X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/9bb5f1864f27faca2b172bf6941a53f268847919..989042c42c15e746278c98212cd16e655d03f182:/src/smpi/mpi/smpi_request.cpp

diff --git a/src/smpi/mpi/smpi_request.cpp b/src/smpi/mpi/smpi_request.cpp
index 519b0e0abc..b782759c68 100644
--- a/src/smpi/mpi/smpi_request.cpp
+++ b/src/smpi/mpi/smpi_request.cpp
@@ -9,6 +9,8 @@
 #include "private.hpp"
 #include "simgrid/Exception.hpp"
 #include "simgrid/s4u/Exec.hpp"
+#include "simgrid/s4u/Mutex.hpp"
+#include "simgrid/s4u/ConditionVariable.hpp"
 #include "smpi_comm.hpp"
 #include "smpi_datatype.hpp"
 #include "smpi_host.hpp"
@@ -34,7 +36,7 @@ extern void (*smpi_comm_copy_data_callback)(simgrid::kernel::activity::CommImpl*
 namespace simgrid{
 namespace smpi{
 
-Request::Request(const void* buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
+Request::Request(const void* buf, int count, MPI_Datatype datatype, aid_t src, aid_t dst, int tag, MPI_Comm comm,
                  unsigned flags, MPI_Op op)
     : buf_(const_cast<void*>(buf))
     , old_type_(datatype)
@@ -61,10 +63,10 @@ Request::Request(const void* buf, int count, MPI_Datatype datatype, int src, int
     refcount_ = 1;
   else
     refcount_ = 0;
-  cancelled_ = 0;
   nbc_requests_=nullptr;
   nbc_requests_size_=0;
   init_buffer(count);
+  this->add_f();
 }
 
 void Request::ref(){
@@ -73,30 +75,29 @@ void Request::ref(){
 
 void Request::unref(MPI_Request* request)
 {
-  if((*request) != MPI_REQUEST_NULL){
-    (*request)->refcount_--;
-    if((*request)->refcount_ < 0) {
-      (*request)->print_request("wrong refcount");
-      xbt_die("Whoops, wrong refcount");
-    }
-    if((*request)->refcount_==0){
-      if ((*request)->flags_ & MPI_REQ_GENERALIZED){
-        ((*request)->generalized_funcs)->free_fn(((*request)->generalized_funcs)->extra_state);
-      }else{
-        Comm::unref((*request)->comm_);
-        Datatype::unref((*request)->old_type_);
-      }
-      if ((*request)->op_!=MPI_REPLACE && (*request)->op_!=MPI_OP_NULL)
-        Op::unref(&(*request)->op_);
+  xbt_assert(*request != MPI_REQUEST_NULL, "freeing an already free request");
 
-      (*request)->print_request("Destroying");
-      delete *request;
-      *request = MPI_REQUEST_NULL;
-    }else{
-      (*request)->print_request("Decrementing");
+  (*request)->refcount_--;
+  if ((*request)->refcount_ < 0) {
+    (*request)->print_request("wrong refcount");
+    xbt_die("Whoops, wrong refcount");
+  }
+  if ((*request)->refcount_ == 0) {
+    if ((*request)->flags_ & MPI_REQ_GENERALIZED) {
+      ((*request)->generalized_funcs)->free_fn(((*request)->generalized_funcs)->extra_state);
+    } else {
+      Comm::unref((*request)->comm_);
+      Datatype::unref((*request)->old_type_);
     }
-  }else{
-    xbt_die("freeing an already free request");
+    if ((*request)->op_ != MPI_REPLACE && (*request)->op_ != MPI_OP_NULL)
+      Op::unref(&(*request)->op_);
+
+    (*request)->print_request("Destroying");
+    F2C::free_f((*request)->f2c_id());
+    delete *request;
+    *request = MPI_REQUEST_NULL;
+  } else {
+    (*request)->print_request("Decrementing");
   }
 }
 
@@ -104,7 +105,7 @@ bool Request::match_common(MPI_Request req, MPI_Request sender, MPI_Request rece
 {
   xbt_assert(sender, "Cannot match against null sender");
   xbt_assert(receiver, "Cannot match against null receiver");
-  XBT_DEBUG("Trying to match %s of sender src %d against %d, tag %d against %d, id %d against %d",
+  XBT_DEBUG("Trying to match %s of sender src %ld against %ld, tag %d against %d, id %d against %d",
             (req == receiver ? "send" : "recv"), sender->src_, receiver->src_, sender->tag_, receiver->tag_,
             sender->comm_->id(), receiver->comm_->id());
 
@@ -118,15 +119,18 @@ bool Request::match_common(MPI_Request req, MPI_Request sender, MPI_Request rece
       receiver->real_src_ = sender->src_;
     if (receiver->tag_ == MPI_ANY_TAG)
       receiver->real_tag_ = sender->tag_;
-    if (receiver->real_size_ < sender->real_size_){
-      XBT_DEBUG("Truncating message - should not happen: receiver size : %zu < sender size : %zu", receiver->real_size_, sender->real_size_);
-      receiver->truncated_ = true;
+    if ((receiver->flags_ & MPI_REQ_PROBE) == 0 ){
+      if (receiver->real_size_ < sender->real_size_){
+        XBT_DEBUG("Truncating message - should not happen: receiver size : %zu < sender size : %zu", receiver->real_size_, sender->real_size_);
+        receiver->truncated_ = true;
+      } else if (receiver->real_size_ > sender->real_size_){
+        receiver->real_size_=sender->real_size_;
+      }
     }
     if (sender->detached_)
       receiver->detached_sender_ = sender; // tie the sender to the receiver, as it is detached and has to be freed in
                                            // the receiver
-    if (req->cancelled_ == 0)
-      req->cancelled_ = -1; // mark as uncancelable
+    req->flags_ |= MPI_REQ_MATCHED; // mark as impossible to cancel anymore
     XBT_DEBUG("match succeeded");
     return true;
   }
@@ -168,61 +172,66 @@ bool Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl*)
 
 void Request::print_request(const char* message) const
 {
-  XBT_VERB("%s  request %p  [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
-       message, this, buf_, size_, src_, dst_, tag_, flags_);
+  XBT_VERB("%s  request %p  [buf = %p, size = %zu, src = %ld, dst = %ld, tag = %d, flags = %x]", message, this, buf_,
+           size_, src_, dst_, tag_, flags_);
 }
 
 /* factories, to hide the internal flags from the caller */
 MPI_Request Request::bsend_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
 {
   return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
-                     comm->group()->actor(dst)->get_pid(), tag, comm,
+                     dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
                      MPI_REQ_PERSISTENT | MPI_REQ_SEND | MPI_REQ_PREPARED | MPI_REQ_BSEND);
 }
 
 MPI_Request Request::send_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
 {
   return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
-                     comm->group()->actor(dst)->get_pid(), tag, comm,
+                     dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
                      MPI_REQ_PERSISTENT | MPI_REQ_SEND | MPI_REQ_PREPARED);
 }
 
 MPI_Request Request::ssend_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
 {
   return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
-                     comm->group()->actor(dst)->get_pid(), tag, comm,
+                     dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
                      MPI_REQ_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
 }
 
 MPI_Request Request::isend_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
 {
   return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
-                     comm->group()->actor(dst)->get_pid(), tag, comm,
+                     dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
                      MPI_REQ_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
 }
 
-
 MPI_Request Request::rma_send_init(const void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
                                MPI_Op op)
 {
   MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
   if(op==MPI_OP_NULL){
-    request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(),
-                          comm->group()->actor(dst)->get_pid(), tag, comm,
+    request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src),
+                          dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
                           MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
   }else{
-    request      = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(),
-                          comm->group()->actor(dst)->get_pid(), tag, comm,
+    request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src),
+                          dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
                           MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED |
-                              MPI_REQ_ACCUMULATE, op);
+                              MPI_REQ_ACCUMULATE,
+                          op);
   }
   return request;
 }
 
 MPI_Request Request::recv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
 {
+  aid_t source = MPI_PROC_NULL;
+  if (src == MPI_ANY_SOURCE)
+    source = MPI_ANY_SOURCE;
+  else if (src != MPI_PROC_NULL)
+    source = comm->group()->actor(src);
   return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
-                     src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(src)->get_pid(),
+                     source,
                      simgrid::s4u::this_actor::get_pid(), tag, comm,
                      MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
 }
@@ -231,23 +240,33 @@ MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype,
                                MPI_Op op)
 {
   MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
+  aid_t source        = MPI_PROC_NULL;
+  if (src == MPI_ANY_SOURCE)
+    source = MPI_ANY_SOURCE;
+  else if (src != MPI_PROC_NULL)
+    source = comm->group()->actor(src);
   if(op==MPI_OP_NULL){
-    request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(),
-                          comm->group()->actor(dst)->get_pid(), tag, comm,
+    request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, source,
+                          dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
                           MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
   }else{
-    request      = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(),
-                          comm->group()->actor(dst)->get_pid(), tag, comm,
-                          MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED | MPI_REQ_ACCUMULATE, op);
+    request =
+        new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, source,
+                    dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
+                    MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED | MPI_REQ_ACCUMULATE, op);
   }
   return request;
 }
 
 MPI_Request Request::irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
 {
+  aid_t source = MPI_PROC_NULL;
+  if (src == MPI_ANY_SOURCE)
+    source = MPI_ANY_SOURCE;
+  else if (src != MPI_PROC_NULL)
+    source = comm->group()->actor(src);
   return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
-                     src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(src)->get_pid(),
-                     simgrid::s4u::this_actor::get_pid(), tag, comm,
+                     source, simgrid::s4u::this_actor::get_pid(), tag, comm,
                      MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
 }
 
@@ -255,9 +274,10 @@ MPI_Request Request::ibsend(const void *buf, int count, MPI_Datatype datatype, i
 {
   MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
   request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
-                        comm->group()->actor(dst)->get_pid(), tag, comm,
+                        dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
                         MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_BSEND);
-  request->start();
+  if(dst != MPI_PROC_NULL)
+    request->start();
   return request;
 }
 
@@ -265,9 +285,10 @@ MPI_Request Request::isend(const void *buf, int count, MPI_Datatype datatype, in
 {
   MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
   request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
-                        comm->group()->actor(dst)->get_pid(), tag, comm,
+                        dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
                         MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND);
-  request->start();
+  if(dst != MPI_PROC_NULL)
+    request->start();
   return request;
 }
 
@@ -275,38 +296,47 @@ MPI_Request Request::issend(const void *buf, int count, MPI_Datatype datatype, i
 {
   MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
   request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
-                        comm->group()->actor(dst)->get_pid(), tag, comm,
+                        dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
                         MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SSEND | MPI_REQ_SEND);
-  request->start();
+  if(dst != MPI_PROC_NULL)
+    request->start();
   return request;
 }
 
-
 MPI_Request Request::irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
 {
   MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
+  aid_t source        = MPI_PROC_NULL;
+  if (src == MPI_ANY_SOURCE)
+    source = MPI_ANY_SOURCE;
+  else if (src != MPI_PROC_NULL)
+    source = comm->group()->actor(src);
   request             = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
-                        src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(src)->get_pid(),
-                        simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV);
-  request->start();
+                        source, simgrid::s4u::this_actor::get_pid(), tag, comm, 
+                        MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV);
+  if(src != MPI_PROC_NULL)
+    request->start();
   return request;
 }
 
-void Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
+int Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
 {
   MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
   request = irecv(buf, count, datatype, src, tag, comm);
-  wait(&request,status);
+  int retval = wait(&request,status);
   request = nullptr;
+  return retval;
 }
 
 void Request::bsend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
 {
   MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
   request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
-                        comm->group()->actor(dst)->get_pid(), tag, comm, MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND | MPI_REQ_BSEND);
+                        dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
+                        MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND | MPI_REQ_BSEND);
 
-  request->start();
+  if(dst != MPI_PROC_NULL)
+   request->start();
   wait(&request, MPI_STATUS_IGNORE);
   request = nullptr;
 }
@@ -315,9 +345,10 @@ void Request::send(const void *buf, int count, MPI_Datatype datatype, int dst, i
 {
   MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
   request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
-                        comm->group()->actor(dst)->get_pid(), tag, comm, MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND);
-
-  request->start();
+                        dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
+                        MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND);
+  if(dst != MPI_PROC_NULL)
+   request->start();
   wait(&request, MPI_STATUS_IGNORE);
   request = nullptr;
 }
@@ -326,10 +357,11 @@ void Request::ssend(const void *buf, int count, MPI_Datatype datatype, int dst,
 {
   MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
   request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
-                        comm->group()->actor(dst)->get_pid(), tag, comm,
+                        dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
                         MPI_REQ_NON_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND);
 
-  request->start();
+  if(dst != MPI_PROC_NULL)
+   request->start();
   wait(&request,MPI_STATUS_IGNORE);
   request = nullptr;
 }
@@ -338,13 +370,20 @@ void Request::sendrecv(const void *sendbuf, int sendcount, MPI_Datatype sendtype
                        void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
                        MPI_Comm comm, MPI_Status * status)
 {
+  aid_t source = MPI_PROC_NULL;
+  if (src == MPI_ANY_SOURCE)
+    source = MPI_ANY_SOURCE;
+  else if (src != MPI_PROC_NULL)
+    source = comm->group()->actor(src);
+  aid_t destination = dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL;
+
   std::array<MPI_Request, 2> requests;
   std::array<MPI_Status, 2> stats;
-  int myid = simgrid::s4u::this_actor::get_pid();
-  if ((comm->group()->actor(dst)->get_pid() == myid) && (comm->group()->actor(src)->get_pid() == myid)) {
+  aid_t myid = simgrid::s4u::this_actor::get_pid();
+  if ((destination == myid) && (source == myid)) {
     Datatype::copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
     if (status != MPI_STATUS_IGNORE) {
-      status->MPI_SOURCE = src;
+      status->MPI_SOURCE = source;
       status->MPI_TAG    = recvtag;
       status->MPI_ERROR  = MPI_SUCCESS;
       status->count      = sendcount * sendtype->size();
@@ -431,11 +470,9 @@ void Request::start()
       mut->unlock();
   } else { /* the RECV flag was not set, so this is a send */
     const simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
-    xbt_assert(process, "Actor pid=%d is gone??", dst_);
-    int rank = src_;
-    if (TRACE_smpi_view_internals()) {
-      TRACE_smpi_send(rank, rank, dst_, tag_, size_);
-    }
+    xbt_assert(process, "Actor pid=%ld is gone??", dst_);
+    if (TRACE_smpi_view_internals())
+      TRACE_smpi_send(src_, src_, dst_, tag_, size_);
     this->print_request("New send");
 
     void* buf = buf_;
@@ -513,8 +550,9 @@ void Request::start()
     }
 
     size_t payload_size_ = size_ + 16;//MPI enveloppe size (tag+dest+communicator)
-    action_   = simcall_comm_isend(
-        simgrid::s4u::Actor::by_pid(src_)->get_impl(), mailbox->get_impl(), payload_size_, -1.0, buf, real_size_, &match_send,
+    action_              = simcall_comm_isend(
+        simgrid::kernel::actor::ActorImpl::by_pid(src_), mailbox->get_impl(), payload_size_, -1.0, buf, real_size_,
+        &match_send,
         &xbt_free_f, // how to free the userdata if a detached send fails
         process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this,
         // detach if msg size < eager/rdv switch limit
@@ -538,14 +576,14 @@ void Request::startall(int count, MPI_Request * requests)
     return;
 
   for(int i = 0; i < count; i++) {
-    requests[i]->start();
+    if(requests[i]->src_ != MPI_PROC_NULL && requests[i]->dst_ != MPI_PROC_NULL)
+      requests[i]->start();
   }
 }
 
 void Request::cancel()
 {
-  if(cancelled_!=-1)
-    cancelled_=1;
+  this->flags_ |= MPI_REQ_CANCELLED;
   if (this->action_ != nullptr)
     (boost::static_pointer_cast<simgrid::kernel::activity::CommImpl>(this->action_))->cancel();
 }
@@ -560,26 +598,14 @@ int Request::test(MPI_Request * request, MPI_Status * status, int* flag) {
 
   static int nsleeps = 1;
   int ret = MPI_SUCCESS;
-  
-  // Are we testing a request meant for non blocking collectives ?
-  // If so, test all the subrequests.
-  if ((*request)->nbc_requests_size_>0){
-    ret = testall((*request)->nbc_requests_size_, (*request)->nbc_requests_, flag, MPI_STATUSES_IGNORE);
-    if(*flag){
-      delete[] (*request)->nbc_requests_;
-      (*request)->nbc_requests_size_=0;
-      unref(request);
-    }
-    return ret;
-  }
-  
+
   if(smpi_test_sleep > 0)
     simgrid::s4u::this_actor::sleep_for(nsleeps * smpi_test_sleep);
 
   Status::empty(status);
   *flag = 1;
   if (((*request)->flags_ & (MPI_REQ_PREPARED | MPI_REQ_FINISHED)) == 0) {
-    if ((*request)->action_ != nullptr && (*request)->cancelled_ != 1){
+    if ((*request)->action_ != nullptr && ((*request)->flags_ & MPI_REQ_CANCELLED) == 0){
       try{
         *flag = simcall_comm_test((*request)->action_.get());
       } catch (const Exception&) {
@@ -763,9 +789,9 @@ void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status*
   static int nsleeps = 1;
   double speed        = s4u::this_actor::get_host()->get_speed();
   double maxrate      = smpi_cfg_iprobe_cpu_usage();
-  auto request        = new Request(nullptr, 0, MPI_CHAR,
-                             source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(source)->get_pid(),
-                             simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_PERSISTENT | MPI_REQ_RECV);
+  auto request =
+      new Request(nullptr, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(source),
+                  simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PROBE);
   if (smpi_iprobe_sleep > 0) {
     /** Compute the number of flops we will sleep **/
     s4u::this_actor::exec_init(/*nsleeps: See comment above */ nsleeps *
@@ -818,12 +844,35 @@ void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status*
   xbt_assert(request == MPI_REQUEST_NULL);
 }
 
+int Request::finish_nbc_requests(MPI_Request* request){
+  int ret = waitall((*request)->nbc_requests_size_, (*request)->nbc_requests_, MPI_STATUSES_IGNORE);
+  XBT_DEBUG("finish non blocking collective request with %d sub-requests", (*request)->nbc_requests_size_);
+  for (int i = 0; i < (*request)->nbc_requests_size_; i++) {
+    if((*request)->buf_!=nullptr && (*request)->nbc_requests_[i]!=MPI_REQUEST_NULL){//reduce case
+      void * buf=(*request)->nbc_requests_[i]->buf_;
+      if((*request)->old_type_->flags() & DT_FLAG_DERIVED)
+        buf=(*request)->nbc_requests_[i]->old_buf_;
+      if((*request)->nbc_requests_[i]->flags_ & MPI_REQ_RECV ){
+        if((*request)->op_!=MPI_OP_NULL){
+          int count=(*request)->size_/ (*request)->old_type_->size();
+          (*request)->op_->apply(buf, (*request)->buf_, &count, (*request)->old_type_);
+        }
+        smpi_free_tmp_buffer(static_cast<unsigned char*>(buf));
+      }
+    }
+    if((*request)->nbc_requests_[i]!=MPI_REQUEST_NULL)
+      Request::unref(&((*request)->nbc_requests_[i]));
+  }
+  delete[] (*request)->nbc_requests_;
+  (*request)->nbc_requests_size_=0;
+  return ret;
+}
+
 void Request::finish_wait(MPI_Request* request, MPI_Status * status)
 {
   MPI_Request req = *request;
   Status::empty(status);
-  
-  if (req->cancelled_==1){
+  if((req->flags_ & MPI_REQ_CANCELLED) != 0 && (req->flags_ & MPI_REQ_MATCHED) == 0) {
     if (status!=MPI_STATUS_IGNORE)
       status->cancelled=1;
     if(req->detached_sender_ != nullptr)
@@ -832,12 +881,23 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status)
     return;
   }
 
+  if ((*request)->flags() & MPI_REQ_NBC){
+    int ret = finish_nbc_requests(request);
+    if (ret != MPI_SUCCESS)
+      xbt_die("error when finishing non blocking collective requests");
+  }
+
   if ((req->flags_ & (MPI_REQ_PREPARED | MPI_REQ_GENERALIZED | MPI_REQ_FINISHED)) == 0) {
-    if(status != MPI_STATUS_IGNORE) {
-      int src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
-      status->MPI_SOURCE = req->comm_->group()->rank(src);
-      status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
-      status->MPI_ERROR  = req->truncated_ ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
+    if (status != MPI_STATUS_IGNORE) {
+      if (req->src_== MPI_PROC_NULL || req->dst_== MPI_PROC_NULL){
+        Status::empty(status);
+        status->MPI_SOURCE = MPI_PROC_NULL;
+      } else {
+        aid_t src          = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
+        status->MPI_SOURCE = req->comm_->group()->rank(src);
+        status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
+        status->MPI_ERROR  = req->truncated_ ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
+      }
       // this handles the case were size in receive differs from size in send
       status->count = req->real_size_;
     }
@@ -875,8 +935,8 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status)
   }
 
   if (TRACE_smpi_view_internals() && ((req->flags_ & MPI_REQ_RECV) != 0)) {
-    int rank       = simgrid::s4u::this_actor::get_pid();
-    int src_traced = (req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_);
+    aid_t rank       = simgrid::s4u::this_actor::get_pid();
+    aid_t src_traced = (req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_);
     TRACE_smpi_recv(src_traced, rank,req->tag_);
   }
   if(req->detached_sender_ != nullptr){
@@ -908,8 +968,8 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status)
       simgrid::smpi::Errhandler::unref(err);
     MC_assert(not MC_is_active()); /* Only fail in MC mode */
   }
-  unref(request);
-
+  if(req->src_ != MPI_PROC_NULL && req->dst_ != MPI_PROC_NULL)
+    unref(request);
 }
 
 int Request::wait(MPI_Request * request, MPI_Status * status)
@@ -918,29 +978,12 @@ int Request::wait(MPI_Request * request, MPI_Status * status)
   xbt_assert(*request != MPI_REQUEST_NULL);
 
   int ret=MPI_SUCCESS;
-  // Are we waiting on a request meant for non blocking collectives ?
-  // If so, wait for all the subrequests.
-  if ((*request)->nbc_requests_size_>0){
-    ret = waitall((*request)->nbc_requests_size_, (*request)->nbc_requests_, MPI_STATUSES_IGNORE);
-    for (int i = 0; i < (*request)->nbc_requests_size_; i++) {
-      if((*request)->buf_!=nullptr && (*request)->nbc_requests_[i]!=MPI_REQUEST_NULL){//reduce case
-        void * buf=(*request)->nbc_requests_[i]->buf_;
-        if((*request)->old_type_->flags() & DT_FLAG_DERIVED)
-          buf=(*request)->nbc_requests_[i]->old_buf_;
-        if((*request)->nbc_requests_[i]->flags_ & MPI_REQ_RECV ){
-          if((*request)->op_!=MPI_OP_NULL){
-            int count=(*request)->size_/ (*request)->old_type_->size();
-            (*request)->op_->apply(buf, (*request)->buf_, &count, (*request)->old_type_);
-          }
-          smpi_free_tmp_buffer(static_cast<unsigned char*>(buf));
-        }
-      }
-      if((*request)->nbc_requests_[i]!=MPI_REQUEST_NULL)
-        Request::unref(&((*request)->nbc_requests_[i]));
+
+  if((*request)->src_ == MPI_PROC_NULL || (*request)->dst_ == MPI_PROC_NULL){
+    if (status != MPI_STATUS_IGNORE) {
+      Status::empty(status);
+      status->MPI_SOURCE = MPI_PROC_NULL;
     }
-    delete[] (*request)->nbc_requests_;
-    (*request)->nbc_requests_size_=0;
-    unref(request);
     (*request)=MPI_REQUEST_NULL;
     return ret;
   }
@@ -977,6 +1020,9 @@ int Request::wait(MPI_Request * request, MPI_Status * status)
     ret = ((*request)->generalized_funcs)->query_fn(((*request)->generalized_funcs)->extra_state, mystatus);
   }
 
+  if ((*request)->truncated_)
+    ret = MPI_ERR_TRUNCATE;
+
   finish_wait(request, status); // may invalidate *request
   if (*request != MPI_REQUEST_NULL && (((*request)->flags_ & MPI_REQ_NON_PERSISTENT) != 0))
     *request = MPI_REQUEST_NULL;
@@ -1148,7 +1194,7 @@ int Request::get_status(const Request* req, int* flag, MPI_Status* status)
   *flag=0;
 
   if(req != MPI_REQUEST_NULL && req->action_ != nullptr) {
-    req->iprobe(req->src_, req->tag_, req->comm_, flag, status);
+    req->iprobe(req->comm_->group()->rank(req->src_), req->tag_, req->comm_, flag, status);
     if(*flag)
       return MPI_SUCCESS;
   }
@@ -1162,7 +1208,7 @@ int Request::get_status(const Request* req, int* flag, MPI_Status* status)
   *flag=1;
   if(req != MPI_REQUEST_NULL &&
      status != MPI_STATUS_IGNORE) {
-    int src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
+    aid_t src          = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
     status->MPI_SOURCE = req->comm_->group()->rank(src);
     status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
     status->MPI_ERROR = req->truncated_ ? MPI_ERR_TRUNCATE : MPI_SUCCESS;