X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/1ba3b199ce47dca5644315bf6009a795d3816ffd..1b3ef52aafb46cf06757353f5a85f3aa2f5c38a2:/src/smpi/mpi/smpi_request.cpp diff --git a/src/smpi/mpi/smpi_request.cpp b/src/smpi/mpi/smpi_request.cpp index 6711a16797..77885e4bc6 100644 --- a/src/smpi/mpi/smpi_request.cpp +++ b/src/smpi/mpi/smpi_request.cpp @@ -1,4 +1,4 @@ -/* Copyright (c) 2007-2021. The SimGrid Team. All rights reserved. */ +/* Copyright (c) 2007-2022. 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. */ @@ -8,14 +8,16 @@ #include "mc/mc.h" #include "private.hpp" #include "simgrid/Exception.hpp" +#include "simgrid/s4u/ConditionVariable.hpp" #include "simgrid/s4u/Exec.hpp" #include "simgrid/s4u/Mutex.hpp" -#include "simgrid/s4u/ConditionVariable.hpp" #include "smpi_comm.hpp" #include "smpi_datatype.hpp" #include "smpi_host.hpp" #include "smpi_op.hpp" #include "src/kernel/activity/CommImpl.hpp" +#include "src/kernel/actor/ActorImpl.hpp" +#include "src/kernel/actor/SimcallObserver.hpp" #include "src/mc/mc_replay.hpp" #include "src/smpi/include/smpi_actor.hpp" @@ -31,7 +33,7 @@ static simgrid::config::Flag smpi_test_sleep( std::vector smpi_ois_values; -extern void (*smpi_comm_copy_data_callback)(simgrid::kernel::activity::CommImpl*, void*, size_t); +extern std::function smpi_comm_copy_data_callback; namespace simgrid{ namespace smpi{ @@ -69,6 +71,7 @@ Request::Request(const void* buf, int count, MPI_Datatype datatype, aid_t src, a refcount_ = 1; else refcount_ = 0; + message_id_ = 0; init_buffer(count); this->add_f(); } @@ -142,13 +145,10 @@ bool Request::match_common(MPI_Request req, MPI_Request sender, MPI_Request rece } if (receiver->tag_ == MPI_ANY_TAG) receiver->real_tag_ = sender->tag_; - if ((receiver->flags_ & MPI_REQ_PROBE) == 0 ){ - if (receiver->real_size_ < sender->real_size_){ - XBT_DEBUG("Truncating message - should not happen: receiver size : %zu < sender size : %zu", receiver->real_size_, sender->real_size_); - receiver->truncated_ = true; - } else if (receiver->real_size_ > sender->real_size_){ - receiver->real_size_=sender->real_size_; - } + if ((receiver->flags_ & MPI_REQ_PROBE) == 0 && receiver->real_size_ < sender->real_size_) { + XBT_DEBUG("Truncating message - should not happen: receiver size : %zu < sender size : %zu", receiver->real_size_, + sender->real_size_); + receiver->truncated_ = true; } //0-sized datatypes/counts should not interfere and match if ( sender->real_size_ != 0 && receiver->real_size_ != 0 && @@ -186,7 +186,35 @@ bool Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl*) { auto ref = static_cast(a); auto req = static_cast(b); - return match_common(req, req, ref); + bool match = match_common(req, req, ref); + if (not match || ref->comm_ == MPI_COMM_UNINITIALIZED || ref->comm_->is_smp_comm()) + return match; + + if (ref->comm_->get_received_messages_count(ref->comm_->group()->rank(req->src_), + ref->comm_->group()->rank(req->dst_), req->tag_) == req->message_id_) { + if (((ref->flags_ & MPI_REQ_PROBE) == 0) && ((req->flags_ & MPI_REQ_PROBE) == 0)) { + XBT_DEBUG("increasing count in comm %p, which was %u from pid %ld, to pid %ld with tag %d", ref->comm_, + ref->comm_->get_received_messages_count(ref->comm_->group()->rank(req->src_), + ref->comm_->group()->rank(req->dst_), req->tag_), + req->src_, req->dst_, req->tag_); + ref->comm_->increment_received_messages_count(ref->comm_->group()->rank(req->src_), + ref->comm_->group()->rank(req->dst_), req->tag_); + if (ref->real_size_ > req->real_size_) { + ref->real_size_ = req->real_size_; + } + } + } else { + match = false; + req->flags_ &= ~MPI_REQ_MATCHED; + ref->detached_sender_ = nullptr; + XBT_DEBUG("Refusing to match message, as its ID is not the one I expect. in comm %p, %u != %u, " + "from pid %ld to pid %ld, with tag %d", + ref->comm_, + ref->comm_->get_received_messages_count(ref->comm_->group()->rank(req->src_), + ref->comm_->group()->rank(req->dst_), req->tag_), + req->message_id_, req->src_, req->dst_, req->tag_); + } + return match; } bool Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl*) @@ -444,6 +472,9 @@ void Request::start() if (smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0) mut->lock(); + bool is_probe = ((flags_ & MPI_REQ_PROBE) != 0); + flags_ |= MPI_REQ_PROBE; + if (smpi_cfg_async_small_thresh() == 0 && (flags_ & MPI_REQ_RMA) == 0) { mailbox = process->mailbox(); } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast(size_) < smpi_cfg_async_small_thresh()) { @@ -463,7 +494,7 @@ void Request::start() mailbox = process->mailbox_small(); } } else { - XBT_DEBUG("yes there was something for us in the large mailbox"); + XBT_DEBUG("yes there was something for us in the small mailbox"); } } else { mailbox = process->mailbox_small(); @@ -477,10 +508,22 @@ void Request::start() XBT_DEBUG("yes there was something for us in the small mailbox"); } } + if(!is_probe) + flags_ &= ~MPI_REQ_PROBE; + kernel::actor::CommIrecvSimcall observer{process->get_actor()->get_impl(), + mailbox->get_impl(), + static_cast(buf_), + &real_size_, + &match_recv, + process->replaying() ? &smpi_comm_null_copy_buffer_callback + : smpi_comm_copy_data_callback, + this, + -1.0}; + observer.set_tag(tag_); + + action_ = kernel::actor::simcall_answered([&observer] { return kernel::activity::CommImpl::irecv(&observer); }, + &observer); - action_ = simcall_comm_irecv( - process->get_actor()->get_impl(), mailbox->get_impl(), buf_, &real_size_, &match_recv, - process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this, -1.0); XBT_DEBUG("recv simcall posted"); if (smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0) @@ -492,6 +535,9 @@ void Request::start() TRACE_smpi_send(src_, src_, dst_, tag_, size_); this->print_request("New send"); + message_id_=comm_->get_sent_messages_count(comm_->group()->rank(src_), comm_->group()->rank(dst_), tag_); + comm_->increment_sent_messages_count(comm_->group()->rank(src_), comm_->group()->rank(dst_), tag_); + void* buf = buf_; if ((flags_ & MPI_REQ_SSEND) == 0 && ((flags_ & MPI_REQ_RMA) != 0 || (flags_ & MPI_REQ_BSEND) != 0 || @@ -541,6 +587,9 @@ void Request::start() 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(size_) < smpi_cfg_async_small_thresh()) { // eager mode + bool is_probe = ((flags_ & MPI_REQ_PROBE) != 0); + flags_ |= MPI_REQ_PROBE; + mailbox = process->mailbox(); XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %s?", mailbox->get_cname()); simgrid::kernel::activity::ActivityImplPtr action = mailbox->iprobe(1, &match_send, static_cast(this)); @@ -562,19 +611,24 @@ void Request::start() } else { XBT_DEBUG("Yes there was something for us in the large mailbox"); } + if(!is_probe) + flags_ &= ~MPI_REQ_PROBE; } else { mailbox = process->mailbox(); XBT_DEBUG("Send request %p is in the large mailbox %s (buf: %p)", this, mailbox->get_cname(), buf_); } size_t payload_size_ = size_ + 16;//MPI enveloppe size (tag+dest+communicator) - action_ = simcall_comm_isend( - simgrid::kernel::actor::ActorImpl::by_pid(src_), mailbox->get_impl(), payload_size_, -1.0, buf, real_size_, - &match_send, + kernel::actor::CommIsendSimcall observer{ + simgrid::kernel::actor::ActorImpl::by_pid(src_), mailbox->get_impl(), static_cast(payload_size_), -1, + static_cast(buf), real_size_, &match_send, &xbt_free_f, // how to free the userdata if a detached send fails process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this, // detach if msg size < eager/rdv switch limit - detached_); + detached_}; + observer.set_tag(tag_); + action_ = kernel::actor::simcall_answered([&observer] { return kernel::activity::CommImpl::isend(&observer); }, + &observer); XBT_DEBUG("send simcall posted"); /* FIXME: detached sends are not traceable (action_ == nullptr) */ @@ -630,7 +684,10 @@ int Request::test(MPI_Request * request, MPI_Status * status, int* flag) { if (((*request)->flags_ & (MPI_REQ_PREPARED | MPI_REQ_FINISHED)) == 0) { if ((*request)->action_ != nullptr && ((*request)->flags_ & MPI_REQ_CANCELLED) == 0){ try{ - *flag = simcall_comm_test((*request)->action_.get()); + kernel::actor::ActorImpl* issuer = kernel::actor::ActorImpl::self(); + kernel::actor::ActivityTestSimcall observer{issuer, (*request)->action_.get()}; + *flag = kernel::actor::simcall_answered( + [&observer] { return observer.get_activity()->test(observer.get_issuer()); }, &observer); } catch (const Exception&) { *flag = 0; return ret; @@ -696,7 +753,7 @@ int Request::testsome(int incount, MPI_Request requests[], int *count, int *indi int Request::testany(int count, MPI_Request requests[], int *index, int* flag, MPI_Status * status) { - std::vector comms; + std::vector comms; comms.reserve(count); *flag = 0; @@ -706,7 +763,7 @@ int Request::testany(int count, MPI_Request requests[], int *index, int* flag, M std::vector map; /** Maps all matching comms back to their location in requests **/ for (int i = 0; i < count; i++) { if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action_ && not(requests[i]->flags_ & MPI_REQ_PREPARED)) { - comms.push_back(static_cast(requests[i]->action_.get())); + comms.push_back(requests[i]->action_.get()); map.push_back(i); } } @@ -717,7 +774,13 @@ int Request::testany(int count, MPI_Request requests[], int *index, int* flag, M simgrid::s4u::this_actor::sleep_for(nsleeps * smpi_test_sleep); ssize_t i; try{ - i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches! + kernel::actor::ActorImpl* issuer = kernel::actor::ActorImpl::self(); + kernel::actor::ActivityTestanySimcall observer{issuer, comms}; + i = kernel::actor::simcall_answered( + [&observer] { + return kernel::activity::ActivityImpl::test_any(observer.get_issuer(), observer.get_activities()); + }, + &observer); } catch (const Exception&) { XBT_DEBUG("Exception in testany"); return 0; @@ -1027,7 +1090,11 @@ int Request::wait(MPI_Request * request, MPI_Status * status) if ((*request)->action_ != nullptr){ try{ // this is not a detached send - simcall_comm_wait((*request)->action_.get(), -1.0); + kernel::actor::ActorImpl* issuer = kernel::actor::ActorImpl::self(); + kernel::actor::ActivityWaitSimcall observer{issuer, (*request)->action_.get(), -1}; + kernel::actor::simcall_blocking( + [&observer] { observer.get_activity()->wait_for(observer.get_issuer(), observer.get_timeout()); }, + &observer); } catch (const CancelException&) { XBT_VERB("Request cancelled"); } @@ -1068,7 +1135,7 @@ int Request::waitany(int count, MPI_Request requests[], MPI_Status * status) if(count > 0) { // Wait for a request to complete - std::vector comms; + std::vector comms; std::vector map; XBT_DEBUG("Wait for one of %d", count); for(int i = 0; i < count; i++) { @@ -1076,7 +1143,7 @@ int Request::waitany(int count, MPI_Request requests[], MPI_Status * status) not(requests[i]->flags_ & MPI_REQ_FINISHED)) { if (requests[i]->action_ != nullptr) { XBT_DEBUG("Waiting any %p ", requests[i]); - comms.push_back(static_cast(requests[i]->action_.get())); + comms.push_back(requests[i]->action_.get()); map.push_back(i); } else { // This is a finished detached request, let's return this one @@ -1095,7 +1162,14 @@ int Request::waitany(int count, MPI_Request requests[], MPI_Status * status) XBT_DEBUG("Enter waitany for %zu comms", comms.size()); ssize_t i; try{ - i = simcall_comm_waitany(comms.data(), comms.size(), -1); + kernel::actor::ActorImpl* issuer = kernel::actor::ActorImpl::self(); + kernel::actor::ActivityWaitanySimcall observer{issuer, comms, -1}; + i = kernel::actor::simcall_blocking( + [&observer] { + kernel::activity::ActivityImpl::wait_any_for(observer.get_issuer(), observer.get_activities(), + observer.get_timeout()); + }, + &observer); } catch (const CancelException&) { XBT_INFO("request cancelled"); i = -1;