add self in the group */
         recvbuf[2 * i] = MPI_UNDEFINED;
-        rankmap.push_back({recvbuf[2 * i + 1], i});
+        rankmap.emplace_back(recvbuf[2 * i + 1], i);
         std::sort(begin(rankmap), end(rankmap));
         group_out = new Group(rankmap.size());
         if (i == 0) {
@@ -269,7 +288,7 @@ MPI_Comm Comm::split(int color, int key)
           s4u::Actor* actor = group->actor(rankmap[j].second);
           group_out->set_mapping(actor, j);
         }
-        MPI_Request* requests = xbt_new(MPI_Request, rankmap.size());
+        std::vector<MPI_Request> requests(rankmap.size());
         int reqs              = 0;
         for (auto const& rank : rankmap) {
           if (rank.second != 0) {
@@ -281,19 +300,16 @@ MPI_Comm Comm::split(int color, int key)
         if(i != 0 && group_out != MPI_COMM_WORLD->group() && group_out != MPI_GROUP_EMPTY)
           Group::unref(group_out);
 
-        Request::waitall(reqs, requests, MPI_STATUS_IGNORE);
-        xbt_free(requests);
+        Request::waitall(reqs, requests.data(), MPI_STATUS_IGNORE);
       }
     }
-    xbt_free(recvbuf);
-    xbt_free(group_snd);
     group_out = group_root; /* exit with root's group */
   } else {
     if(color != MPI_UNDEFINED) {
       Request::recv(&group_out, 1, MPI_PTR, 0, system_tag, this, MPI_STATUS_IGNORE);
     } /* otherwise, exit with group_out == nullptr */
   }
-  return group_out!=nullptr ? new  Comm(group_out, nullptr) : MPI_COMM_NULL;
+  return group_out!=nullptr ? new  Comm(group_out, topo_) : MPI_COMM_NULL;
 }
 
 void Comm::ref(){
@@ -320,44 +336,50 @@ void Comm::unref(Comm* comm){
     return;
   }
   comm->refcount_--;
-  Group::unref(comm->group_);
 
   if(comm->refcount_==0){
+    if(simgrid::smpi::F2C::lookup() != nullptr)
+      F2C::free_f(comm->f2c_id());
     comm->cleanup_smp();
     comm->cleanup_attr<Comm>();
     if (comm->info_ != MPI_INFO_NULL)
       simgrid::smpi::Info::unref(comm->info_);
-    if (comm->errhandler_ != MPI_ERRHANDLER_NULL)
+    if(comm->errhandlers_!=nullptr){
+      for (int i=0; i<comm->size(); i++)
+	if (comm->errhandlers_[i]!=MPI_ERRHANDLER_NULL)
+          simgrid::smpi::Errhandler::unref(comm->errhandlers_[i]);
+      delete[] comm->errhandlers_;
+    } else if (comm->errhandler_ != MPI_ERRHANDLER_NULL)
       simgrid::smpi::Errhandler::unref(comm->errhandler_);
-    delete comm->topo_; // there's no use count on topos
-    delete comm;
   }
+  Group::unref(comm->group_);
+  if(comm->refcount_==0)
+    delete comm;
 }
 
 MPI_Comm Comm::find_intra_comm(int * leader){
   //get the indices of all processes sharing the same simix host
-  auto& actor_list        = sg_host_self()->pimpl_->actor_list_;
   int intra_comm_size     = 0;
   int min_index           = INT_MAX; // the minimum index will be the leader
-  for (auto& actor : actor_list) {
+  sg_host_self()->get_impl()->foreach_actor([this, &intra_comm_size, &min_index](auto& actor) {
     int index = actor.get_pid();
-    if (this->group()->rank(actor.ciface()) != MPI_UNDEFINED) { // Is this process in the current group?
+    if (this->group()->rank(actor.get_ciface()) != MPI_UNDEFINED) { // Is this process in the current group?
       intra_comm_size++;
       if (index < min_index)
         min_index = index;
     }
-  }
+  });
   XBT_DEBUG("number of processes deployed on my node : %d", intra_comm_size);
-  MPI_Group group_intra = new  Group(intra_comm_size);
+  auto* group_intra = new Group(intra_comm_size);
   int i = 0;
-  for (auto& actor : actor_list) {
-    if (this->group()->rank(actor.ciface()) != MPI_UNDEFINED) {
-      group_intra->set_mapping(actor.ciface(), i);
+  sg_host_self()->get_impl()->foreach_actor([this, group_intra, &i](auto& actor) {
+    if (this->group()->rank(actor.get_ciface()) != MPI_UNDEFINED) {
+      group_intra->set_mapping(actor.get_ciface(), i);
       i++;
     }
-  }
+  });
   *leader=min_index;
-  return new  Comm(group_intra, nullptr, 1);
+  return new Comm(group_intra, nullptr, true);
 }
 
 void Comm::init_smp(){
@@ -383,9 +405,8 @@ void Comm::init_smp(){
   // identify neighbors in comm
   MPI_Comm comm_intra = find_intra_comm(&leader);
 
-
-  int* leaders_map = new int[comm_size];
-  int* leader_list = new int[comm_size];
+  auto* leaders_map = new int[comm_size];
+  auto* leader_list = new int[comm_size];
   std::fill_n(leaders_map, comm_size, 0);
   std::fill_n(leader_list, comm_size, -1);
 
@@ -417,14 +438,14 @@ void Comm::init_smp(){
   xbt_assert(leader_group_size > 0);
   std::sort(leader_list, leader_list + leader_group_size);
 
-  MPI_Group leaders_group = new  Group(leader_group_size);
+  auto* leaders_group = new Group(leader_group_size);
 
   MPI_Comm leader_comm = MPI_COMM_NULL;
   if(MPI_COMM_WORLD!=MPI_COMM_UNINITIALIZED && this!=MPI_COMM_WORLD){
     //create leader_communicator
     for (i=0; i< leader_group_size;i++)
       leaders_group->set_mapping(s4u::Actor::by_pid(leader_list[i]).get(), i);
-    leader_comm = new  Comm(leaders_group, nullptr,1);
+    leader_comm = new Comm(leaders_group, nullptr, true);
     this->set_leaders_comm(leader_comm);
     this->set_intra_comm(comm_intra);
 
@@ -434,7 +455,7 @@ void Comm::init_smp(){
       leaders_group->set_mapping(s4u::Actor::by_pid(leader_list[i]).get(), i);
 
     if(this->get_leaders_comm()==MPI_COMM_NULL){
-      leader_comm = new  Comm(leaders_group, nullptr,1);
+      leader_comm = new Comm(leaders_group, nullptr, true);
       this->set_leaders_comm(leader_comm);
     }else{
       leader_comm=this->get_leaders_comm();
@@ -456,14 +477,14 @@ void Comm::init_smp(){
         break;
       }
     }
-    if(is_uniform==0 && this->is_uniform()!=0){
+    if (is_uniform == 0 && this->is_uniform()) {
       non_uniform_map_ = non_uniform_map;
     }else{
       xbt_free(non_uniform_map);
     }
     is_uniform_=is_uniform;
   }
-  bcast__scatter_LR_allgather(&(is_uniform_),1, MPI_INT, 0, comm_intra );
+  bcast__scatter_LR_allgather(&is_uniform_, 1, MPI_INT, 0, comm_intra);
 
   if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) {
     // we need to switch as the called function may silently touch global variables
@@ -482,7 +503,7 @@ void Comm::init_smp(){
   }
 
   int global_blocked;
-  allreduce__default(&is_blocked, &(global_blocked), 1, MPI_INT, MPI_LAND, this);
+  allreduce__default(&is_blocked, &global_blocked, 1, MPI_INT, MPI_LAND, this);
 
   if(MPI_COMM_WORLD==MPI_COMM_UNINITIALIZED || this==MPI_COMM_WORLD){
     if(this->rank()==0){
@@ -502,10 +523,9 @@ MPI_Comm Comm::f2c(int id) {
     return MPI_COMM_SELF;
   } else if(id==0){
     return MPI_COMM_WORLD;
-  } else if(F2C::f2c_lookup() != nullptr && id >= 0) {
-    char key[KEY_SIZE];
-    const auto& lookup = F2C::f2c_lookup();
-    auto comm          = lookup->find(get_key(key, id));
+  } else if (F2C::lookup() != nullptr && id >= 0) {
+    const auto& lookup = F2C::lookup();
+    auto comm          = lookup->find(id);
     return comm == lookup->end() ? MPI_COMM_NULL : static_cast<MPI_Comm>(comm->second);
   } else {
     return MPI_COMM_NULL;
@@ -513,8 +533,7 @@ MPI_Comm Comm::f2c(int id) {
 }
 
 void Comm::free_f(int id) {
-  char key[KEY_SIZE];
-  F2C::f2c_lookup()->erase(get_key(key, id));
+  F2C::lookup()->erase(id);
 }
 
 void Comm::add_rma_win(MPI_Win win){
@@ -525,7 +544,8 @@ void Comm::remove_rma_win(MPI_Win win){
   rma_wins_.remove(win);
 }
 
-void Comm::finish_rma_calls(){
+void Comm::finish_rma_calls() const
+{
   for (auto const& it : rma_wins_) {
     if(it->rank()==this->rank()){//is it ours (for MPI_COMM_WORLD)?
       int finished = it->finish_comms();
@@ -553,21 +573,39 @@ void Comm::set_info(MPI_Info info)
 
 MPI_Errhandler Comm::errhandler()
 {
-  if (errhandler_ != MPI_ERRHANDLER_NULL)
-    errhandler_->ref();
-  return errhandler_;
+  if (this != MPI_COMM_WORLD){
+    if (errhandler_ != MPI_ERRHANDLER_NULL)
+      errhandler_->ref();
+    return errhandler_;
+  } else {
+    if(errhandlers_==nullptr)
+      return MPI_ERRORS_ARE_FATAL;
+    else {
+      if(errhandlers_[this->rank()] != MPI_ERRHANDLER_NULL)
+        errhandlers_[this->rank()]->ref();
+      return errhandlers_[this->rank()];
+    }
+  }
 }
 
 void Comm::set_errhandler(MPI_Errhandler errhandler)
 {
-  if (errhandler_ != MPI_ERRHANDLER_NULL)
-    simgrid::smpi::Errhandler::unref(errhandler_);
-  errhandler_ = errhandler;
-  if (errhandler_ != MPI_ERRHANDLER_NULL)
-    errhandler_->ref();
+  if(this != MPI_COMM_WORLD){
+    if (errhandler_ != MPI_ERRHANDLER_NULL)
+      simgrid::smpi::Errhandler::unref(errhandler_);
+    errhandler_ = errhandler;
+  }else{
+    if(errhandlers_==nullptr)
+      errhandlers_= new MPI_Errhandler[this->size()]{MPI_ERRHANDLER_NULL};
+    if(errhandlers_[this->rank()] != MPI_ERRHANDLER_NULL)
+      simgrid::smpi::Errhandler::unref(errhandlers_[this->rank()]);
+    errhandlers_[this->rank()]=errhandler;
+  }
+  if (errhandler != MPI_ERRHANDLER_NULL)
+    errhandler->ref();
 }
 
-MPI_Comm Comm::split_type(int type, int /*key*/, MPI_Info)
+MPI_Comm Comm::split_type(int type, int /*key*/, const Info*)
 {
   //MPI_UNDEFINED can be given to some nodes... but we need them to still perform the smp part which is collective
   if(type != MPI_COMM_TYPE_SHARED && type != MPI_UNDEFINED){