X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/84402e8e2ee2a2d0bef25fdceb0a263ed8b471f6..ea02d47c2ee462e84798356e4dda87b1467e22b8:/src/smpi/mpi/smpi_request.cpp

diff --git a/src/smpi/mpi/smpi_request.cpp b/src/smpi/mpi/smpi_request.cpp
index 6a8ac60b4c..7fc8421b9b 100644
--- a/src/smpi/mpi/smpi_request.cpp
+++ b/src/smpi/mpi/smpi_request.cpp
@@ -1,4 +1,4 @@
-/* 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. */
@@ -18,6 +18,7 @@
 #include "src/smpi/include/smpi_actor.hpp"
 
 #include <algorithm>
+#include <array>
 
 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_request, smpi, "Logging specific to SMPI (request)");
 
@@ -33,26 +34,18 @@ 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, unsigned flags, MPI_Op op)
-    : buf_(const_cast<void*>(buf)), old_type_(datatype), src_(src), dst_(dst), tag_(tag), comm_(comm), flags_(flags), op_(op)
+Request::Request(const void* buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
+                 unsigned flags, MPI_Op op)
+    : buf_(const_cast<void*>(buf))
+    , old_type_(datatype)
+    , size_(datatype->size() * count)
+    , src_(src)
+    , dst_(dst)
+    , tag_(tag)
+    , comm_(comm)
+    , flags_(flags)
+    , op_(op)
 {
-  void *old_buf = nullptr;
-// FIXME Handle the case of a partial shared malloc.
-  if ((((flags & MPI_REQ_RECV) != 0) && ((flags & MPI_REQ_ACCUMULATE) != 0)) || (datatype->flags() & DT_FLAG_DERIVED)) {
-    // This part handles the problem of non-contiguous memory
-    old_buf = const_cast<void*>(buf);
-    if (count==0){
-      buf_ = nullptr;
-    }else {
-      buf_ = xbt_malloc(count*datatype->size());
-      if ((datatype->flags() & DT_FLAG_DERIVED) && ((flags & MPI_REQ_SEND) != 0)) {
-        datatype->serialize(old_buf, buf_, count);
-      }
-    }
-  }
-  // This part handles the problem of non-contiguous memory (for the unserialization at the reception)
-  old_buf_  = old_buf;
-  size_ = datatype->size() * count;
   datatype->ref();
   comm_->ref();
   if(op != MPI_REPLACE && op != MPI_OP_NULL)
@@ -68,10 +61,10 @@ Request::Request(const void* buf, int count, MPI_Datatype datatype, int src, int
     refcount_ = 1;
   else
     refcount_ = 0;
-  cancelled_ = 0;
-  generalized_funcs=nullptr;
   nbc_requests_=nullptr;
   nbc_requests_size_=0;
+  init_buffer(count);
+  this->add_f();
 }
 
 void Request::ref(){
@@ -80,31 +73,29 @@ 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);
-        delete (*request)->generalized_funcs;
-      }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");
   }
 }
 
@@ -126,34 +117,58 @@ 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_)
-      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;
   }
   return false;
 }
 
+void Request::init_buffer(int count){
+  void *old_buf = nullptr;
+// FIXME Handle the case of a partial shared malloc.
+  // This part handles the problem of non-contiguous memory (for the unserialization at the reception)
+  if ((((flags_ & MPI_REQ_RECV) != 0) && ((flags_ & MPI_REQ_ACCUMULATE) != 0)) || (old_type_->flags() & DT_FLAG_DERIVED)) {
+    // This part handles the problem of non-contiguous memory
+    old_buf = buf_;
+    if (count==0){
+      buf_ = nullptr;
+    }else {
+      buf_ = xbt_malloc(count*old_type_->size());
+      if ((old_type_->flags() & DT_FLAG_DERIVED) && ((flags_ & MPI_REQ_SEND) != 0)) {
+        old_type_->serialize(old_buf, buf_, count);
+      }
+    }
+  }
+  old_buf_  = old_buf;
+}
+
 bool Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl*)
 {
-  MPI_Request ref = static_cast<MPI_Request>(a);
-  MPI_Request req = static_cast<MPI_Request>(b);
+  auto ref = static_cast<MPI_Request>(a);
+  auto req = static_cast<MPI_Request>(b);
   return match_common(req, req, ref);
 }
 
 bool Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl*)
 {
-  MPI_Request ref = static_cast<MPI_Request>(a);
-  MPI_Request req = static_cast<MPI_Request>(b);
+  auto ref = static_cast<MPI_Request>(a);
+  auto req = static_cast<MPI_Request>(b);
   return match_common(req, ref, req);
 }
 
-void Request::print_request(const char *message)
+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_);
@@ -163,28 +178,28 @@ void Request::print_request(const char *message)
 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)->get_pid() : 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)->get_pid() : 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)->get_pid() : 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)->get_pid() : MPI_PROC_NULL, tag, comm,
                      MPI_REQ_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
 }
 
@@ -195,11 +210,11 @@ MPI_Request Request::rma_send_init(const void *buf, int count, MPI_Datatype data
   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,
+                          dst != MPI_PROC_NULL ? comm->group()->actor(dst)->get_pid() : 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,
+                          dst != MPI_PROC_NULL ? comm->group()->actor(dst)->get_pid() : 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);
   }
@@ -208,8 +223,13 @@ MPI_Request Request::rma_send_init(const void *buf, int count, MPI_Datatype data
 
 MPI_Request Request::recv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
 {
+  int source = MPI_PROC_NULL;
+  if (src == MPI_ANY_SOURCE)
+    source = MPI_ANY_SOURCE;
+  else if (src != MPI_PROC_NULL)
+    source = comm->group()->actor(src)->get_pid();
   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);
 }
@@ -218,13 +238,18 @@ 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 */
+  int source = MPI_PROC_NULL;
+  if (src == MPI_ANY_SOURCE)
+    source = MPI_ANY_SOURCE;
+  else if (src != MPI_PROC_NULL)
+    source = comm->group()->actor(src)->get_pid();
   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)->get_pid() : 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,
+    request      = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, source,
+                          dst != MPI_PROC_NULL ? comm->group()->actor(dst)->get_pid() : MPI_PROC_NULL, tag, comm,
                           MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED | MPI_REQ_ACCUMULATE, op);
   }
   return request;
@@ -232,9 +257,13 @@ MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype,
 
 MPI_Request Request::irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
 {
+  int source = MPI_PROC_NULL;
+  if (src == MPI_ANY_SOURCE)
+    source = MPI_ANY_SOURCE;
+  else if (src != MPI_PROC_NULL)
+    source = comm->group()->actor(src)->get_pid();
   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);
 }
 
@@ -242,9 +271,10 @@ 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)->get_pid() : 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;
 }
 
@@ -252,9 +282,10 @@ 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)->get_pid() : 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;
 }
 
@@ -262,9 +293,10 @@ 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)->get_pid() : 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;
 }
 
@@ -272,28 +304,37 @@ MPI_Request Request::issend(const void *buf, int count, MPI_Datatype datatype, i
 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 */
+  int source = MPI_PROC_NULL;
+  if (src == MPI_ANY_SOURCE)
+    source = MPI_ANY_SOURCE;
+  else if (src != MPI_PROC_NULL)
+    source = comm->group()->actor(src)->get_pid();
   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)->get_pid() : 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;
 }
@@ -302,9 +343,10 @@ 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)->get_pid() : 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;
 }
@@ -313,10 +355,11 @@ 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)->get_pid() : 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;
 }
@@ -325,13 +368,20 @@ 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)
 {
-  MPI_Request requests[2];
-  MPI_Status stats[2];
+  int source = MPI_PROC_NULL;
+  if (src == MPI_ANY_SOURCE)
+    source = MPI_ANY_SOURCE;
+  else if (src != MPI_PROC_NULL)
+    source = comm->group()->actor(src)->get_pid();
+  int destination = dst != MPI_PROC_NULL ? comm->group()->actor(dst)->get_pid() : 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)) {
+  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();
@@ -340,8 +390,8 @@ void Request::sendrecv(const void *sendbuf, int sendcount, MPI_Datatype sendtype
   }
   requests[0] = isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
   requests[1] = irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
-  startall(2, requests);
-  waitall(2, requests, stats);
+  startall(2, requests.data());
+  waitall(2, requests.data(), stats.data());
   unref(&requests[0]);
   unref(&requests[1]);
   if(status != MPI_STATUS_IGNORE) {
@@ -355,10 +405,17 @@ void Request::start()
   s4u::Mailbox* mailbox;
 
   xbt_assert(action_ == nullptr, "Cannot (re-)start unfinished communication");
+  //reinitialize temporary buffer for persistent requests
+  if(real_size_ > 0 && flags_ & MPI_REQ_FINISHED){
+    buf_ = old_buf_;
+    init_buffer(real_size_/old_type_->size());
+  }
   flags_ &= ~MPI_REQ_PREPARED;
   flags_ &= ~MPI_REQ_FINISHED;
   this->ref();
 
+  // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
+  real_size_=size_;
   if ((flags_ & MPI_REQ_RECV) != 0) {
     this->print_request("New recv");
 
@@ -376,7 +433,7 @@ void Request::start()
       mailbox = process->mailbox_small();
       XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %s (in case of SSEND)?",
                 mailbox->get_cname());
-      smx_activity_t action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
+      simgrid::kernel::activity::ActivityImplPtr action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
 
       if (action == nullptr) {
         mailbox = process->mailbox();
@@ -392,7 +449,7 @@ void Request::start()
     } else {
       mailbox = process->mailbox_small();
       XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
-      smx_activity_t action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
+      simgrid::kernel::activity::ActivityImplPtr action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
 
       if (action == nullptr) {
         XBT_DEBUG("No, nothing in the permanent receive mailbox");
@@ -402,8 +459,6 @@ void Request::start()
       }
     }
 
-    // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
-    real_size_=size_;
     action_   = simcall_comm_irecv(
         process->get_actor()->get_impl(), mailbox->get_impl(), buf_, &real_size_, &match_recv,
         process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this, -1.0);
@@ -412,7 +467,7 @@ void Request::start()
     if (smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)
       mut->unlock();
   } else { /* the RECV flag was not set, so this is a send */
-    simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
+    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()) {
@@ -470,7 +525,7 @@ void Request::start()
     } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast<int>(size_) < smpi_cfg_async_small_thresh()) { // eager mode
       mailbox = process->mailbox();
       XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %s?", mailbox->get_cname());
-      smx_activity_t action = mailbox->iprobe(1, &match_send, static_cast<void*>(this));
+      simgrid::kernel::activity::ActivityImplPtr action = mailbox->iprobe(1, &match_send, static_cast<void*>(this));
       if (action == nullptr) {
         if ((flags_ & MPI_REQ_SSEND) == 0) {
           mailbox = process->mailbox_small();
@@ -494,10 +549,10 @@ void Request::start()
       XBT_DEBUG("Send request %p is in the large mailbox %s (buf: %p)", this, mailbox->get_cname(), buf_);
     }
 
-    // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
-    real_size_=size_;
-    action_   = simcall_comm_isend(
-        simgrid::s4u::Actor::by_pid(src_)->get_impl(), mailbox->get_impl(), size_, -1.0, buf, real_size_, &match_send,
+    size_t payload_size_ = size_ + 16;//MPI enveloppe size (tag+dest+communicator)
+    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
@@ -521,24 +576,26 @@ 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();
 }
 
 int Request::test(MPI_Request * request, MPI_Status * status, int* flag) {
-  //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or testall before)
+  // assume that *request is not MPI_REQUEST_NULL (filtered in PMPI_Test or testall before)
   // to avoid deadlocks if used as a break condition, such as
   //     while (MPI_Test(request, flag, status) && flag) dostuff...
   // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
   // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
+  xbt_assert(*request != MPI_REQUEST_NULL);
+
   static int nsleeps = 1;
   int ret = MPI_SUCCESS;
   
@@ -559,32 +616,29 @@ int Request::test(MPI_Request * request, MPI_Status * status, int* flag) {
 
   Status::empty(status);
   *flag = 1;
-  if (((*request)->flags_ & MPI_REQ_PREPARED) == 0) {
-    if ((*request)->action_ != nullptr && (*request)->cancelled_ != 1){
+  if (((*request)->flags_ & (MPI_REQ_PREPARED | MPI_REQ_FINISHED)) == 0) {
+    if ((*request)->action_ != nullptr && ((*request)->flags_ & MPI_REQ_CANCELLED) == 0){
       try{
-        *flag = simcall_comm_test((*request)->action_);
+        *flag = simcall_comm_test((*request)->action_.get());
       } catch (const Exception&) {
         *flag = 0;
         return ret;
       }
     }
-    if (*request != MPI_REQUEST_NULL && 
-        ((*request)->flags_ & MPI_REQ_GENERALIZED)
-        && !((*request)->flags_ & MPI_REQ_COMPLETE)) 
+    if (((*request)->flags_ & MPI_REQ_GENERALIZED) && !((*request)->flags_ & MPI_REQ_COMPLETE))
       *flag=0;
     if (*flag) {
-      finish_wait(request,status);
+      finish_wait(request, status); // may invalidate *request
       if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_GENERALIZED)){
+        MPI_Status tmp_status;
         MPI_Status* mystatus;
-        if(status==MPI_STATUS_IGNORE){
-          mystatus=new MPI_Status();
+        if (status == MPI_STATUS_IGNORE) {
+          mystatus = &tmp_status;
           Status::empty(mystatus);
-        }else{
-          mystatus=status;
+        } else {
+          mystatus = status;
         }
         ret = ((*request)->generalized_funcs)->query_fn(((*request)->generalized_funcs)->extra_state, mystatus);
-        if(status==MPI_STATUS_IGNORE) 
-          delete mystatus;
       }
       nsleeps=1;//reset the number of sleeps we will do next time
       if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_PERSISTENT) == 0)
@@ -667,16 +721,15 @@ int Request::testany(int count, MPI_Request requests[], int *index, int* flag, M
       } else {
         finish_wait(&requests[*index],status);
       if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_GENERALIZED)){
+        MPI_Status tmp_status;
         MPI_Status* mystatus;
-        if(status==MPI_STATUS_IGNORE){
-          mystatus=new MPI_Status();
+        if (status == MPI_STATUS_IGNORE) {
+          mystatus = &tmp_status;
           Status::empty(mystatus);
-        }else{
-          mystatus=status;
+        } else {
+          mystatus = status;
         }
         ret=(requests[*index]->generalized_funcs)->query_fn((requests[*index]->generalized_funcs)->extra_state, mystatus);
-        if(status==MPI_STATUS_IGNORE) 
-          delete mystatus;
       }
 
         if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_NON_PERSISTENT)) 
@@ -748,9 +801,9 @@ 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();
-  MPI_Request 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)->get_pid(),
+                             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 *
@@ -807,8 +860,7 @@ 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)
@@ -818,11 +870,16 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status)
   }
 
   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 {
+        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;
+      }
       // this handles the case were size in receive differs from size in send
       status->count = req->real_size_;
     }
@@ -846,12 +903,14 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status)
           if ((req->flags_ & MPI_REQ_RECV) && datatype->size() != 0)
             datatype->unserialize(req->buf_, req->old_buf_, req->real_size_/datatype->size() , req->op_);
           xbt_free(req->buf_);
+          req->buf_=nullptr;
         } else if (req->flags_ & MPI_REQ_RECV) { // apply op on contiguous buffer for accumulate
           if (datatype->size() != 0) {
             int n = req->real_size_ / datatype->size();
             req->op_->apply(req->buf_, req->old_buf_, &n, datatype);
           }
           xbt_free(req->buf_);
+          req->buf_=nullptr;
         }
       }
     }
@@ -875,12 +934,41 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status)
   if (req->flags_ & MPI_REQ_PERSISTENT)
     req->action_ = nullptr;
   req->flags_ |= MPI_REQ_FINISHED;
-  unref(request);
+
+  if (req->truncated_) {
+    char error_string[MPI_MAX_ERROR_STRING];
+    int error_size;
+    PMPI_Error_string(MPI_ERR_TRUNCATE, error_string, &error_size);
+    MPI_Errhandler err = (req->comm_) ? (req->comm_)->errhandler() : MPI_ERRHANDLER_NULL;
+    if (err == MPI_ERRHANDLER_NULL || err == MPI_ERRORS_RETURN)
+      XBT_WARN("recv - returned %.*s instead of MPI_SUCCESS", error_size, error_string);
+    else if (err == MPI_ERRORS_ARE_FATAL)
+      xbt_die("recv - returned %.*s instead of MPI_SUCCESS", error_size, error_string);
+    else
+      err->call((req->comm_), MPI_ERR_TRUNCATE);
+    if (err != MPI_ERRHANDLER_NULL)
+      simgrid::smpi::Errhandler::unref(err);
+    MC_assert(not MC_is_active()); /* Only fail in MC mode */
+  }
+  if(req->src_ != MPI_PROC_NULL && req->dst_ != MPI_PROC_NULL)
+    unref(request);
 }
 
 int Request::wait(MPI_Request * request, MPI_Status * status)
 {
+  // assume that *request is not MPI_REQUEST_NULL (filtered in PMPI_Wait before)
+  xbt_assert(*request != MPI_REQUEST_NULL);
+
   int ret=MPI_SUCCESS;
+
+  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;
+    }
+    (*request)=MPI_REQUEST_NULL;
+    return ret;
+  }
   // Are we waiting on a request meant for non blocking collectives ?
   // If so, wait for all the subrequests.
   if ((*request)->nbc_requests_size_>0){
@@ -909,7 +997,7 @@ int Request::wait(MPI_Request * request, MPI_Status * status)
   }
 
   (*request)->print_request("Waiting");
-  if ((*request)->flags_ & MPI_REQ_PREPARED) {
+  if ((*request)->flags_ & (MPI_REQ_PREPARED | MPI_REQ_FINISHED)) {
     Status::empty(status);
     return ret;
   }
@@ -917,31 +1005,33 @@ int Request::wait(MPI_Request * request, MPI_Status * status)
   if ((*request)->action_ != nullptr){
       try{
         // this is not a detached send
-        simcall_comm_wait((*request)->action_, -1.0);
+        simcall_comm_wait((*request)->action_.get(), -1.0);
       } catch (const Exception&) {
         XBT_VERB("Request cancelled");
       }
   }
 
-  if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_GENERALIZED)){
-    MPI_Status* mystatus;
+  if ((*request)->flags_ & MPI_REQ_GENERALIZED) {
     if(!((*request)->flags_ & MPI_REQ_COMPLETE)){
       ((*request)->generalized_funcs)->mutex->lock();
       ((*request)->generalized_funcs)->cond->wait(((*request)->generalized_funcs)->mutex);
       ((*request)->generalized_funcs)->mutex->unlock();
-      }
-    if(status==MPI_STATUS_IGNORE){
-      mystatus=new MPI_Status();
+    }
+    MPI_Status tmp_status;
+    MPI_Status* mystatus;
+    if (status == MPI_STATUS_IGNORE) {
+      mystatus = &tmp_status;
       Status::empty(mystatus);
-    }else{
-      mystatus=status;
+    } else {
+      mystatus = status;
     }
     ret = ((*request)->generalized_funcs)->query_fn(((*request)->generalized_funcs)->extra_state, mystatus);
-    if(status==MPI_STATUS_IGNORE) 
-      delete mystatus;
   }
 
-  finish_wait(request,status);
+  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;
   return ret;
@@ -949,12 +1039,11 @@ int Request::wait(MPI_Request * request, MPI_Status * status)
 
 int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
 {
-  std::vector<simgrid::kernel::activity::CommImpl*> comms;
-  comms.reserve(count);
   int index = MPI_UNDEFINED;
 
   if(count > 0) {
     // Wait for a request to complete
+    std::vector<simgrid::kernel::activity::CommImpl*> comms;
     std::vector<int> map;
     XBT_DEBUG("Wait for one of %d", count);
     for(int i = 0; i < count; i++) {
@@ -966,7 +1055,7 @@ int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
           map.push_back(i);
         } else {
           // This is a finished detached request, let's return this one
-          comms.clear(); // so we free don't do the waitany call
+          comms.clear(); // don't do the waitany call afterwards
           index = i;
           finish_wait(&requests[i], status); // cleanup if refcount = 0
           if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
@@ -977,13 +1066,12 @@ int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
     }
     if (not comms.empty()) {
       XBT_DEBUG("Enter waitany for %zu comms", comms.size());
-      int i=MPI_UNDEFINED;
+      int i;
       try{
-        // this is not a detached send
         i = simcall_comm_waitany(comms.data(), comms.size(), -1);
       } catch (const Exception&) {
-        XBT_INFO("request %d cancelled ", i);
-        return i;
+        XBT_INFO("request cancelled");
+        i = -1;
       }
 
       // not MPI_UNDEFINED, as this is a simix return code
@@ -1006,7 +1094,7 @@ int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
   return index;
 }
 
-static int sort_accumulates(MPI_Request a, MPI_Request b)
+static int sort_accumulates(const Request* a, const Request* b)
 {
   return (a->tag() > b->tag());
 }
@@ -1035,8 +1123,8 @@ int Request::waitall(int count, MPI_Request requests[], MPI_Status status[])
       wait(&requests[c],pstat);
       index = c;
     } else {
-      index = waitany(count, (MPI_Request*)requests, pstat);
-      
+      index = waitany(count, requests, pstat);
+
       if (index == MPI_UNDEFINED)
         break;
 
@@ -1070,7 +1158,7 @@ int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Sta
   int index = 0;
   MPI_Status stat;
   MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
-  index = waitany(incount, (MPI_Request*)requests, pstat);
+  index             = waitany(incount, requests, pstat);
   if(index==MPI_UNDEFINED) return MPI_UNDEFINED;
   if(status != MPI_STATUSES_IGNORE) {
     status[count] = *pstat;
@@ -1097,17 +1185,15 @@ int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Sta
 
 MPI_Request Request::f2c(int id)
 {
-  char key[KEY_SIZE];
   if(id==MPI_FORTRAN_REQUEST_NULL)
-    return static_cast<MPI_Request>(MPI_REQUEST_NULL);
-  return static_cast<MPI_Request>(F2C::f2c_lookup()->at(get_key(key,id)));
+    return MPI_REQUEST_NULL;
+  return static_cast<MPI_Request>(F2C::lookup()->at(id));
 }
 
 void Request::free_f(int id)
 {
   if (id != MPI_FORTRAN_REQUEST_NULL) {
-    char key[KEY_SIZE];
-    F2C::f2c_lookup()->erase(get_key(key, id));
+    F2C::lookup()->erase(id);
   }
 }
 
@@ -1116,7 +1202,7 @@ 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;
   }
@@ -1146,7 +1232,7 @@ int Request::grequest_start(MPI_Grequest_query_function* query_fn, MPI_Grequest_
   (*request)->flags_ |= MPI_REQ_GENERALIZED;
   (*request)->flags_ |= MPI_REQ_PERSISTENT;
   (*request)->refcount_ = 1;
-  ((*request)->generalized_funcs) = new s_smpi_mpi_generalized_request_funcs_t;
+  ((*request)->generalized_funcs)             = std::make_unique<smpi_mpi_generalized_request_funcs_t>();
   ((*request)->generalized_funcs)->query_fn=query_fn;
   ((*request)->generalized_funcs)->free_fn=free_fn;
   ((*request)->generalized_funcs)->cancel_fn=cancel_fn;
@@ -1158,7 +1244,7 @@ int Request::grequest_start(MPI_Grequest_query_function* query_fn, MPI_Grequest_
 
 int Request::grequest_complete(MPI_Request request)
 {
-  if ((!(request->flags_ & MPI_REQ_GENERALIZED)) || request->generalized_funcs->mutex==NULL) 
+  if ((!(request->flags_ & MPI_REQ_GENERALIZED)) || request->generalized_funcs->mutex == nullptr)
     return MPI_ERR_REQUEST;
   request->generalized_funcs->mutex->lock();
   request->flags_ |= MPI_REQ_COMPLETE; // in case wait would be called after complete
@@ -1177,13 +1263,14 @@ void Request::set_nbc_requests(MPI_Request* reqs, int size){
   }
 }
 
-int Request::get_nbc_requests_size(){
+int Request::get_nbc_requests_size() const
+{
   return nbc_requests_size_;
 }
 
-MPI_Request* Request::get_nbc_requests(){
+MPI_Request* Request::get_nbc_requests() const
+{
   return nbc_requests_;
 }
-
 }
 }