X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/e51fd157704f026607d5ef1b79334cb2decf8d5f..35f18859ce88a96b7e07de4d7d70867424feb8df:/src/smpi/mpi/smpi_request.cpp

diff --git a/src/smpi/mpi/smpi_request.cpp b/src/smpi/mpi/smpi_request.cpp
index 988dc1dd9f..05cdc18c1e 100644
--- a/src/smpi/mpi/smpi_request.cpp
+++ b/src/smpi/mpi/smpi_request.cpp
@@ -1,26 +1,28 @@
-/* Copyright (c) 2007-2018. The SimGrid Team. All rights reserved.          */
+/* Copyright (c) 2007-2019. 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 "simgrid/s4u/Mutex.hpp"
+#include "simgrid/s4u/ConditionVariable.hpp"
 #include "smpi_request.hpp"
 
-#include "smpi_host.hpp"
 #include "mc/mc.h"
 #include "private.hpp"
+#include "simgrid/Exception.hpp"
+#include "simgrid/s4u/Exec.hpp"
 #include "smpi_comm.hpp"
 #include "smpi_datatype.hpp"
+#include "smpi_host.hpp"
 #include "smpi_op.hpp"
-#include "smpi_process.hpp"
 #include "src/kernel/activity/CommImpl.hpp"
 #include "src/mc/mc_replay.hpp"
-#include "src/simix/ActorImpl.hpp"
+#include "src/smpi/include/smpi_actor.hpp"
 #include "xbt/config.hpp"
-#include <xbt/ex.hpp>
+
 
 #include <algorithm>
 
-XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_request, smpi, "Logging specific to SMPI (reques)");
+XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_request, smpi, "Logging specific to SMPI (request)");
 
 static simgrid::config::Flag<double> smpi_iprobe_sleep(
   "smpi/iprobe", "Minimum time to inject inside a call to MPI_Iprobe", 1e-4);
@@ -29,19 +31,19 @@ static simgrid::config::Flag<double> smpi_test_sleep(
 
 std::vector<s_smpi_factor_t> smpi_ois_values;
 
-extern void (*smpi_comm_copy_data_callback) (smx_activity_t, void*, size_t);
+extern void (*smpi_comm_copy_data_callback)(simgrid::kernel::activity::CommImpl*, void*, size_t);
 
 namespace simgrid{
 namespace smpi{
 
-Request::Request(void* buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm, unsigned flags)
-    : buf_(buf), old_type_(datatype), src_(src), dst_(dst), tag_(tag), comm_(comm), flags_(flags)
+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)
 {
   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 = buf;
+    old_buf = const_cast<void*>(buf);
     if (count==0){
       buf_ = nullptr;
     }else {
@@ -56,55 +58,27 @@ Request::Request(void* buf, int count, MPI_Datatype datatype, int src, int dst,
   size_ = datatype->size() * count;
   datatype->ref();
   comm_->ref();
+  if(op != MPI_REPLACE && op != MPI_OP_NULL)
+    op_->ref();
   action_          = nullptr;
-  detached_        = 0;
+  detached_        = false;
   detached_sender_ = nullptr;
   real_src_        = 0;
-  truncated_       = 0;
+  truncated_       = false;
   real_size_       = 0;
   real_tag_        = 0;
   if (flags & MPI_REQ_PERSISTENT)
     refcount_ = 1;
   else
     refcount_ = 0;
-  op_   = MPI_REPLACE;
   cancelled_ = 0;
+  generalized_funcs=nullptr;
+  nbc_requests_=nullptr;
+  nbc_requests_size_=0;
 }
 
-MPI_Comm Request::comm(){
-  return comm_;
-}
-
-int Request::src(){
-  return src_;
-}
-
-int Request::dst(){
-  return dst_;
-}
-
-int Request::tag(){
-  return tag_;
-}
-
-int Request::flags(){
-  return flags_;
-}
-
-int Request::detached(){
-  return detached_;
-}
-
-MPI_Datatype Request::type(){
-  return old_type_;
-}
-
-size_t Request::size(){
-  return size_;
-}
-
-size_t Request::real_size(){
-  return real_size_;
+void Request::ref(){
+  refcount_++;
 }
 
 void Request::unref(MPI_Request* request)
@@ -116,11 +90,19 @@ void Request::unref(MPI_Request* request)
       xbt_die("Whoops, wrong refcount");
     }
     if((*request)->refcount_==0){
-        Datatype::unref((*request)->old_type_);
+      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_);
-        (*request)->print_request("Destroying");
-        delete *request;
-        *request = MPI_REQUEST_NULL;
+        Datatype::unref((*request)->old_type_);
+      }
+      if ((*request)->op_!=MPI_REPLACE && (*request)->op_!=MPI_OP_NULL)
+        Op::unref(&(*request)->op_);
+
+      (*request)->print_request("Destroying");
+      delete *request;
+      *request = MPI_REQUEST_NULL;
     }else{
       (*request)->print_request("Decrementing");
     }
@@ -129,15 +111,16 @@ void Request::unref(MPI_Request* request)
   }
 }
 
-int Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl* ignored)
+int 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);
-  XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
+  XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d, id %d against %d",ref->src_,req->src_, ref->tag_, req->tag_,ref->comm_->id(),req->comm_->id());
 
   xbt_assert(ref, "Cannot match recv against null reference");
   xbt_assert(req, "Cannot match recv against null request");
-  if((ref->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
+  if((ref->comm_->id()==MPI_UNDEFINED || req->comm_->id() == MPI_UNDEFINED || (ref->comm_->id()==req->comm_->id()))
+    && ((ref->src_ == MPI_ANY_SOURCE  && (ref->comm_->group()->rank(req->src_) != MPI_UNDEFINED)) || req->src_ == ref->src_)
     && ((ref->tag_ == MPI_ANY_TAG && req->tag_ >=0) || req->tag_ == ref->tag_)){
     //we match, we can transfer some values
     if(ref->src_ == MPI_ANY_SOURCE)
@@ -145,8 +128,8 @@ int Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl* i
     if(ref->tag_ == MPI_ANY_TAG)
       ref->real_tag_ = req->tag_;
     if(ref->real_size_ < req->real_size_)
-      ref->truncated_ = 1;
-    if(req->detached_==1)
+      ref->truncated_ = true;
+    if (req->detached_)
       ref->detached_sender_=req; //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 uncancellable
@@ -155,23 +138,24 @@ int Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl* i
   }else return 0;
 }
 
-int Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl* ignored)
+int 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);
-  XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
+  XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d, id %d against %d",ref->src_,req->src_, ref->tag_, req->tag_,ref->comm_->id(),req->comm_->id());
   xbt_assert(ref, "Cannot match send against null reference");
   xbt_assert(req, "Cannot match send against null request");
 
-  if((req->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
+  if((ref->comm_->id()==MPI_UNDEFINED || req->comm_->id() == MPI_UNDEFINED || (ref->comm_->id()==req->comm_->id()))
+      && ((req->src_ == MPI_ANY_SOURCE  && (req->comm_->group()->rank(ref->src_) != MPI_UNDEFINED)) || req->src_ == ref->src_)
       && ((req->tag_ == MPI_ANY_TAG && ref->tag_ >=0)|| req->tag_ == ref->tag_)){
     if(req->src_ == MPI_ANY_SOURCE)
       req->real_src_ = ref->src_;
     if(req->tag_ == MPI_ANY_TAG)
       req->real_tag_ = ref->tag_;
     if(req->real_size_ < ref->real_size_)
-      req->truncated_ = 1;
-    if(ref->detached_==1)
+      req->truncated_ = true;
+    if (ref->detached_)
       req->detached_sender_=ref; //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 uncancellable
@@ -189,7 +173,7 @@ void Request::print_request(const char *message)
 
 
 /* factories, to hide the internal flags from the caller */
-MPI_Request Request::send_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
+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(),
@@ -197,14 +181,14 @@ MPI_Request Request::send_init(void *buf, int count, MPI_Datatype datatype, int
                      MPI_REQ_PERSISTENT | MPI_REQ_SEND | MPI_REQ_PREPARED);
 }
 
-MPI_Request Request::ssend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
+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,
                      MPI_REQ_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
 }
 
-MPI_Request Request::isend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
+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,
@@ -212,7 +196,7 @@ MPI_Request Request::isend_init(void *buf, int count, MPI_Datatype datatype, int
 }
 
 
-MPI_Request Request::rma_send_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
+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 */
@@ -224,8 +208,7 @@ MPI_Request Request::rma_send_init(void *buf, int count, MPI_Datatype datatype,
     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_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED |
-                              MPI_REQ_ACCUMULATE);
-    request->op_ = op;
+                              MPI_REQ_ACCUMULATE, op);
   }
   return request;
 }
@@ -249,8 +232,7 @@ MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype,
   }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);
-    request->op_ = op;
+                          MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED | MPI_REQ_ACCUMULATE, op);
   }
   return request;
 }
@@ -263,7 +245,7 @@ MPI_Request Request::irecv_init(void *buf, int count, MPI_Datatype datatype, int
                      MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
 }
 
-MPI_Request Request::isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
+MPI_Request Request::isend(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(),
@@ -273,7 +255,7 @@ MPI_Request Request::isend(void *buf, int count, MPI_Datatype datatype, int dst,
   return request;
 }
 
-MPI_Request Request::issend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
+MPI_Request Request::issend(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(),
@@ -302,7 +284,7 @@ void Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag
   request = nullptr;
 }
 
-void Request::send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
+void Request::send(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(),
@@ -313,7 +295,7 @@ void Request::send(void *buf, int count, MPI_Datatype datatype, int dst, int tag
   request = nullptr;
 }
 
-void Request::ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
+void Request::ssend(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(),
@@ -325,7 +307,7 @@ void Request::ssend(void *buf, int count, MPI_Datatype datatype, int dst, int ta
   request = nullptr;
 }
 
-void Request::sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
+void Request::sendrecv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
                        void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
                        MPI_Comm comm, MPI_Status * status)
 {
@@ -356,23 +338,23 @@ void Request::sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int d
 
 void Request::start()
 {
-  smx_mailbox_t mailbox;
+  s4u::Mailbox* mailbox;
 
   xbt_assert(action_ == nullptr, "Cannot (re-)start unfinished communication");
   flags_ &= ~MPI_REQ_PREPARED;
   flags_ &= ~MPI_REQ_FINISHED;
-  refcount_++;
+  this->ref();
 
   if ((flags_ & MPI_REQ_RECV) != 0) {
     this->print_request("New recv");
 
-    simgrid::smpi::Process* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
+    simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
 
     int async_small_thresh = simgrid::config::get_value<int>("smpi/async-small-thresh");
 
-    xbt_mutex_t mut = process->mailboxes_mutex();
+    simgrid::s4u::MutexPtr mut = process->mailboxes_mutex();
     if (async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0)
-      xbt_mutex_acquire(mut);
+      mut->lock();
 
     if (async_small_thresh == 0 && (flags_ & MPI_REQ_RMA) == 0) {
       mailbox = process->mailbox();
@@ -380,15 +362,16 @@ void Request::start()
       //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
       //begin with the more appropriate one : the small one.
       mailbox = process->mailbox_small();
-      XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
-      smx_activity_t action = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(this));
+      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));
 
       if (action == nullptr) {
         mailbox = process->mailbox();
-        XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
-        action = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(this));
+        XBT_DEBUG("No, nothing in the small mailbox test the other one : %s", mailbox->get_cname());
+        action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
         if (action == nullptr) {
-          XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox);
+          XBT_DEBUG("Still nothing, switch back to the small mailbox : %s", mailbox->get_cname());
           mailbox = process->mailbox_small();
         }
       } else {
@@ -397,7 +380,7 @@ void Request::start()
     } else {
       mailbox = process->mailbox_small();
       XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
-      smx_activity_t action = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(this));
+      smx_activity_t action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
 
       if (action == nullptr) {
         XBT_DEBUG("No, nothing in the permanent receive mailbox");
@@ -410,14 +393,14 @@ 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, buf_, &real_size_, &match_recv,
+        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);
     XBT_DEBUG("recv simcall posted");
 
     if (async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0)
-      xbt_mutex_release(mut);
+      mut->unlock();
   } else { /* the RECV flag was not set, so this is a send */
-    simgrid::smpi::Process* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
+    simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
     int rank = src_;
     if (TRACE_smpi_view_internals()) {
       TRACE_smpi_send(rank, rank, dst_, tag_, size_);
@@ -429,13 +412,13 @@ void Request::start()
         ((flags_ & MPI_REQ_RMA) != 0 ||
          static_cast<int>(size_) < simgrid::config::get_value<int>("smpi/send-is-detached-thresh"))) {
       void *oldbuf = nullptr;
-      detached_ = 1;
+      detached_    = true;
       XBT_DEBUG("Send request %p is detached", this);
-      refcount_++;
+      this->ref();
       if (not(old_type_->flags() & DT_FLAG_DERIVED)) {
         oldbuf = buf_;
         if (not process->replaying() && oldbuf != nullptr && size_ != 0) {
-          if ((smpi_privatize_global_variables != SmpiPrivStrategies::None) &&
+          if ((smpi_privatize_global_variables != SmpiPrivStrategies::NONE) &&
               (static_cast<char*>(buf_) >= smpi_data_exe_start) &&
               (static_cast<char*>(buf_) < smpi_data_exe_start + smpi_data_exe_size)) {
             XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
@@ -450,7 +433,7 @@ void Request::start()
 
     //if we are giving back the control to the user without waiting for completion, we have to inject timings
     double sleeptime = 0.0;
-    if (detached_ != 0 || ((flags_ & (MPI_REQ_ISEND | MPI_REQ_SSEND)) != 0)) { // issend should be treated as isend
+    if (detached_ || ((flags_ & (MPI_REQ_ISEND | MPI_REQ_SSEND)) != 0)) { // issend should be treated as isend
       // isend and send timings may be different
       sleeptime = ((flags_ & MPI_REQ_ISEND) != 0)
                       ? simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->oisend(size_)
@@ -464,25 +447,27 @@ void Request::start()
 
     int async_small_thresh = simgrid::config::get_value<int>("smpi/async-small-thresh");
 
-    xbt_mutex_t mut=process->mailboxes_mutex();
+    simgrid::s4u::MutexPtr mut = process->mailboxes_mutex();
 
     if (async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0)
-      xbt_mutex_acquire(mut);
+      mut->lock();
 
     if (not(async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0)) {
       mailbox = process->mailbox();
     } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast<int>(size_) < async_small_thresh) { // eager mode
       mailbox = process->mailbox();
-      XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
-      smx_activity_t action = simcall_comm_iprobe(mailbox, 1, &match_send, static_cast<void*>(this));
+      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));
       if (action == nullptr) {
         if ((flags_ & MPI_REQ_SSEND) == 0) {
           mailbox = process->mailbox_small();
-          XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox);
+          XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %s",
+                    mailbox->get_cname());
         } else {
           mailbox = process->mailbox_small();
-          XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox);
-          action = simcall_comm_iprobe(mailbox, 1, &match_send, static_cast<void*>(this));
+          XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %s?",
+                    mailbox->get_cname());
+          action = mailbox->iprobe(1, &match_send, static_cast<void*>(this));
           if (action == nullptr) {
             XBT_DEBUG("No, we are first, send to large mailbox");
             mailbox = process->mailbox();
@@ -493,13 +478,13 @@ void Request::start()
       }
     } else {
       mailbox = process->mailbox();
-      XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, this,buf_);
+      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, size_, -1.0, buf, real_size_, &match_send,
+        simgrid::s4u::Actor::by_pid(src_)->get_impl(), mailbox->get_impl(), size_, -1.0, buf, real_size_, &match_send,
         &xbt_free_f, // how to free the userdata if a detached send fails
         not process->replaying() ? smpi_comm_copy_data_callback : &smpi_comm_null_copy_buffer_callback, this,
         // detach if msg size < eager/rdv switch limit
@@ -507,10 +492,13 @@ void Request::start()
     XBT_DEBUG("send simcall posted");
 
     /* FIXME: detached sends are not traceable (action_ == nullptr) */
-    if (action_ != nullptr)
-      simcall_set_category(action_, TRACE_internal_smpi_get_category());
+    if (action_ != nullptr) {
+      boost::static_pointer_cast<kernel::activity::CommImpl>(action_)->set_tracing_category(
+          smpi_process()->get_tracing_category());
+    }
+
     if (async_small_thresh != 0 || ((flags_ & MPI_REQ_RMA) != 0))
-      xbt_mutex_release(mut);
+      mut->unlock();
   }
 }
 
@@ -532,28 +520,59 @@ void Request::cancel()
     (boost::static_pointer_cast<simgrid::kernel::activity::CommImpl>(this->action_))->cancel();
 }
 
-int Request::test(MPI_Request * request, MPI_Status * status) {
+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)
   // 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
   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)
     simcall_process_sleep(nsleeps*smpi_test_sleep);
 
   Status::empty(status);
-  int flag = 1;
+  *flag = 1;
   if (((*request)->flags_ & MPI_REQ_PREPARED) == 0) {
     if ((*request)->action_ != nullptr){
       try{
-        flag = simcall_comm_test((*request)->action_);
-      }catch (xbt_ex& e) {
-        return 0;
+        *flag = simcall_comm_test((*request)->action_);
+      } catch (const Exception&) {
+        *flag = 0;
+        return ret;
       }
     }
-    if (flag) {
+    if (*request != MPI_REQUEST_NULL && 
+        ((*request)->flags_ & MPI_REQ_GENERALIZED)
+        && !((*request)->flags_ & MPI_REQ_COMPLETE)) 
+      *flag=0;
+    if (*flag) {
       finish_wait(request,status);
+      if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_GENERALIZED)){
+        MPI_Status* mystatus;
+        if(status==MPI_STATUS_IGNORE){
+          mystatus=new MPI_Status();
+          Status::empty(mystatus);
+        }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)
         *request = MPI_REQUEST_NULL;
@@ -561,21 +580,27 @@ int Request::test(MPI_Request * request, MPI_Status * status) {
       nsleeps++;
     }
   }
-  return flag;
+  return ret;
 }
 
-int Request::testsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
+int Request::testsome(int incount, MPI_Request requests[], int *count, int *indices, MPI_Status status[])
 {
-  int count = 0;
+  int ret = MPI_SUCCESS;
+  int error=0;
   int count_dead = 0;
+  int flag = 0;
   MPI_Status stat;
   MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
 
+  *count = 0;
   for (int i = 0; i < incount; i++) {
     if (requests[i] != MPI_REQUEST_NULL) {
-      if (test(&requests[i], pstat)) {
+      ret = test(&requests[i], pstat, &flag);
+      if(ret!=MPI_SUCCESS)
+        error = 1;
+      if(flag) {
         indices[i] = 1;
-        count++;
+        (*count)++;
         if (status != MPI_STATUSES_IGNORE)
           status[i] = *pstat;
         if ((requests[i] != MPI_REQUEST_NULL) && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
@@ -585,25 +610,27 @@ int Request::testsome(int incount, MPI_Request requests[], int *indices, MPI_Sta
       count_dead++;
     }
   }
-  if(count_dead==incount)
-    return MPI_UNDEFINED;
-  else return count;
+  if(count_dead==incount)*count=MPI_UNDEFINED;
+  if(error!=0)
+    return MPI_ERR_IN_STATUS;
+  else
+    return MPI_SUCCESS;
 }
 
-int Request::testany(int count, MPI_Request requests[], int *index, MPI_Status * status)
+int Request::testany(int count, MPI_Request requests[], int *index, int* flag, MPI_Status * status)
 {
-  std::vector<simgrid::kernel::activity::ActivityImplPtr> comms;
+  std::vector<simgrid::kernel::activity::CommImpl*> comms;
   comms.reserve(count);
 
   int i;
-  int flag = 0;
-
+  *flag = 0;
+  int ret = MPI_SUCCESS;
   *index = MPI_UNDEFINED;
 
   std::vector<int> map; /** Maps all matching comms back to their location in requests **/
   for(i = 0; i < count; i++) {
     if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action_ && not(requests[i]->flags_ & MPI_REQ_PREPARED)) {
-      comms.push_back(requests[i]->action_);
+      comms.push_back(static_cast<simgrid::kernel::activity::CommImpl*>(requests[i]->action_.get()));
       map.push_back(i);
     }
   }
@@ -614,42 +641,71 @@ int Request::testany(int count, MPI_Request requests[], int *index, MPI_Status *
       simcall_process_sleep(nsleeps*smpi_test_sleep);
     try{
       i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches!
-    }catch (xbt_ex& e) {
+    } catch (const Exception&) {
+      XBT_DEBUG("Exception in testany");
       return 0;
     }
     
     if (i != -1) { // -1 is not MPI_UNDEFINED but a SIMIX return code. (nothing matches)
       *index = map[i];
-      finish_wait(&requests[*index],status);
-      flag             = 1;
-      nsleeps          = 1;
-      if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_NON_PERSISTENT)) {
-        requests[*index] = MPI_REQUEST_NULL;
+      if (requests[*index] != MPI_REQUEST_NULL && 
+          (requests[*index]->flags_ & MPI_REQ_GENERALIZED)
+          && !(requests[*index]->flags_ & MPI_REQ_COMPLETE)) {
+        *flag=0;
+      } else {
+        finish_wait(&requests[*index],status);
+      if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_GENERALIZED)){
+        MPI_Status* mystatus;
+        if(status==MPI_STATUS_IGNORE){
+          mystatus=new MPI_Status();
+          Status::empty(mystatus);
+        }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)) 
+          requests[*index] = MPI_REQUEST_NULL;
+        XBT_DEBUG("Testany - returning with index %d", *index);
+        *flag=1;
       }
+      nsleeps = 1;
     } else {
       nsleeps++;
     }
   } else {
+      XBT_DEBUG("Testany on inactive handles, returning flag=1 but empty status");
       //all requests are null or inactive, return true
-      flag = 1;
+      *flag = 1;
+      *index = MPI_UNDEFINED;
       Status::empty(status);
   }
 
-  return flag;
+  return ret;
 }
 
-int Request::testall(int count, MPI_Request requests[], MPI_Status status[])
+int Request::testall(int count, MPI_Request requests[], int* outflag, MPI_Status status[])
 {
   MPI_Status stat;
   MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
-  int flag=1;
+  int flag;
+  int error = 0;
+  int ret=MPI_SUCCESS;
+  *outflag = 1;
   for(int i=0; i<count; i++){
     if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_PREPARED)) {
-      if (test(&requests[i], pstat)!=1){
+      ret = test(&requests[i], pstat, &flag);
+      if (flag){
         flag=0;
+        requests[i]=MPI_REQUEST_NULL;
       }else{
-          requests[i]=MPI_REQUEST_NULL;
+        *outflag=0;
       }
+      if (ret != MPI_SUCCESS) 
+        error = 1;
     }else{
       Status::empty(pstat);
     }
@@ -657,7 +713,10 @@ int Request::testall(int count, MPI_Request requests[], MPI_Status status[])
       status[i] = *pstat;
     }
   }
-  return flag;
+  if(error==1) 
+    return MPI_ERR_IN_STATUS;
+  else 
+    return MPI_SUCCESS;
 }
 
 void Request::probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
@@ -676,38 +735,45 @@ void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status*
   // nsleeps is a multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
   // This can speed up the execution of certain applications by an order of magnitude, such as HPL
   static int nsleeps = 1;
-  double speed        = simgrid::s4u::Actor::self()->get_host()->getSpeed();
+  double speed        = s4u::this_actor::get_host()->get_speed();
   double maxrate      = simgrid::config::get_value<double>("smpi/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);
   if (smpi_iprobe_sleep > 0) {
-    smx_activity_t iprobe_sleep = simcall_execution_start(
-        "iprobe", /* flops to executek*/ nsleeps * smpi_iprobe_sleep * speed * maxrate, /* priority */ 1.0,
-        /* performance bound */ maxrate * speed, smpi_process()->get_actor()->get_host());
-    simcall_execution_wait(iprobe_sleep);
+    /** Compute the number of flops we will sleep **/
+    s4u::this_actor::exec_init(/*nsleeps: See comment above */ nsleeps *
+                               /*(seconds * flop/s -> total flops)*/ smpi_iprobe_sleep * speed * maxrate)
+        ->set_name("iprobe")
+        /* Not the entire CPU can be used when iprobing: This is important for
+         * the energy consumption caused by polling with iprobes. 
+         * Note also that the number of flops that was
+         * computed above contains a maxrate factor and is hence reduced (maxrate < 1)
+         */
+        ->set_bound(maxrate*speed)
+        ->start()
+        ->wait();
   }
   // behave like a receive, but don't do it
-  smx_mailbox_t mailbox;
+  s4u::Mailbox* mailbox;
 
   request->print_request("New iprobe");
   // We have to test both mailboxes as we don't know if we will receive one one or another
   if (simgrid::config::get_value<int>("smpi/async-small-thresh") > 0) {
     mailbox = smpi_process()->mailbox_small();
     XBT_DEBUG("Trying to probe the perm recv mailbox");
-    request->action_ = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(request));
+    request->action_ = mailbox->iprobe(0, &match_recv, static_cast<void*>(request));
   }
 
   if (request->action_ == nullptr){
     mailbox = smpi_process()->mailbox();
     XBT_DEBUG("trying to probe the other mailbox");
-    request->action_ = simcall_comm_iprobe(mailbox, 0, &match_recv, static_cast<void*>(request));
+    request->action_ = mailbox->iprobe(0, &match_recv, static_cast<void*>(request));
   }
 
   if (request->action_ != nullptr){
-    simgrid::kernel::activity::CommImplPtr sync_comm =
-        boost::static_pointer_cast<simgrid::kernel::activity::CommImpl>(request->action_);
-    MPI_Request req                            = static_cast<MPI_Request>(sync_comm->src_data);
+    kernel::activity::CommImplPtr sync_comm = boost::static_pointer_cast<kernel::activity::CommImpl>(request->action_);
+    MPI_Request req                         = static_cast<MPI_Request>(sync_comm->src_data_);
     *flag = 1;
     if (status != MPI_STATUS_IGNORE && (req->flags_ & MPI_REQ_PREPARED) == 0) {
       status->MPI_SOURCE = comm->group()->rank(req->src_);
@@ -723,6 +789,7 @@ void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status*
       nsleeps++;
   }
   unref(&request);
+  xbt_assert(request == MPI_REQUEST_NULL);
 }
 
 void Request::finish_wait(MPI_Request* request, MPI_Status * status)
@@ -733,44 +800,49 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status)
   if (req->cancelled_==1){
     if (status!=MPI_STATUS_IGNORE)
       status->cancelled=1;
+    if(req->detached_sender_ != nullptr)
+      unref(&(req->detached_sender_));
+    unref(request);
     return;
   }
 
-  if (not((req->detached_ != 0) && ((req->flags_ & MPI_REQ_SEND) != 0)) && ((req->flags_ & MPI_REQ_PREPARED) == 0)) {
+  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_ != 0 ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
+      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_;
     }
+    //detached send will be finished at the other end
+    if (not(req->detached_ && ((req->flags_ & MPI_REQ_SEND) != 0))) {
+      req->print_request("Finishing");
+      MPI_Datatype datatype = req->old_type_;
+
+      // FIXME Handle the case of a partial shared malloc.
+      if (((req->flags_ & MPI_REQ_ACCUMULATE) != 0) ||
+          (datatype->flags() & DT_FLAG_DERIVED)) { // && (not smpi_is_shared(req->old_buf_))){
+
+        if (not smpi_process()->replaying() && smpi_privatize_global_variables != SmpiPrivStrategies::NONE &&
+            static_cast<char*>(req->old_buf_) >= smpi_data_exe_start &&
+            static_cast<char*>(req->old_buf_) < smpi_data_exe_start + smpi_data_exe_size) {
+          XBT_VERB("Privatization : We are unserializing to a zone in global memory  Switch data segment ");
+          smpi_switch_data_segment(simgrid::s4u::Actor::self());
+        }
 
-    req->print_request("Finishing");
-    MPI_Datatype datatype = req->old_type_;
-
-// FIXME Handle the case of a partial shared malloc.
-    if (((req->flags_ & MPI_REQ_ACCUMULATE) != 0) ||
-        (datatype->flags() & DT_FLAG_DERIVED)) { // && (not smpi_is_shared(req->old_buf_))){
-
-      if (not smpi_process()->replaying() && smpi_privatize_global_variables != SmpiPrivStrategies::None &&
-          static_cast<char*>(req->old_buf_) >= smpi_data_exe_start &&
-          static_cast<char*>(req->old_buf_) < smpi_data_exe_start + smpi_data_exe_size) {
-        XBT_VERB("Privatization : We are unserializing to a zone in global memory  Switch data segment ");
-        smpi_switch_data_segment(simgrid::s4u::Actor::self());
-      }
-
-      if(datatype->flags() & DT_FLAG_DERIVED){
-        // This part handles the problem of non-contignous memory the unserialization at the reception
-        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_);
-      } 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);
+        if(datatype->flags() & DT_FLAG_DERIVED){
+          // This part handles the problem of non-contignous memory the unserialization at the reception
+          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_);
+        } 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_);
         }
-        xbt_free(req->buf_);
       }
     }
   }
@@ -796,54 +868,95 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status)
   unref(request);
 }
 
-void Request::wait(MPI_Request * request, MPI_Status * status)
+int Request::wait(MPI_Request * request, MPI_Status * status)
 {
+  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]));
+    }
+    delete[] (*request)->nbc_requests_;
+    (*request)->nbc_requests_size_=0;
+    unref(request);
+    (*request)=MPI_REQUEST_NULL;
+    return ret;
+  }
+
   (*request)->print_request("Waiting");
   if ((*request)->flags_ & MPI_REQ_PREPARED) {
     Status::empty(status);
-    return;
+    return ret;
   }
 
   if ((*request)->action_ != nullptr){
       try{
         // this is not a detached send
         simcall_comm_wait((*request)->action_, -1.0);
-      }catch (xbt_ex& e) {
+      } 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_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();
+      Status::empty(mystatus);
+    }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 != MPI_REQUEST_NULL && (((*request)->flags_ & MPI_REQ_NON_PERSISTENT) != 0))
     *request = MPI_REQUEST_NULL;
+  return ret;
 }
 
 int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
 {
-  s_xbt_dynar_t comms; // Keep it on stack to save some extra mallocs
+  std::vector<simgrid::kernel::activity::CommImpl*> comms;
+  comms.reserve(count);
   int index = MPI_UNDEFINED;
 
   if(count > 0) {
-    int size = 0;
     // Wait for a request to complete
-    xbt_dynar_init(&comms, sizeof(smx_activity_t), [](void*ptr){
-      intrusive_ptr_release(*(simgrid::kernel::activity::ActivityImpl**)ptr);
-    });
-    int *map = xbt_new(int, count);
+    std::vector<int> map;
     XBT_DEBUG("Wait for one of %d", count);
     for(int i = 0; i < count; i++) {
       if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_PREPARED) &&
           not(requests[i]->flags_ & MPI_REQ_FINISHED)) {
         if (requests[i]->action_ != nullptr) {
           XBT_DEBUG("Waiting any %p ", requests[i]);
-          intrusive_ptr_add_ref(requests[i]->action_.get());
-          xbt_dynar_push_as(&comms, simgrid::kernel::activity::ActivityImpl*, requests[i]->action_.get());
-          map[size] = i;
-          size++;
+          comms.push_back(static_cast<simgrid::kernel::activity::CommImpl*>(requests[i]->action_.get()));
+          map.push_back(i);
         } else {
           // This is a finished detached request, let's return this one
-          size  = 0; // so we free the dynar but don't do the waitany call
+          comms.clear(); // so we free don't do the waitany call
           index = i;
           finish_wait(&requests[i], status); // cleanup if refcount = 0
           if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
@@ -852,14 +965,14 @@ int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
         }
       }
     }
-    if (size > 0) {
-      XBT_DEBUG("Enter waitany for %lu comms", xbt_dynar_length(&comms));
+    if (not comms.empty()) {
+      XBT_DEBUG("Enter waitany for %zu comms", comms.size());
       int i=MPI_UNDEFINED;
       try{
         // this is not a detached send
-        i = simcall_comm_waitany(&comms, -1);
-      }catch (xbt_ex& e) {
-      XBT_INFO("request %d cancelled ",i);
+        i = simcall_comm_waitany(comms.data(), comms.size(), -1);
+      } catch (const Exception&) {
+        XBT_INFO("request %d cancelled ", i);
         return i;
       }
 
@@ -870,14 +983,11 @@ int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
         if ((requests[index] == MPI_REQUEST_NULL) ||
             (not((requests[index]->flags_ & MPI_REQ_ACCUMULATE) && (requests[index]->flags_ & MPI_REQ_RECV)))) {
           finish_wait(&requests[index],status);
-          if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
+          if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT))
             requests[index] = MPI_REQUEST_NULL;
         }
       }
     }
-
-    xbt_dynar_free_data(&comms);
-    xbt_free(map);
   }
 
   if (index==MPI_UNDEFINED)
@@ -946,21 +1056,32 @@ int Request::waitall(int count, MPI_Request requests[], MPI_Status status[])
 int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
 {
   int count = 0;
+  int flag = 0;
+  int index = 0;
   MPI_Status stat;
   MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
 
+  index = waitany(incount, (MPI_Request*)requests, pstat);
+  if(index==MPI_UNDEFINED) return MPI_UNDEFINED;
+  if(status != MPI_STATUSES_IGNORE) {
+    status[count] = *pstat;
+  }
+  indices[count] = index;
+  count++;
   for (int i = 0; i < incount; i++) {
-    int index = waitany(incount, requests, pstat);
-    if(index!=MPI_UNDEFINED){
-      indices[count] = index;
-      count++;
-      if(status != MPI_STATUSES_IGNORE) {
-        status[index] = *pstat;
+    if (i==index)
+      continue;
+    if (requests[i] != MPI_REQUEST_NULL) {
+      test(&requests[i], pstat,&flag);
+      if (flag==1){
+        indices[count] = i;
+        if(status != MPI_STATUSES_IGNORE) {
+          status[count] = *pstat;
+        }
+        if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
+          requests[i]=MPI_REQUEST_NULL;
+        count++;
       }
-      if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT))
-        requests[index] = MPI_REQUEST_NULL;
-    }else{
-      return MPI_UNDEFINED;
     }
   }
   return count;
@@ -992,5 +1113,77 @@ void Request::free_f(int id)
   }
 }
 
+
+int Request::get_status(MPI_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);
+    if(*flag)
+      return MPI_SUCCESS;
+  }
+  if (req != MPI_REQUEST_NULL && 
+     (req->flags_ & MPI_REQ_GENERALIZED)
+     && !(req->flags_ & MPI_REQ_COMPLETE)) {
+     *flag=0;
+    return MPI_SUCCESS;
+  }
+
+  *flag=1;
+  if(req != MPI_REQUEST_NULL &&
+     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;
+    status->count = req->real_size_;
+  }
+  return MPI_SUCCESS;
+}
+
+int Request::grequest_start( MPI_Grequest_query_function *query_fn, MPI_Grequest_free_function *free_fn, MPI_Grequest_cancel_function *cancel_fn, void *extra_state, MPI_Request *request){
+
+  *request = new Request();
+  (*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)->query_fn=query_fn;
+  ((*request)->generalized_funcs)->free_fn=free_fn;
+  ((*request)->generalized_funcs)->cancel_fn=cancel_fn;
+  ((*request)->generalized_funcs)->extra_state=extra_state;
+  ((*request)->generalized_funcs)->cond = simgrid::s4u::ConditionVariable::create();
+  ((*request)->generalized_funcs)->mutex = simgrid::s4u::Mutex::create();
+  return MPI_SUCCESS;
+}
+
+int Request::grequest_complete( MPI_Request request){
+  if ((!(request->flags_ & MPI_REQ_GENERALIZED)) || request->generalized_funcs->mutex==NULL) 
+    return MPI_ERR_REQUEST;
+  request->generalized_funcs->mutex->lock();
+  request->flags_ |= MPI_REQ_COMPLETE; // in case wait would be called after complete
+  request->generalized_funcs->cond->notify_one();
+  request->generalized_funcs->mutex->unlock();
+  return MPI_SUCCESS;
+}
+
+void Request::set_nbc_requests(MPI_Request* reqs, int size){
+  nbc_requests_size_ = size;
+  if (size > 0) {
+    nbc_requests_ = reqs;
+  } else {
+    delete[] reqs;
+    nbc_requests_ = nullptr;
+  }
+}
+
+int Request::get_nbc_requests_size(){
+  return nbc_requests_size_;
+}
+
+MPI_Request* Request::get_nbc_requests(){
+  return nbc_requests_;
+}
+
 }
 }