smpi_comm_copy_data_callback = callback;
}
+static void print(std::vector<std::pair<size_t, size_t>> vec) {
+ fprintf(stderr, "{");
+ for(auto elt: vec) {
+ fprintf(stderr, "(0x%lx, 0x%lx),", elt.first, elt.second);
+ }
+ fprintf(stderr, "}\n");
+}
+static void memcpy_private(void *dest, const void *src, size_t n, std::vector<std::pair<size_t, size_t>> &private_blocks) {
+ for(auto block : private_blocks) {
+ memcpy((uint8_t*)dest+block.first, (uint8_t*)src+block.first, block.second-block.first);
+ }
+}
+
+static void check_blocks(std::vector<std::pair<size_t, size_t>> &private_blocks, size_t buff_size) {
+ for(auto block : private_blocks) {
+ xbt_assert(block.first <= block.second && block.second <= buff_size, "Oops, bug in shared malloc.");
+ }
+}
+
void smpi_comm_copy_buffer_callback(smx_activity_t synchro, void *buff, size_t buff_size)
{
-
simgrid::kernel::activity::Comm *comm = dynamic_cast<simgrid::kernel::activity::Comm*>(synchro);
-
+ int src_shared=0, dst_shared=0;
+ size_t src_offset=0, dst_offset=0;
+ std::vector<std::pair<size_t, size_t>> src_private_blocks;
+ std::vector<std::pair<size_t, size_t>> dst_private_blocks;
XBT_DEBUG("Copy the data over");
- if(smpi_is_shared(buff)){
+ if((src_shared=smpi_is_shared(buff, src_private_blocks, &src_offset))) {
XBT_DEBUG("Sender %p is shared. Let's ignore it.", buff);
- }else if(smpi_is_shared((char*)comm->dst_buff)){
+ src_private_blocks = shift_and_frame_private_blocks(src_private_blocks, src_offset, buff_size);
+ }
+ else {
+ src_private_blocks.clear();
+ src_private_blocks.push_back(std::make_pair(0, buff_size));
+ }
+ if((dst_shared=smpi_is_shared((char*)comm->dst_buff, dst_private_blocks, &dst_offset))) {
XBT_DEBUG("Receiver %p is shared. Let's ignore it.", (char*)comm->dst_buff);
- }else{
- void* tmpbuff=buff;
- if((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) && (static_cast<char*>(buff) >= smpi_start_data_exe)
- && (static_cast<char*>(buff) < smpi_start_data_exe + smpi_size_data_exe )
- ){
- XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
-
- smpi_switch_data_segment(
- (static_cast<simgrid::smpi::Process*>((static_cast<simgrid::MsgActorExt*>(comm->src_proc->data)->data))->index()));
- tmpbuff = static_cast<void*>(xbt_malloc(buff_size));
- memcpy(tmpbuff, buff, buff_size);
- }
-
- if((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) && ((char*)comm->dst_buff >= smpi_start_data_exe)
- && ((char*)comm->dst_buff < smpi_start_data_exe + smpi_size_data_exe )){
- XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
- smpi_switch_data_segment(
- (static_cast<simgrid::smpi::Process*>((static_cast<simgrid::MsgActorExt*>(comm->dst_proc->data)->data))->index()));
- }
-
- XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff,comm->dst_buff);
- memcpy(comm->dst_buff, tmpbuff, buff_size);
+ dst_private_blocks = shift_and_frame_private_blocks(dst_private_blocks, dst_offset, buff_size);
+ }
+ else {
+ dst_private_blocks.clear();
+ dst_private_blocks.push_back(std::make_pair(0, buff_size));
+ }
+/*
+ fprintf(stderr, "size: 0x%x\n", buff_size);
+ fprintf(stderr, "src: ");
+ print(src_private_blocks);
+ fprintf(stderr, "src_offset = 0x%x\n", src_offset);
+ fprintf(stderr, "dst: ");
+ print(dst_private_blocks);
+ fprintf(stderr, "dst_offset = 0x%x\n", dst_offset);
+*/
+ check_blocks(src_private_blocks, buff_size);
+ check_blocks(dst_private_blocks, buff_size);
+ auto private_blocks = merge_private_blocks(src_private_blocks, dst_private_blocks);
+/*
+ fprintf(stderr, "Private blocks: ");
+ print(private_blocks);
+*/
+ check_blocks(private_blocks, buff_size);
+ void* tmpbuff=buff;
- if((smpi_privatize_global_variables) && (static_cast<char*>(buff) >= smpi_start_data_exe)
++ if((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) && (static_cast<char*>(buff) >= smpi_start_data_exe)
+ && (static_cast<char*>(buff) < smpi_start_data_exe + smpi_size_data_exe )
+ ){
+ XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
+
+ smpi_switch_data_segment(
+ (static_cast<simgrid::smpi::Process*>((static_cast<simgrid::MsgActorExt*>(comm->src_proc->data)->data))->index()));
+ tmpbuff = static_cast<void*>(xbt_malloc(buff_size));
+ memcpy_private(tmpbuff, buff, buff_size, private_blocks);
+ }
- if((smpi_privatize_global_variables) && ((char*)comm->dst_buff >= smpi_start_data_exe)
- if (comm->detached) {
- // if this is a detached send, the source buffer was duplicated by SMPI
- // sender to make the original buffer available to the application ASAP
- xbt_free(buff);
- //It seems that the request is used after the call there this should be free somewhere else but where???
- //xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free
- comm->src_buff = nullptr;
- }
- if(tmpbuff!=buff)xbt_free(tmpbuff);
++ if((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) && ((char*)comm->dst_buff >= smpi_start_data_exe)
+ && ((char*)comm->dst_buff < smpi_start_data_exe + smpi_size_data_exe )){
+ XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
+ smpi_switch_data_segment(
+ (static_cast<simgrid::smpi::Process*>((static_cast<simgrid::MsgActorExt*>(comm->dst_proc->data)->data))->index()));
+ }
-
+ XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff,comm->dst_buff);
+ memcpy_private(comm->dst_buff, tmpbuff, buff_size, private_blocks);
+
+ if (comm->detached) {
+ // if this is a detached send, the source buffer was duplicated by SMPI
+ // sender to make the original buffer available to the application ASAP
+ xbt_free(buff);
+ //It seems that the request is used after the call there this should be free somewhere else but where???
+ //xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free
+ comm->src_buff = nullptr;
}
+ if(tmpbuff!=buff)xbt_free(tmpbuff);
}
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)
{
void *old_buf = nullptr;
- if(((((flags & RECV) != 0) && ((flags & ACCUMULATE) !=0)) || (datatype->flags() & DT_FLAG_DERIVED)) && (!smpi_is_shared(buf_))){
+// FIXME Handle the case of a partial shared malloc.
+ if(((((flags & RECV) != 0) && ((flags & ACCUMULATE) !=0)) || (datatype->flags() & DT_FLAG_DERIVED))) { // && (!smpi_is_shared(buf_))){
// This part handles the problem of non-contiguous memory
old_buf = buf;
- buf_ = count==0 ? nullptr : xbt_malloc(count*datatype->size());
- if ((datatype->flags() & DT_FLAG_DERIVED) && ((flags & SEND) != 0)) {
- datatype->serialize(old_buf, buf_, count);
+ if (count==0){
+ buf_ = nullptr;
+ }else {
+ buf_ = xbt_malloc(count*datatype->size());
+ if ((datatype->flags() & DT_FLAG_DERIVED) && ((flags & SEND) != 0)) {
+ datatype->serialize(old_buf, buf_, count);
+ }
}
}
// This part handles the problem of non-contiguous memory (for the unserialisation at the reception)
req->print_request("Finishing");
MPI_Datatype datatype = req->old_type_;
- if((((req->flags_ & ACCUMULATE) != 0) || (datatype->flags() & DT_FLAG_DERIVED)) && (!smpi_is_shared(req->old_buf_))){
+// FIXME Handle the case of a partial shared malloc.
+ if((((req->flags_ & ACCUMULATE) != 0) || (datatype->flags() & DT_FLAG_DERIVED))){// && (!smpi_is_shared(req->old_buf_))){
if (!smpi_process()->replaying()){
- if( smpi_privatize_global_variables != 0 && (static_cast<char*>(req->old_buf_) >= smpi_start_data_exe)
+ if( smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP && (static_cast<char*>(req->old_buf_) >= smpi_start_data_exe)
&& ((char*)req->old_buf_ < smpi_start_data_exe + smpi_size_data_exe )){
XBT_VERB("Privatization : We are unserializing to a zone in global memory Switch data segment ");
smpi_switch_data_segment(smpi_process()->index());
