X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/f23be1900e1e6d0297407b0916cc37c97b1df269..15e30191fd888064abe15d164cd4dd02229fd75d:/src/smpi/mpi/smpi_request.cpp

diff --git a/src/smpi/mpi/smpi_request.cpp b/src/smpi/mpi/smpi_request.cpp
index 609e0366aa..6418c6585f 100644
--- a/src/smpi/mpi/smpi_request.cpp
+++ b/src/smpi/mpi/smpi_request.cpp
@@ -2,8 +2,7 @@
 
 /* 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 "mc/mc.h"
@@ -17,8 +16,6 @@
 #include "src/kernel/activity/CommImpl.hpp"
 #include "src/mc/mc_replay.hpp"
 #include "src/smpi/include/smpi_actor.hpp"
-#include "xbt/config.hpp"
-
 
 #include <algorithm>
 
@@ -36,14 +33,14 @@ extern void (*smpi_comm_copy_data_callback)(simgrid::kernel::activity::CommImpl*
 namespace simgrid{
 namespace smpi{
 
-Request::Request(void* buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm, unsigned flags, MPI_Op op)
-    : buf_(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), 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 {
@@ -53,7 +50,7 @@ Request::Request(void* buf, int count, MPI_Datatype datatype, int src, int dst,
       }
     }
   }
-  // This part handles the problem of non-contiguous memory (for the unserialisation at the reception)
+  // 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();
@@ -111,56 +108,49 @@ void Request::unref(MPI_Request* request)
   }
 }
 
-int Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl*)
+bool Request::match_common(MPI_Request req, MPI_Request sender, MPI_Request receiver)
+{
+  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",
+            (req == receiver ? "send" : "recv"), sender->src_, receiver->src_, sender->tag_, receiver->tag_,
+            sender->comm_->id(), receiver->comm_->id());
+
+  if ((receiver->comm_->id() == MPI_UNDEFINED || sender->comm_->id() == MPI_UNDEFINED ||
+       receiver->comm_->id() == sender->comm_->id()) &&
+      ((receiver->src_ == MPI_ANY_SOURCE && (receiver->comm_->group()->rank(sender->src_) != MPI_UNDEFINED)) ||
+       receiver->src_ == sender->src_) &&
+      ((receiver->tag_ == MPI_ANY_TAG && sender->tag_ >= 0) || receiver->tag_ == sender->tag_)) {
+    // we match, we can transfer some values
+    if (receiver->src_ == MPI_ANY_SOURCE)
+      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 (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
+    XBT_DEBUG("match succeeded");
+    return true;
+  }
+  return false;
+}
+
+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);
-  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_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_)
-    && ((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)
-      ref->real_src_ = req->src_;
-    if(ref->tag_ == MPI_ANY_TAG)
-      ref->real_tag_ = req->tag_;
-    if(ref->real_size_ < req->real_size_)
-      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
-    XBT_DEBUG("match succeeded");
-    return 1;
-  }else return 0;
+  return match_common(req, req, ref);
 }
 
-int Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl*)
+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);
-  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_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_)
-      && ((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_ = 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
-    XBT_DEBUG("match succeeded");
-    return 1;
-  } else
-    return 0;
+  return match_common(req, ref, req);
 }
 
 void Request::print_request(const char *message)
@@ -169,24 +159,29 @@ void Request::print_request(const char *message)
        message, this, buf_, size_, src_, dst_, tag_, flags_);
 }
 
-
 /* 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::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,
+                     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,
                      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,
@@ -194,7 +189,7 @@ 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 */
@@ -243,7 +238,17 @@ 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::ibsend(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_ISEND | MPI_REQ_SEND | MPI_REQ_BSEND);
+  request->start();
+  return request;
+}
+
+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(),
@@ -253,7 +258,7 @@ 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(),
@@ -282,7 +287,18 @@ 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::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);
+
+  request->start();
+  wait(&request, MPI_STATUS_IGNORE);
+  request = nullptr;
+}
+
+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(),
@@ -293,7 +309,7 @@ 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(),
@@ -305,7 +321,7 @@ 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)
 {
@@ -348,15 +364,13 @@ void Request::start()
 
     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");
-
     simgrid::s4u::MutexPtr mut = process->mailboxes_mutex();
-    if (async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0)
+    if (smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)
       mut->lock();
 
-    if (async_small_thresh == 0 && (flags_ & MPI_REQ_RMA) == 0) {
+    if (smpi_cfg_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) {
+    } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast<int>(size_) < smpi_cfg_async_small_thresh()) {
       //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();
@@ -395,10 +409,11 @@ void Request::start()
         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)
+    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_));
+    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_);
@@ -407,8 +422,8 @@ void Request::start()
 
     void* buf = buf_;
     if ((flags_ & MPI_REQ_SSEND) == 0 &&
-        ((flags_ & MPI_REQ_RMA) != 0 ||
-         static_cast<int>(size_) < simgrid::config::get_value<int>("smpi/send-is-detached-thresh"))) {
+        ((flags_ & MPI_REQ_RMA) != 0 || (flags_ & MPI_REQ_BSEND) != 0 ||
+         static_cast<int>(size_) < smpi_cfg_detached_send_thresh())) {
       void *oldbuf = nullptr;
       detached_    = true;
       XBT_DEBUG("Send request %p is detached", this);
@@ -416,12 +431,14 @@ void Request::start()
       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_cfg_privatization() != 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 ");
             smpi_switch_data_segment(simgrid::s4u::Actor::by_pid(src_));
           }
+          //we need this temporary buffer even for bsend, as it will be released in the copy callback and we don't have a way to differentiate it
+          //so actually ... don't use manually attached buffer space.
           buf = xbt_malloc(size_);
           memcpy(buf,oldbuf,size_);
           XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
@@ -439,20 +456,18 @@ void Request::start()
     }
 
     if(sleeptime > 0.0){
-      simcall_process_sleep(sleeptime);
+      simgrid::s4u::this_actor::sleep_for(sleeptime);
       XBT_DEBUG("sending size of %zu : sleep %f ", size_, sleeptime);
     }
 
-    int async_small_thresh = simgrid::config::get_value<int>("smpi/async-small-thresh");
-
     simgrid::s4u::MutexPtr mut = process->mailboxes_mutex();
 
-    if (async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0)
+    if (smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)
       mut->lock();
 
-    if (not(async_small_thresh != 0 || (flags_ & MPI_REQ_RMA) != 0)) {
+    if (not(smpi_cfg_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
+    } 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));
@@ -484,7 +499,7 @@ void Request::start()
     action_   = simcall_comm_isend(
         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,
+        process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this,
         // detach if msg size < eager/rdv switch limit
         detached_);
     XBT_DEBUG("send simcall posted");
@@ -495,7 +510,7 @@ void Request::start()
           smpi_process()->get_tracing_category());
     }
 
-    if (async_small_thresh != 0 || ((flags_ & MPI_REQ_RMA) != 0))
+    if (smpi_cfg_async_small_thresh() != 0 || ((flags_ & MPI_REQ_RMA) != 0))
       mut->unlock();
   }
 }
@@ -540,15 +555,15 @@ int Request::test(MPI_Request * request, MPI_Status * status, int* flag) {
   }
   
   if(smpi_test_sleep > 0)
-    simcall_process_sleep(nsleeps*smpi_test_sleep);
+    simgrid::s4u::this_actor::sleep_for(nsleeps * smpi_test_sleep);
 
   Status::empty(status);
   *flag = 1;
   if (((*request)->flags_ & MPI_REQ_PREPARED) == 0) {
-    if ((*request)->action_ != nullptr){
+    if ((*request)->action_ != nullptr && (*request)->cancelled_ != 1){
       try{
         *flag = simcall_comm_test((*request)->action_);
-      }catch (xbt_ex& e) {
+      } catch (const Exception&) {
         *flag = 0;
         return ret;
       }
@@ -574,7 +589,7 @@ int Request::test(MPI_Request * request, MPI_Status * status, int* flag) {
       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;
-    } else if (simgrid::config::get_value<bool>("smpi/grow-injected-times")) {
+    } else if (smpi_cfg_grow_injected_times()) {
       nsleeps++;
     }
   }
@@ -583,7 +598,6 @@ int Request::test(MPI_Request * request, MPI_Status * status, int* flag) {
 
 int Request::testsome(int incount, MPI_Request requests[], int *count, int *indices, MPI_Status status[])
 {
-  int ret = MPI_SUCCESS;
   int error=0;
   int count_dead = 0;
   int flag = 0;
@@ -592,15 +606,15 @@ int Request::testsome(int incount, MPI_Request requests[], int *count, int *indi
 
   *count = 0;
   for (int i = 0; i < incount; i++) {
-    if (requests[i] != MPI_REQUEST_NULL) {
-      ret = test(&requests[i], pstat, &flag);
+    if (requests[i] != MPI_REQUEST_NULL && not (requests[i]->flags_ & MPI_REQ_FINISHED)) {
+      int ret = test(&requests[i], pstat, &flag);
       if(ret!=MPI_SUCCESS)
         error = 1;
       if(flag) {
-        indices[i] = 1;
-        (*count)++;
+        indices[*count] = i;
         if (status != MPI_STATUSES_IGNORE)
-          status[i] = *pstat;
+          status[*count] = *pstat;
+        (*count)++;
         if ((requests[i] != MPI_REQUEST_NULL) && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
           requests[i] = MPI_REQUEST_NULL;
       }
@@ -636,10 +650,11 @@ int Request::testany(int count, MPI_Request requests[], int *index, int* flag, M
     //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
     static int nsleeps = 1;
     if(smpi_test_sleep > 0)
-      simcall_process_sleep(nsleeps*smpi_test_sleep);
+      simgrid::s4u::this_actor::sleep_for(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;
     }
     
@@ -666,6 +681,7 @@ int Request::testany(int count, MPI_Request requests[], int *index, int* flag, M
 
         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;
@@ -673,8 +689,10 @@ int Request::testany(int count, MPI_Request requests[], int *index, int* flag, M
       nsleeps++;
     }
   } else {
+      XBT_DEBUG("Testany on inactive handles, returning flag=1 but empty status");
       //all requests are null or inactive, return true
       *flag = 1;
+      *index = MPI_UNDEFINED;
       Status::empty(status);
   }
 
@@ -687,11 +705,10 @@ int Request::testall(int count, MPI_Request requests[], int* outflag, MPI_Status
   MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
   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)) {
-      ret = test(&requests[i], pstat, &flag);
+      int ret = test(&requests[i], pstat, &flag);
       if (flag){
         flag=0;
         requests[i]=MPI_REQUEST_NULL;
@@ -730,7 +747,7 @@ void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status*
   // This can speed up the execution of certain applications by an order of magnitude, such as HPL
   static int nsleeps = 1;
   double speed        = s4u::this_actor::get_host()->get_speed();
-  double maxrate      = simgrid::config::get_value<double>("smpi/iprobe-cpu-usage");
+  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);
@@ -752,8 +769,8 @@ void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status*
   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) {
+  // We have to test both mailboxes as we don't know if we will receive one or another
+  if (smpi_cfg_async_small_thresh() > 0) {
     mailbox = smpi_process()->mailbox_small();
     XBT_DEBUG("Trying to probe the perm recv mailbox");
     request->action_ = mailbox->iprobe(0, &match_recv, static_cast<void*>(request));
@@ -779,7 +796,7 @@ void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status*
   }
   else {
     *flag = 0;
-    if (simgrid::config::get_value<bool>("smpi/grow-injected-times"))
+    if (smpi_cfg_grow_injected_times())
       nsleeps++;
   }
   unref(&request);
@@ -800,8 +817,7 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status)
     return;
   }
 
-  if (not(req->detached_ && ((req->flags_ & MPI_REQ_SEND) != 0)) && ((req->flags_ & MPI_REQ_PREPARED) == 0) &&
-      ((req->flags_ & MPI_REQ_GENERALIZED) == 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);
@@ -810,32 +826,33 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status)
       // 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_cfg_privatization() != 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-contiguous 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_);
       }
     }
   }
@@ -849,8 +866,8 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status)
     //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
     double sleeptime =
         simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->orecv(req->real_size());
-    if(sleeptime > 0.0){
-      simcall_process_sleep(sleeptime);
+    if (sleeptime > 0.0) {
+      simgrid::s4u::this_actor::sleep_for(sleeptime);
       XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size_, sleeptime);
     }
     unref(&(req->detached_sender_));
@@ -878,7 +895,7 @@ int Request::wait(MPI_Request * request, MPI_Status * status)
             int count=(*request)->size_/ (*request)->old_type_->size();
             (*request)->op_->apply(buf, (*request)->buf_, &count, (*request)->old_type_);
           }
-          smpi_free_tmp_buffer(buf);
+          smpi_free_tmp_buffer(static_cast<unsigned char*>(buf));
         }
       }
       if((*request)->nbc_requests_[i]!=MPI_REQUEST_NULL)
@@ -901,7 +918,7 @@ int Request::wait(MPI_Request * request, MPI_Status * status)
       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");
       }
   }
@@ -964,8 +981,8 @@ int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
       try{
         // this is not a detached send
         i = simcall_comm_waitany(comms.data(), comms.size(), -1);
-      }catch (xbt_ex& e) {
-      XBT_INFO("request %d cancelled ",i);
+      } catch (const Exception&) {
+        XBT_INFO("request %d cancelled ", i);
         return i;
       }
 
@@ -1053,7 +1070,6 @@ 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);
   if(index==MPI_UNDEFINED) return MPI_UNDEFINED;
   if(status != MPI_STATUSES_IGNORE) {
@@ -1062,7 +1078,8 @@ int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Sta
   indices[count] = index;
   count++;
   for (int i = 0; i < incount; i++) {
-    if (requests[i] != MPI_REQUEST_NULL) {
+    if (i!=index && requests[i] != MPI_REQUEST_NULL 
+        && not(requests[i]->flags_ & MPI_REQ_FINISHED)) {
       test(&requests[i], pstat,&flag);
       if (flag==1){
         indices[count] = i;
@@ -1078,34 +1095,24 @@ int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Sta
   return count;
 }
 
-MPI_Request Request::f2c(int id) {
+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_id(key, id)));
-}
-
-int Request::add_f()
-{
-  if (F2C::f2c_lookup() == nullptr) {
-    F2C::set_f2c_lookup(new std::unordered_map<std::string, F2C*>);
-  }
-  char key[KEY_SIZE];
-  (*(F2C::f2c_lookup()))[get_key_id(key, F2C::f2c_id())] = this;
-  F2C::f2c_id_increment();
-  return F2C::f2c_id()-1;
+  return static_cast<MPI_Request>(F2C::f2c_lookup()->at(get_key(key,id)));
 }
 
 void Request::free_f(int id)
 {
   if (id != MPI_FORTRAN_REQUEST_NULL) {
     char key[KEY_SIZE];
-    F2C::f2c_lookup()->erase(get_key_id(key, id));
+    F2C::f2c_lookup()->erase(get_key(key, id));
   }
 }
 
-
-int Request::get_status(MPI_Request req, int* flag, MPI_Status * status){
+int Request::get_status(MPI_Request req, int* flag, MPI_Status* status)
+{
   *flag=0;
 
   if(req != MPI_REQUEST_NULL && req->action_ != nullptr) {
@@ -1132,8 +1139,9 @@ int Request::get_status(MPI_Request req, int* flag, MPI_Status * status){
   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){
-
+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;
@@ -1148,7 +1156,8 @@ int Request::grequest_start( MPI_Grequest_query_function *query_fn, MPI_Grequest
   return MPI_SUCCESS;
 }
 
-int Request::grequest_complete( MPI_Request request){
+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();