X-Git-Url: http://bilbo.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/b10b767ac387effc02cf42174436acab4457f9f3..89d7a404bbe64b7f19e6743d2979468f1cac42a2:/src/smpi/mpi/smpi_win.cpp diff --git a/src/smpi/mpi/smpi_win.cpp b/src/smpi/mpi/smpi_win.cpp index a7300f171b..7b33c80243 100644 --- a/src/smpi/mpi/smpi_win.cpp +++ b/src/smpi/mpi/smpi_win.cpp @@ -1,8 +1,9 @@ -/* Copyright (c) 2007-2021. The SimGrid Team. All rights reserved. */ +/* Copyright (c) 2007-2023. 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 #include "smpi_win.hpp" #include "private.hpp" @@ -13,8 +14,10 @@ #include "smpi_keyvals.hpp" #include "smpi_request.hpp" #include "src/smpi/include/smpi_actor.hpp" +#include "src/mc/mc_replay.hpp" #include +#include // std::scoped_lock XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_rma, smpi, "Logging specific to SMPI (RMA operations)"); @@ -33,8 +36,7 @@ XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_rma, smpi, "Logging specific to SMPI (RMA o return MPI_ERR_WIN; \ } -namespace simgrid{ -namespace smpi{ +namespace simgrid::smpi { std::unordered_map Win::keyvals_; int Win::keyval_id_=0; @@ -53,47 +55,54 @@ Win::Win(void* base, MPI_Aint size, int disp_unit, MPI_Info info, MPI_Comm comm, if(info!=MPI_INFO_NULL) info->ref(); connected_wins_[rank_] = this; - if(rank_==0){ - bar_ = new s4u::Barrier(comm->size()); - } errhandler_->ref(); comm->add_rma_win(this); comm->ref(); colls::allgather(&connected_wins_[rank_], sizeof(MPI_Win), MPI_BYTE, connected_wins_.data(), sizeof(MPI_Win), MPI_BYTE, comm); - - colls::bcast(&bar_, sizeof(s4u::Barrier*), MPI_BYTE, 0, comm); - - colls::barrier(comm); + if (MC_is_active() || MC_record_replay_is_active()){ + s4u::Barrier* bar_ptr; + if (rank_ == 0) { + bar_ = s4u::Barrier::create(comm->size()); + bar_ptr = bar_.get(); + } + colls::bcast(&bar_ptr, sizeof(s4u::Barrier*), MPI_BYTE, 0, comm); + if (rank_ != 0) + bar_ = s4u::BarrierPtr(bar_ptr); + } this->add_f(); } -Win::~Win(){ +int Win::del(Win* win){ //As per the standard, perform a barrier to ensure every async comm is finished - bar_->wait(); - - int finished = finish_comms(); - XBT_DEBUG("Win destructor - Finished %d RMA calls", finished); - - if (info_ != MPI_INFO_NULL) - simgrid::smpi::Info::unref(info_); - if (errhandler_ != MPI_ERRHANDLER_NULL) - simgrid::smpi::Errhandler::unref(errhandler_); - - comm_->remove_rma_win(this); - - colls::barrier(comm_); - Comm::unref(comm_); - - if (rank_ == 0) - delete bar_; + if (MC_is_active() || MC_record_replay_is_active()) + win->bar_->wait(); + else + colls::barrier(win->comm_); + win->flush_local_all(); + + if (win->info_ != MPI_INFO_NULL) + simgrid::smpi::Info::unref(win->info_); + if (win->errhandler_ != MPI_ERRHANDLER_NULL) + simgrid::smpi::Errhandler::unref(win->errhandler_); + + win->comm_->remove_rma_win(win); + + colls::barrier(win->comm_); + Comm::unref(win->comm_); + if (not win->lockers_.empty() || win->opened_ < 0) { + XBT_WARN("Freeing a locked or opened window"); + return MPI_ERR_WIN; + } + if (win->allocated_) + xbt_free(win->base_); - if (allocated_) - xbt_free(base_); + F2C::free_f(win->f2c_id()); + win->cleanup_attr(); - F2C::free_f(this->f2c_id()); - cleanup_attr(); + delete win; + return MPI_SUCCESS; } int Win::attach(void* /*base*/, MPI_Aint size) @@ -131,9 +140,6 @@ void Win::get_group(MPI_Group* group){ MPI_Info Win::info() { - if (info_ == MPI_INFO_NULL) - info_ = new Info(); - info_->ref(); return info_; } @@ -183,32 +189,24 @@ void Win::set_name(const char* name){ int Win::fence(int assert) { XBT_DEBUG("Entering fence"); - if (opened_ == 0) - opened_=1; + opened_++; if (not (assert & MPI_MODE_NOPRECEDE)) { // This is not the first fence => finalize what came before - bar_->wait(); - mut_->lock(); - // This (simulated) mutex ensures that no process pushes to the vector of requests during the waitall. - // Without this, the vector could get redimensioned when another process pushes. - // This would result in the array used by Request::waitall() to be invalidated. - // Another solution would be to copy the data and cleanup the vector *before* Request::waitall - - // start all requests that have been prepared by another process - if (not requests_.empty()) { - int size = static_cast(requests_.size()); - MPI_Request* treqs = requests_.data(); - Request::waitall(size, treqs, MPI_STATUSES_IGNORE); - } + if (MC_is_active() || MC_record_replay_is_active()) + bar_->wait(); + else + colls::barrier(comm_); + flush_local_all(); count_=0; - mut_->unlock(); } if (assert & MPI_MODE_NOSUCCEED) // there should be no ops after this one, tell we are closed. opened_=0; assert_ = assert; - - bar_->wait(); + if (MC_is_active() || MC_record_replay_is_active()) + bar_->wait(); + else + colls::barrier(comm_); XBT_DEBUG("Leaving fence"); return MPI_SUCCESS; @@ -242,16 +240,14 @@ int Win::put(const void *origin_addr, int origin_count, MPI_Datatype origin_data if(request!=nullptr){ *request=sreq; }else{ - mut_->lock(); + const std::scoped_lock lock(*mut_); requests_.push_back(sreq); - mut_->unlock(); } //push request to receiver's win - recv_win->mut_->lock(); + const std::scoped_lock recv_lock(*recv_win->mut_); recv_win->requests_.push_back(rreq); rreq->start(); - recv_win->mut_->unlock(); } else { XBT_DEBUG("Entering MPI_Put from myself to myself, rank %d", target_rank); Datatype::copy(origin_addr, origin_count, origin_datatype, recv_addr, target_count, target_datatype); @@ -285,10 +281,10 @@ int Win::get( void *origin_addr, int origin_count, MPI_Datatype origin_datatype, //start the send, with another process than us as sender. sreq->start(); - //push request to receiver's win - send_win->mut_->lock(); - send_win->requests_.push_back(sreq); - send_win->mut_->unlock(); + // push request to sender's win + if (const std::scoped_lock send_lock(*send_win->mut_); true) { + send_win->requests_.push_back(sreq); + } //start recv rreq->start(); @@ -296,9 +292,8 @@ int Win::get( void *origin_addr, int origin_count, MPI_Datatype origin_datatype, if(request!=nullptr){ *request=rreq; }else{ - mut_->lock(); + const std::scoped_lock lock(*mut_); requests_.push_back(rreq); - mut_->unlock(); } } else { Datatype::copy(send_addr, target_count, target_datatype, origin_addr, origin_count, origin_datatype); @@ -337,19 +332,21 @@ int Win::accumulate(const void *origin_addr, int origin_count, MPI_Datatype orig // start send sreq->start(); // push request to receiver's win - recv_win->mut_->lock(); - recv_win->requests_.push_back(rreq); - rreq->start(); - recv_win->mut_->unlock(); + if (const std::scoped_lock recv_lock(*recv_win->mut_); true) { + recv_win->requests_.push_back(rreq); + rreq->start(); + } if (request != nullptr) { *request = sreq; } else { - mut_->lock(); + const std::scoped_lock lock(*mut_); requests_.push_back(sreq); - mut_->unlock(); } + // FIXME: The current implementation fails to ensure the correct ordering of the accumulate requests. The following + // 'flush' is a workaround to fix that. + flush(target_rank); XBT_DEBUG("Leaving MPI_Win_Accumulate"); return MPI_SUCCESS; } @@ -366,8 +363,8 @@ int Win::get_accumulate(const void* origin_addr, int origin_count, MPI_Datatype XBT_DEBUG("Entering MPI_Get_accumulate from %d", target_rank); //need to be sure ops are correctly ordered, so finish request here ? slow. - MPI_Request req; - send_win->atomic_mut_->lock(); + MPI_Request req = MPI_REQUEST_NULL; + const std::scoped_lock lock(*send_win->atomic_mut_); get(result_addr, result_count, result_datatype, target_rank, target_disp, target_count, target_datatype, &req); if (req != MPI_REQUEST_NULL) @@ -377,7 +374,6 @@ int Win::get_accumulate(const void* origin_addr, int origin_count, MPI_Datatype target_disp, target_count, target_datatype, op, &req); if (req != MPI_REQUEST_NULL) Request::wait(&req, MPI_STATUS_IGNORE); - send_win->atomic_mut_->unlock(); return MPI_SUCCESS; } @@ -391,7 +387,7 @@ int Win::compare_and_swap(const void* origin_addr, const void* compare_addr, voi XBT_DEBUG("Entering MPI_Compare_and_swap with %d", target_rank); MPI_Request req = MPI_REQUEST_NULL; - send_win->atomic_mut_->lock(); + const std::scoped_lock lock(*send_win->atomic_mut_); get(result_addr, 1, datatype, target_rank, target_disp, 1, datatype, &req); if (req != MPI_REQUEST_NULL) @@ -400,7 +396,6 @@ int Win::compare_and_swap(const void* origin_addr, const void* compare_addr, voi put(origin_addr, 1, datatype, target_rank, target_disp, 1, datatype); } - send_win->atomic_mut_->unlock(); return MPI_SUCCESS; } @@ -436,7 +431,7 @@ int Win::start(MPI_Group group, int /*assert*/) group->ref(); dst_group_ = group; - opened_++; // we're open for business ! + opened_--; // we're open for business ! XBT_DEBUG("Leaving MPI_Win_Start"); return MPI_SUCCESS; } @@ -461,7 +456,7 @@ int Win::post(MPI_Group group, int /*assert*/) group->ref(); src_group_ = group; - opened_++; // we're open for business ! + opened_--; // we're open for business ! XBT_DEBUG("Leaving MPI_Win_Post"); return MPI_SUCCESS; } @@ -485,10 +480,9 @@ int Win::complete(){ for (auto& req : reqs) Request::unref(&req); - int finished = finish_comms(); - XBT_DEBUG("Win_complete - Finished %d RMA calls", finished); + flush_local_all(); - opened_--; //we're closed for business ! + opened_++; //we're closed for business ! Group::unref(dst_group_); dst_group_ = MPI_GROUP_NULL; return MPI_SUCCESS; @@ -512,10 +506,9 @@ int Win::wait(){ for (auto& req : reqs) Request::unref(&req); - int finished = finish_comms(); - XBT_DEBUG("Win_wait - Finished %d RMA calls", finished); + flush_local_all(); - opened_--; //we're closed for business ! + opened_++; //we're closed for business ! Group::unref(src_group_); src_group_ = MPI_GROUP_NULL; return MPI_SUCCESS; @@ -536,10 +529,7 @@ int Win::lock(int lock_type, int rank, int /*assert*/) target_win->lockers_.push_back(rank_); - int finished = finish_comms(rank); - XBT_DEBUG("Win_lock %d - Finished %d RMA calls", rank, finished); - finished = target_win->finish_comms(rank_); - XBT_DEBUG("Win_lock target %d - Finished %d RMA calls", rank, finished); + flush(rank); return MPI_SUCCESS; } @@ -562,10 +552,7 @@ int Win::unlock(int rank){ target_win->lock_mut_->unlock(); } - int finished = finish_comms(rank); - XBT_DEBUG("Win_unlock %d - Finished %d RMA calls", rank, finished); - finished = target_win->finish_comms(rank_); - XBT_DEBUG("Win_unlock target %d - Finished %d RMA calls", rank, finished); + flush(rank); return MPI_SUCCESS; } @@ -580,33 +567,36 @@ int Win::unlock_all(){ } int Win::flush(int rank){ - MPI_Win target_win = connected_wins_[rank]; - int finished = finish_comms(rank_); - XBT_DEBUG("Win_flush on local %d - Finished %d RMA calls", rank_, finished); - finished = target_win->finish_comms(rank); - XBT_DEBUG("Win_flush on remote %d - Finished %d RMA calls", rank, finished); + int finished = finish_comms(rank); + XBT_DEBUG("Win_flush on local %d for remote %d - Finished %d RMA calls", rank_, rank, finished); + if (rank != rank_) { + finished = connected_wins_[rank]->finish_comms(rank_); + XBT_DEBUG("Win_flush on remote %d for local %d - Finished %d RMA calls", rank, rank_, finished); + } return MPI_SUCCESS; } int Win::flush_local(int rank){ int finished = finish_comms(rank); - XBT_DEBUG("Win_flush_local for rank %d - Finished %d RMA calls", rank, finished); + XBT_DEBUG("Win_flush_local on local %d for remote %d - Finished %d RMA calls", rank_, rank, finished); return MPI_SUCCESS; } int Win::flush_all(){ int finished = finish_comms(); - XBT_DEBUG("Win_flush_all on local - Finished %d RMA calls", finished); + XBT_DEBUG("Win_flush_all on local %d - Finished %d RMA calls", rank_, finished); for (int i = 0; i < comm_->size(); i++) { - finished = connected_wins_[i]->finish_comms(rank_); - XBT_DEBUG("Win_flush_all on %d - Finished %d RMA calls", i, finished); + if (i != rank_) { + finished = connected_wins_[i]->finish_comms(rank_); + XBT_DEBUG("Win_flush_all on remote %d for local %d - Finished %d RMA calls", i, rank_, finished); + } } return MPI_SUCCESS; } int Win::flush_local_all(){ int finished = finish_comms(); - XBT_DEBUG("Win_flush_local_all - Finished %d RMA calls", finished); + XBT_DEBUG("Win_flush_local_all on local %d - Finished %d RMA calls", rank_, finished); return MPI_SUCCESS; } @@ -615,7 +605,11 @@ Win* Win::f2c(int id){ } int Win::finish_comms(){ - mut_->lock(); + // This (simulated) mutex ensures that no process pushes to the vector of requests during the waitall. + // Without this, the vector could get redimensioned when another process pushes. + // This would result in the array used by Request::waitall() to be invalidated. + // Another solution would be to copy the data and cleanup the vector *before* Request::waitall + const std::scoped_lock lock(*mut_); //Finish own requests int size = static_cast(requests_.size()); if (size > 0) { @@ -623,12 +617,12 @@ int Win::finish_comms(){ Request::waitall(size, treqs, MPI_STATUSES_IGNORE); requests_.clear(); } - mut_->unlock(); return size; } int Win::finish_comms(int rank){ - mut_->lock(); + // See comment about the mutex in finish_comms() above + const std::scoped_lock lock(*mut_); // Finish own requests // Let's see if we're either the destination or the sender of this request // because we only wait for requests that we are responsible for. @@ -646,7 +640,6 @@ int Win::finish_comms(int rank){ Request::waitall(size, treqs, MPI_STATUSES_IGNORE); myreqqs.clear(); } - mut_->unlock(); return size; } @@ -683,5 +676,4 @@ void Win::set_errhandler(MPI_Errhandler errhandler) if (errhandler_ != MPI_ERRHANDLER_NULL) errhandler_->ref(); } -} // namespace smpi -} // namespace simgrid +} // namespace simgrid::smpi