1 /* Copyright (c) 2007-2019. The SimGrid Team. All rights reserved. */
3 /* This program is free software; you can redistribute it and/or modify it
4 * under the terms of the license (GNU LGPL) which comes with this package. */
7 #include "smpi_comm.hpp"
8 #include "smpi_coll.hpp"
9 #include "smpi_datatype.hpp"
10 #include "smpi_info.hpp"
11 #include "smpi_win.hpp"
12 #include "smpi_request.hpp"
14 #include "smpi_file.hpp"
15 #include "smpi_status.hpp"
16 #include "simgrid/plugins/file_system.h"
18 #define FP_SIZE sizeof(MPI_Offset)
20 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_io, smpi, "Logging specific to SMPI (RMA operations)");
22 MPI_Errhandler SMPI_default_File_Errhandler = MPI_ERRORS_RETURN;
27 File::File(MPI_Comm comm, const char *filename, int amode, MPI_Info info): comm_(comm), flags_(amode), info_(info) {
28 file_= new simgrid::s4u::File(filename, nullptr);
30 if (comm_->rank() == 0) {
31 int size= comm_->size() + FP_SIZE;
32 list_ = new char[size];
33 errhandler_= SMPI_default_File_Errhandler;
34 memset(list_, 0, size);
35 shared_file_pointer_ = new MPI_Offset();
36 shared_mutex_ = s4u::Mutex::create();
37 *shared_file_pointer_ = 0;
38 win_=new Win(list_, size, 1, MPI_INFO_NULL, comm_);
40 win_=new Win(list_, 0, 1, MPI_INFO_NULL, comm_);
42 simgrid::smpi::Colls::bcast(&shared_file_pointer_, 1, MPI_AINT, 0, comm);
43 simgrid::smpi::Colls::bcast(&shared_mutex_, 1, MPI_AINT, 0, comm);
44 if(comm_->rank() != 0)
45 intrusive_ptr_add_ref(&*shared_mutex_);
49 if(comm_->rank() == 0){
50 delete shared_file_pointer_;
57 int File::close(MPI_File *fh){
58 XBT_DEBUG("Closing MPI_File %s", (*fh)->file_->get_path());
60 if((*fh)->flags() & MPI_MODE_DELETE_ON_CLOSE)
61 (*fh)->file_->unlink();
66 int File::del(const char* filename, MPI_Info)
68 //get the file with MPI_MODE_DELETE_ON_CLOSE and then close it
69 File* f = new File(MPI_COMM_SELF,filename,MPI_MODE_DELETE_ON_CLOSE|MPI_MODE_RDWR, nullptr);
74 int File::get_position(MPI_Offset* offset){
75 *offset=file_->tell();
79 int File::get_position_shared(MPI_Offset* offset){
80 shared_mutex_->lock();
81 *offset=*shared_file_pointer_;
82 shared_mutex_->unlock();
86 int File::seek(MPI_Offset offset, int whence){
89 XBT_VERB("Seeking in MPI_File %s, setting offset %lld", file_->get_path(), offset);
90 file_->seek(offset,SEEK_SET);
93 XBT_VERB("Seeking in MPI_File %s, current offset + %lld", file_->get_path(), offset);
94 file_->seek(offset,SEEK_CUR);
97 XBT_VERB("Seeking in MPI_File %s, end offset + %lld", file_->get_path(), offset);
98 file_->seek(offset,SEEK_END);
106 int File::seek_shared(MPI_Offset offset, int whence){
107 shared_mutex_->lock();
109 *shared_file_pointer_=offset;
110 shared_mutex_->unlock();
114 int File::read(MPI_File fh, void* /*buf*/, int count, MPI_Datatype datatype, MPI_Status* status)
116 //get position first as we may be doing non contiguous reads and it will probably be updated badly
117 MPI_Offset position = fh->file_->tell();
118 MPI_Offset movesize = datatype->get_extent()*count;
119 MPI_Offset readsize = datatype->size()*count;
120 XBT_DEBUG("Position before read in MPI_File %s : %llu",fh->file_->get_path(),fh->file_->tell());
121 MPI_Offset read = fh->file_->read(readsize);
122 XBT_VERB("Read in MPI_File %s, %lld bytes read, readsize %lld bytes, movesize %lld", fh->file_->get_path(), read, readsize, movesize);
123 if(readsize!=movesize){
124 fh->file_->seek(position+movesize, SEEK_SET);
126 XBT_VERB("Position after read in MPI_File %s : %llu",fh->file_->get_path(), fh->file_->tell());
127 status->count=count*datatype->size();
131 /*Ordered and Shared Versions, with RMA-based locks : Based on the model described in :*/
132 /* @InProceedings{10.1007/11557265_15,*/
133 /* author="Latham, Robert and Ross, Robert and Thakur, Rajeev and Toonen, Brian",*/
134 /* title="Implementing MPI-IO Shared File Pointers Without File System Support",*/
135 /* booktitle="Recent Advances in Parallel Virtual Machine and Message Passing Interface",*/
137 /* publisher="Springer Berlin Heidelberg",*/
138 /* address="Berlin, Heidelberg",*/
141 int File::read_shared(MPI_File fh, void *buf, int count, MPI_Datatype datatype, MPI_Status *status){
142 fh->shared_mutex_->lock();
143 fh->seek(*(fh->shared_file_pointer_),MPI_SEEK_SET);
144 read(fh, buf, count, datatype, status);
145 *(fh->shared_file_pointer_)=fh->file_->tell();
146 fh->shared_mutex_->unlock();
150 int File::read_ordered(MPI_File fh, void *buf, int count, MPI_Datatype datatype, MPI_Status *status){
151 //0 needs to get the shared pointer value
153 if(fh->comm_->rank()==0){
154 val=*(fh->shared_file_pointer_);
156 val=count*datatype->size();
160 simgrid::smpi::Colls::scan(&val, &result, 1, MPI_OFFSET, MPI_SUM, fh->comm_);
161 fh->seek(result, MPI_SEEK_SET);
162 int ret = fh->op_all<simgrid::smpi::File::read>(buf, count, datatype, status);
163 if(fh->comm_->rank()==fh->comm_->size()-1){
164 fh->shared_mutex_->lock();
165 *(fh->shared_file_pointer_)=fh->file_->tell();
166 fh->shared_mutex_->unlock();
169 simgrid::smpi::Colls::bcast(&c, 1, MPI_BYTE, fh->comm_->size()-1, fh->comm_);
173 int File::write(MPI_File fh, void* /*buf*/, int count, MPI_Datatype datatype, MPI_Status* status)
175 //get position first as we may be doing non contiguous reads and it will probably be updated badly
176 MPI_Offset position = fh->file_->tell();
177 MPI_Offset movesize = datatype->get_extent()*count;
178 MPI_Offset writesize = datatype->size()*count;
179 XBT_DEBUG("Position before write in MPI_File %s : %llu",fh->file_->get_path(),fh->file_->tell());
180 MPI_Offset write = fh->file_->write(writesize, 1);
181 XBT_VERB("Write in MPI_File %s, %lld bytes written, readsize %lld bytes, movesize %lld", fh->file_->get_path(), write, writesize, movesize);
182 if(writesize!=movesize){
183 fh->file_->seek(position+movesize, SEEK_SET);
185 XBT_VERB("Position after write in MPI_File %s : %llu",fh->file_->get_path(), fh->file_->tell());
186 status->count=count*datatype->size();
190 int File::write_shared(MPI_File fh, const void *buf, int count, MPI_Datatype datatype, MPI_Status *status){
191 fh->shared_mutex_->lock();
192 fh->seek(*(fh->shared_file_pointer_),MPI_SEEK_SET);
193 write(fh, const_cast<void*>(buf), count, datatype, status);
194 *(fh->shared_file_pointer_)=fh->file_->tell();
195 fh->shared_mutex_->unlock();
199 int File::write_ordered(MPI_File fh, const void *buf, int count, MPI_Datatype datatype, MPI_Status *status){
200 //0 needs to get the shared pointer value
202 if(fh->comm_->rank()==0){
203 val=*(fh->shared_file_pointer_);
205 val=count*datatype->size();
208 simgrid::smpi::Colls::scan(&val, &result, 1, MPI_OFFSET, MPI_SUM, fh->comm_);
209 fh->seek(result, MPI_SEEK_SET);
210 int ret = fh->op_all<simgrid::smpi::File::write>(const_cast<void*>(buf), count, datatype, status);
211 if(fh->comm_->rank()==fh->comm_->size()-1){
212 fh->shared_mutex_->lock();
213 *(fh->shared_file_pointer_)=fh->file_->tell();
214 fh->shared_mutex_->unlock();
217 simgrid::smpi::Colls::bcast(&c, 1, MPI_BYTE, fh->comm_->size()-1, fh->comm_);
222 return file_->size();
231 return simgrid::smpi::Colls::barrier(comm_);
234 MPI_Info File::info(){
235 if(info_== MPI_INFO_NULL)
241 void File::set_info(MPI_Info info){
242 if(info_!= MPI_INFO_NULL)
247 MPI_Comm File::comm(){
251 MPI_Errhandler File::errhandler(){
255 void File::set_errhandler(MPI_Errhandler errhandler){
256 errhandler_=errhandler;
257 if(errhandler_!= MPI_ERRHANDLER_NULL)