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)");
25 File::File(MPI_Comm comm, const char *filename, int amode, MPI_Info info): comm_(comm), flags_(amode), info_(info) {
26 file_= new simgrid::s4u::File(filename, nullptr);
28 if (comm_->rank() == 0) {
29 int size= comm_->size() + FP_SIZE;
30 list_ = new char[size];
31 errhandler_=MPI_ERRORS_RETURN;
32 memset(list_, 0, size);
33 shared_file_pointer_ = new MPI_Offset();
34 shared_mutex_ = s4u::Mutex::create();
35 *shared_file_pointer_ = 0;
36 win_=new Win(list_, size, 1, MPI_INFO_NULL, comm_);
38 win_=new Win(list_, 0, 1, MPI_INFO_NULL, comm_);
40 simgrid::smpi::Colls::bcast(&shared_file_pointer_, 1, MPI_AINT, 0, comm);
41 simgrid::smpi::Colls::bcast(&shared_mutex_, 1, MPI_AINT, 0, comm);
42 if(comm_->rank() != 0)
43 intrusive_ptr_add_ref(&*shared_mutex_);
47 if(comm_->rank() == 0){
48 delete shared_file_pointer_;
55 int File::close(MPI_File *fh){
56 XBT_DEBUG("Closing MPI_File %s", (*fh)->file_->get_path());
58 if((*fh)->flags() & MPI_MODE_DELETE_ON_CLOSE)
59 (*fh)->file_->unlink();
64 int File::del(const char* filename, MPI_Info)
66 //get the file with MPI_MODE_DELETE_ON_CLOSE and then close it
67 File* f = new File(MPI_COMM_SELF,filename,MPI_MODE_DELETE_ON_CLOSE|MPI_MODE_RDWR, nullptr);
72 int File::get_position(MPI_Offset* offset){
73 *offset=file_->tell();
77 int File::get_position_shared(MPI_Offset* offset){
78 shared_mutex_->lock();
79 *offset=*shared_file_pointer_;
80 shared_mutex_->unlock();
84 int File::seek(MPI_Offset offset, int whence){
87 XBT_VERB("Seeking in MPI_File %s, setting offset %lld", file_->get_path(), offset);
88 file_->seek(offset,SEEK_SET);
91 XBT_VERB("Seeking in MPI_File %s, current offset + %lld", file_->get_path(), offset);
92 file_->seek(offset,SEEK_CUR);
95 XBT_VERB("Seeking in MPI_File %s, end offset + %lld", file_->get_path(), offset);
96 file_->seek(offset,SEEK_END);
104 int File::seek_shared(MPI_Offset offset, int whence){
105 shared_mutex_->lock();
107 *shared_file_pointer_=offset;
108 shared_mutex_->unlock();
112 int File::read(MPI_File fh, void* /*buf*/, int count, MPI_Datatype datatype, MPI_Status* status)
114 //get position first as we may be doing non contiguous reads and it will probably be updated badly
115 MPI_Offset position = fh->file_->tell();
116 MPI_Offset movesize = datatype->get_extent()*count;
117 MPI_Offset readsize = datatype->size()*count;
118 XBT_DEBUG("Position before read in MPI_File %s : %llu",fh->file_->get_path(),fh->file_->tell());
119 MPI_Offset read = fh->file_->read(readsize);
120 XBT_VERB("Read in MPI_File %s, %lld bytes read, readsize %lld bytes, movesize %lld", fh->file_->get_path(), read, readsize, movesize);
121 if(readsize!=movesize){
122 fh->file_->seek(position+movesize, SEEK_SET);
124 XBT_VERB("Position after read in MPI_File %s : %llu",fh->file_->get_path(), fh->file_->tell());
125 status->count=count*datatype->size();
129 /*Ordered and Shared Versions, with RMA-based locks : Based on the model described in :*/
130 /* @InProceedings{10.1007/11557265_15,*/
131 /* author="Latham, Robert and Ross, Robert and Thakur, Rajeev and Toonen, Brian",*/
132 /* title="Implementing MPI-IO Shared File Pointers Without File System Support",*/
133 /* booktitle="Recent Advances in Parallel Virtual Machine and Message Passing Interface",*/
135 /* publisher="Springer Berlin Heidelberg",*/
136 /* address="Berlin, Heidelberg",*/
139 int File::read_shared(MPI_File fh, void *buf, int count, MPI_Datatype datatype, MPI_Status *status){
140 fh->shared_mutex_->lock();
141 fh->seek(*(fh->shared_file_pointer_),MPI_SEEK_SET);
142 read(fh, buf, count, datatype, status);
143 *(fh->shared_file_pointer_)=fh->file_->tell();
144 fh->shared_mutex_->unlock();
148 int File::read_ordered(MPI_File fh, void *buf, int count, MPI_Datatype datatype, MPI_Status *status){
149 //0 needs to get the shared pointer value
151 if(fh->comm_->rank()==0){
152 val=*(fh->shared_file_pointer_);
154 val=count*datatype->size();
158 simgrid::smpi::Colls::scan(&val, &result, 1, MPI_OFFSET, MPI_SUM, fh->comm_);
159 fh->seek(result, MPI_SEEK_SET);
160 int ret = fh->op_all<simgrid::smpi::File::read>(buf, count, datatype, status);
161 if(fh->comm_->rank()==fh->comm_->size()-1){
162 fh->shared_mutex_->lock();
163 *(fh->shared_file_pointer_)=fh->file_->tell();
164 fh->shared_mutex_->unlock();
167 simgrid::smpi::Colls::bcast(&c, 1, MPI_BYTE, fh->comm_->size()-1, fh->comm_);
171 int File::write(MPI_File fh, void* /*buf*/, int count, MPI_Datatype datatype, MPI_Status* status)
173 //get position first as we may be doing non contiguous reads and it will probably be updated badly
174 MPI_Offset position = fh->file_->tell();
175 MPI_Offset movesize = datatype->get_extent()*count;
176 MPI_Offset writesize = datatype->size()*count;
177 XBT_DEBUG("Position before write in MPI_File %s : %llu",fh->file_->get_path(),fh->file_->tell());
178 MPI_Offset write = fh->file_->write(writesize, 1);
179 XBT_VERB("Write in MPI_File %s, %lld bytes written, readsize %lld bytes, movesize %lld", fh->file_->get_path(), write, writesize, movesize);
180 if(writesize!=movesize){
181 fh->file_->seek(position+movesize, SEEK_SET);
183 XBT_VERB("Position after write in MPI_File %s : %llu",fh->file_->get_path(), fh->file_->tell());
184 status->count=count*datatype->size();
188 int File::write_shared(MPI_File fh, const void *buf, int count, MPI_Datatype datatype, MPI_Status *status){
189 fh->shared_mutex_->lock();
190 fh->seek(*(fh->shared_file_pointer_),MPI_SEEK_SET);
191 write(fh, const_cast<void*>(buf), count, datatype, status);
192 *(fh->shared_file_pointer_)=fh->file_->tell();
193 fh->shared_mutex_->unlock();
197 int File::write_ordered(MPI_File fh, const void *buf, int count, MPI_Datatype datatype, MPI_Status *status){
198 //0 needs to get the shared pointer value
200 if(fh->comm_->rank()==0){
201 val=*(fh->shared_file_pointer_);
203 val=count*datatype->size();
206 simgrid::smpi::Colls::scan(&val, &result, 1, MPI_OFFSET, MPI_SUM, fh->comm_);
207 fh->seek(result, MPI_SEEK_SET);
208 int ret = fh->op_all<simgrid::smpi::File::write>(const_cast<void*>(buf), count, datatype, status);
209 if(fh->comm_->rank()==fh->comm_->size()-1){
210 fh->shared_mutex_->lock();
211 *(fh->shared_file_pointer_)=fh->file_->tell();
212 fh->shared_mutex_->unlock();
215 simgrid::smpi::Colls::bcast(&c, 1, MPI_BYTE, fh->comm_->size()-1, fh->comm_);
220 return file_->size();
229 return simgrid::smpi::Colls::barrier(comm_);
232 MPI_Info File::info(){
233 if(info_== MPI_INFO_NULL)
239 void File::set_info(MPI_Info info){
240 if(info_!= MPI_INFO_NULL)
245 MPI_Comm File::comm(){
249 MPI_Errhandler File::errhandler(){
253 void File::set_errhandler(MPI_Errhandler errhandler){
254 errhandler_=errhandler;
255 if(errhandler_!= MPI_ERRHANDLER_NULL)