X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/0b709e78b05186146867b7b5a841bdec06507bf8..121e8462d6b912a252a9016683d2e3e021cde9a0:/src/kernel/activity/CommImpl.cpp diff --git a/src/kernel/activity/CommImpl.cpp b/src/kernel/activity/CommImpl.cpp index f55d4d445c..ff80213f39 100644 --- a/src/kernel/activity/CommImpl.cpp +++ b/src/kernel/activity/CommImpl.cpp @@ -4,22 +4,368 @@ * under the terms of the license (GNU LGPL) which comes with this package. */ #include "src/kernel/activity/CommImpl.hpp" +#include "simgrid/Exception.hpp" #include "simgrid/kernel/resource/Action.hpp" #include "simgrid/modelchecker.h" #include "simgrid/s4u/Host.hpp" #include "src/kernel/activity/MailboxImpl.hpp" +#include "src/kernel/context/Context.hpp" #include "src/mc/mc_replay.hpp" -#include "src/simix/smx_network_private.hpp" +#include "src/surf/cpu_interface.hpp" #include "src/surf/network_interface.hpp" #include "src/surf/surf_interface.hpp" -XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(simix_network); +#include +XBT_LOG_NEW_DEFAULT_SUBCATEGORY(simix_network, simix, "SIMIX network-related synchronization"); + +XBT_PRIVATE void simcall_HANDLER_comm_send(smx_simcall_t simcall, smx_actor_t src, smx_mailbox_t mbox, double task_size, + double rate, void* src_buff, size_t src_buff_size, + int (*match_fun)(void*, void*, simgrid::kernel::activity::CommImpl*), + void (*copy_data_fun)(smx_activity_t, void*, size_t), void* data, + double timeout) +{ + smx_activity_t comm = simcall_HANDLER_comm_isend(simcall, src, mbox, task_size, rate, src_buff, src_buff_size, + match_fun, nullptr, copy_data_fun, data, 0); + SIMCALL_SET_MC_VALUE(simcall, 0); + simcall_HANDLER_comm_wait(simcall, comm, timeout); +} + +XBT_PRIVATE smx_activity_t simcall_HANDLER_comm_isend( + smx_simcall_t /*simcall*/, smx_actor_t src_proc, smx_mailbox_t mbox, double task_size, double rate, void* src_buff, + size_t src_buff_size, int (*match_fun)(void*, void*, simgrid::kernel::activity::CommImpl*), + void (*clean_fun)(void*), // used to free the synchro in case of problem after a detached send + void (*copy_data_fun)(smx_activity_t, void*, size_t), // used to copy data if not default one + void* data, int detached) +{ + XBT_DEBUG("send from mailbox %p", mbox); + + /* Prepare a synchro describing us, so that it gets passed to the user-provided filter of other side */ + simgrid::kernel::activity::CommImplPtr this_comm = simgrid::kernel::activity::CommImplPtr( + new simgrid::kernel::activity::CommImpl(simgrid::kernel::activity::CommImpl::Type::SEND)); + + /* Look for communication synchro matching our needs. We also provide a description of + * ourself so that the other side also gets a chance of choosing if it wants to match with us. + * + * If it is not found then push our communication into the rendez-vous point */ + simgrid::kernel::activity::CommImplPtr other_comm = + mbox->find_matching_comm(simgrid::kernel::activity::CommImpl::Type::RECEIVE, match_fun, data, this_comm, + /*done*/ false, /*remove_matching*/ true); + + if (not other_comm) { + other_comm = std::move(this_comm); + + if (mbox->permanent_receiver_ != nullptr) { + // this mailbox is for small messages, which have to be sent right now + other_comm->state_ = SIMIX_READY; + other_comm->dst_actor_ = mbox->permanent_receiver_.get(); + mbox->done_comm_queue_.push_back(other_comm); + XBT_DEBUG("pushing a message into the permanent receive list %p, comm %p", mbox, other_comm.get()); + + } else { + mbox->push(other_comm); + } + } else { + XBT_DEBUG("Receive already pushed"); + + other_comm->state_ = SIMIX_READY; + other_comm->type = simgrid::kernel::activity::CommImpl::Type::READY; + } + src_proc->comms.push_back(other_comm); + + if (detached) { + other_comm->detached = true; + other_comm->clean_fun = clean_fun; + } else { + other_comm->clean_fun = nullptr; + } + + /* Setup the communication synchro */ + other_comm->src_actor_ = src_proc; + other_comm->task_size_ = task_size; + other_comm->rate_ = rate; + other_comm->src_buff_ = src_buff; + other_comm->src_buff_size_ = src_buff_size; + other_comm->src_data_ = data; + + other_comm->match_fun = match_fun; + other_comm->copy_data_fun = copy_data_fun; + + if (MC_is_active() || MC_record_replay_is_active()) { + other_comm->state_ = SIMIX_RUNNING; + return (detached ? nullptr : other_comm); + } + + other_comm->start(); + + return (detached ? nullptr : other_comm); +} + +XBT_PRIVATE void simcall_HANDLER_comm_recv(smx_simcall_t simcall, smx_actor_t receiver, smx_mailbox_t mbox, + void* dst_buff, size_t* dst_buff_size, + int (*match_fun)(void*, void*, simgrid::kernel::activity::CommImpl*), + void (*copy_data_fun)(smx_activity_t, void*, size_t), void* data, + double timeout, double rate) +{ + smx_activity_t comm = simcall_HANDLER_comm_irecv(simcall, receiver, mbox, dst_buff, dst_buff_size, match_fun, + copy_data_fun, data, rate); + SIMCALL_SET_MC_VALUE(simcall, 0); + simcall_HANDLER_comm_wait(simcall, comm, timeout); +} + +XBT_PRIVATE smx_activity_t simcall_HANDLER_comm_irecv(smx_simcall_t /*simcall*/, smx_actor_t receiver, + smx_mailbox_t mbox, void* dst_buff, size_t* dst_buff_size, + simix_match_func_t match_fun, + void (*copy_data_fun)(smx_activity_t, void*, size_t), void* data, + double rate) +{ + simgrid::kernel::activity::CommImplPtr this_synchro = simgrid::kernel::activity::CommImplPtr( + new simgrid::kernel::activity::CommImpl(simgrid::kernel::activity::CommImpl::Type::RECEIVE)); + XBT_DEBUG("recv from mbox %p. this_synchro=%p", mbox, this_synchro.get()); + + simgrid::kernel::activity::CommImplPtr other_comm; + // communication already done, get it inside the list of completed comms + if (mbox->permanent_receiver_ != nullptr && not mbox->done_comm_queue_.empty()) { + + XBT_DEBUG("We have a comm that has probably already been received, trying to match it, to skip the communication"); + // find a match in the list of already received comms + other_comm = mbox->find_matching_comm(simgrid::kernel::activity::CommImpl::Type::SEND, match_fun, data, + this_synchro, /*done*/ true, + /*remove_matching*/ true); + // if not found, assume the receiver came first, register it to the mailbox in the classical way + if (not other_comm) { + XBT_DEBUG("We have messages in the permanent receive list, but not the one we are looking for, pushing request " + "into list"); + other_comm = std::move(this_synchro); + mbox->push(other_comm); + } else { + if (other_comm->surf_action_ && other_comm->remains() < 1e-12) { + XBT_DEBUG("comm %p has been already sent, and is finished, destroy it", other_comm.get()); + other_comm->state_ = SIMIX_DONE; + other_comm->type = simgrid::kernel::activity::CommImpl::Type::DONE; + other_comm->mbox = nullptr; + } + } + } else { + /* Prepare a comm describing us, so that it gets passed to the user-provided filter of other side */ + + /* Look for communication activity matching our needs. We also provide a description of + * ourself so that the other side also gets a chance of choosing if it wants to match with us. + * + * If it is not found then push our communication into the rendez-vous point */ + other_comm = mbox->find_matching_comm(simgrid::kernel::activity::CommImpl::Type::SEND, match_fun, data, + this_synchro, /*done*/ false, + /*remove_matching*/ true); + + if (other_comm == nullptr) { + XBT_DEBUG("Receive pushed first (%zu comm enqueued so far)", mbox->comm_queue_.size()); + other_comm = std::move(this_synchro); + mbox->push(other_comm); + } else { + XBT_DEBUG("Match my %p with the existing %p", this_synchro.get(), other_comm.get()); + + other_comm->state_ = SIMIX_READY; + other_comm->type = simgrid::kernel::activity::CommImpl::Type::READY; + } + receiver->comms.push_back(other_comm); + } + + /* Setup communication synchro */ + other_comm->dst_actor_ = receiver; + other_comm->dst_buff_ = dst_buff; + other_comm->dst_buff_size_ = dst_buff_size; + other_comm->dst_data_ = data; + + if (rate > -1.0 && (other_comm->rate_ < 0.0 || rate < other_comm->rate_)) + other_comm->rate_ = rate; + + other_comm->match_fun = match_fun; + other_comm->copy_data_fun = copy_data_fun; + + if (MC_is_active() || MC_record_replay_is_active()) { + other_comm->state_ = SIMIX_RUNNING; + return other_comm; + } + other_comm->start(); + return other_comm; +} + +void simcall_HANDLER_comm_wait(smx_simcall_t simcall, smx_activity_t synchro, double timeout) +{ + /* Associate this simcall to the wait synchro */ + XBT_DEBUG("simcall_HANDLER_comm_wait, %p", synchro.get()); + + synchro->simcalls_.push_back(simcall); + simcall->issuer->waiting_synchro = synchro; + + if (MC_is_active() || MC_record_replay_is_active()) { + int idx = SIMCALL_GET_MC_VALUE(simcall); + if (idx == 0) { + synchro->state_ = SIMIX_DONE; + } else { + /* If we reached this point, the wait simcall must have a timeout */ + /* Otherwise it shouldn't be enabled and executed by the MC */ + if (timeout < 0.0) + THROW_IMPOSSIBLE; + + simgrid::kernel::activity::CommImplPtr comm = + boost::static_pointer_cast(synchro); + if (comm->src_actor_ == simcall->issuer) + comm->state_ = SIMIX_SRC_TIMEOUT; + else + comm->state_ = SIMIX_DST_TIMEOUT; + } + + boost::static_pointer_cast(synchro)->finish(); + return; + } + + /* If the synchro has already finish perform the error handling, */ + /* otherwise set up a waiting timeout on the right side */ + if (synchro->state_ != SIMIX_WAITING && synchro->state_ != SIMIX_RUNNING) { + boost::static_pointer_cast(synchro)->finish(); + } else { /* we need a sleep action (even when there is no timeout) to be notified of host failures */ + simgrid::kernel::resource::Action* sleep = simcall->issuer->get_host()->pimpl_cpu->sleep(timeout); + sleep->set_data(synchro.get()); + + simgrid::kernel::activity::CommImplPtr comm = + boost::static_pointer_cast(synchro); + if (simcall->issuer == comm->src_actor_) + comm->src_timeout_ = sleep; + else + comm->dst_timeout_ = sleep; + } +} + +void simcall_HANDLER_comm_test(smx_simcall_t simcall, smx_activity_t synchro) +{ + simgrid::kernel::activity::CommImplPtr comm = + boost::static_pointer_cast(synchro); + + int res; + + if (MC_is_active() || MC_record_replay_is_active()) { + res = comm->src_actor_ && comm->dst_actor_; + if (res) + synchro->state_ = SIMIX_DONE; + } else { + res = synchro->state_ != SIMIX_WAITING && synchro->state_ != SIMIX_RUNNING; + } + + simcall_comm_test__set__result(simcall, res); + if (simcall_comm_test__get__result(simcall)) { + synchro->simcalls_.push_back(simcall); + boost::static_pointer_cast(synchro)->finish(); + } else { + SIMIX_simcall_answer(simcall); + } +} + +void simcall_HANDLER_comm_testany(smx_simcall_t simcall, simgrid::kernel::activity::ActivityImplPtr comms[], + size_t count) +{ + // The default result is -1 -- this means, "nothing is ready". + // It can be changed below, but only if something matches. + simcall_comm_testany__set__result(simcall, -1); + + if (MC_is_active() || MC_record_replay_is_active()) { + int idx = SIMCALL_GET_MC_VALUE(simcall); + if (idx == -1) { + SIMIX_simcall_answer(simcall); + } else { + simgrid::kernel::activity::ActivityImplPtr synchro = comms[idx]; + simcall_comm_testany__set__result(simcall, idx); + synchro->simcalls_.push_back(simcall); + synchro->state_ = SIMIX_DONE; + boost::static_pointer_cast(synchro)->finish(); + } + return; + } + + for (std::size_t i = 0; i != count; ++i) { + simgrid::kernel::activity::ActivityImplPtr synchro = comms[i]; + if (synchro->state_ != SIMIX_WAITING && synchro->state_ != SIMIX_RUNNING) { + simcall_comm_testany__set__result(simcall, i); + synchro->simcalls_.push_back(simcall); + boost::static_pointer_cast(synchro)->finish(); + return; + } + } + SIMIX_simcall_answer(simcall); +} + +static void SIMIX_waitany_remove_simcall_from_actions(smx_simcall_t simcall) +{ + unsigned int cursor = 0; + xbt_dynar_t synchros = simcall_comm_waitany__get__comms(simcall); + + simgrid::kernel::activity::ActivityImpl* ptr; + xbt_dynar_foreach (synchros, cursor, ptr) { + smx_activity_t synchro = simgrid::kernel::activity::ActivityImplPtr(ptr); + + // Remove the first occurence of simcall: + auto i = boost::range::find(synchro->simcalls_, simcall); + if (i != synchro->simcalls_.end()) + synchro->simcalls_.erase(i); + } +} +void simcall_HANDLER_comm_waitany(smx_simcall_t simcall, xbt_dynar_t synchros, double timeout) +{ + if (MC_is_active() || MC_record_replay_is_active()) { + if (timeout > 0.0) + xbt_die("Timeout not implemented for waitany in the model-checker"); + int idx = SIMCALL_GET_MC_VALUE(simcall); + smx_activity_t synchro = xbt_dynar_get_as(synchros, idx, smx_activity_t); + synchro->simcalls_.push_back(simcall); + simcall_comm_waitany__set__result(simcall, idx); + synchro->state_ = SIMIX_DONE; + boost::static_pointer_cast(synchro)->finish(); + return; + } + + if (timeout < 0.0) { + simcall->timer = NULL; + } else { + simcall->timer = simgrid::simix::Timer::set(SIMIX_get_clock() + timeout, [simcall]() { + SIMIX_waitany_remove_simcall_from_actions(simcall); + simcall_comm_waitany__set__result(simcall, -1); + SIMIX_simcall_answer(simcall); + }); + } + + unsigned int cursor; + simgrid::kernel::activity::ActivityImpl* ptr; + xbt_dynar_foreach (synchros, cursor, ptr) { + smx_activity_t synchro = simgrid::kernel::activity::ActivityImplPtr(ptr); + /* associate this simcall to the the synchro */ + synchro->simcalls_.push_back(simcall); + + /* see if the synchro is already finished */ + if (synchro->state_ != SIMIX_WAITING && synchro->state_ != SIMIX_RUNNING) { + boost::static_pointer_cast(synchro)->finish(); + break; + } + } +} /******************************************************************************/ /* SIMIX_comm_copy_data callbacks */ /******************************************************************************/ static void (*SIMIX_comm_copy_data_callback)(smx_activity_t, void*, size_t) = &SIMIX_comm_copy_pointer_callback; +void SIMIX_comm_copy_buffer_callback(smx_activity_t synchro, void* buff, size_t buff_size) +{ + simgrid::kernel::activity::CommImplPtr comm = + boost::static_pointer_cast(synchro); + + XBT_DEBUG("Copy the data over"); + memcpy(comm->dst_buff_, buff, buff_size); + 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); + comm->src_buff_ = nullptr; + } +} + void SIMIX_comm_set_copy_data_callback(void (*callback)(smx_activity_t, void*, size_t)) { SIMIX_comm_copy_data_callback = callback; @@ -38,7 +384,7 @@ namespace simgrid { namespace kernel { namespace activity { -CommImpl::CommImpl(e_smx_comm_type_t _type) : type(_type) +CommImpl::CommImpl(CommImpl::Type type) : type(type) { state_ = SIMIX_WAITING; src_data_ = nullptr; @@ -70,8 +416,8 @@ void CommImpl::start() /* If both the sender and the receiver are already there, start the communication */ if (state_ == SIMIX_READY) { - s4u::Host* sender = src_actor_->host_; - s4u::Host* receiver = dst_actor_->host_; + s4u::Host* sender = src_actor_->get_host(); + s4u::Host* receiver = dst_actor_->get_host(); surf_action_ = surf_network_model->communicate(sender, receiver, task_size_, rate_); surf_action_->set_data(this); @@ -86,19 +432,18 @@ void CommImpl::start() receiver->get_cname()); state_ = SIMIX_LINK_FAILURE; cleanupSurf(); - } - /* If any of the process is suspended, create the synchro but stop its execution, - it will be restarted when the sender process resume */ - if (src_actor_->is_suspended() || dst_actor_->is_suspended()) { + } else if (src_actor_->is_suspended() || dst_actor_->is_suspended()) { + /* If any of the process is suspended, create the synchro but stop its execution, + it will be restarted when the sender process resume */ if (src_actor_->is_suspended()) XBT_DEBUG("The communication is suspended on startup because src (%s@%s) was suspended since it initiated the " "communication", - src_actor_->get_cname(), src_actor_->host_->get_cname()); + src_actor_->get_cname(), src_actor_->get_host()->get_cname()); else XBT_DEBUG("The communication is suspended on startup because dst (%s@%s) was suspended since it initiated the " "communication", - dst_actor_->get_cname(), dst_actor_->host_->get_cname()); + dst_actor_->get_cname(), dst_actor_->get_host()->get_cname()); surf_action_->suspend(); } @@ -114,8 +459,8 @@ void CommImpl::copy_data() return; XBT_DEBUG("Copying comm %p data from %s (%p) -> %s (%p) (%zu bytes)", this, - src_actor_ ? src_actor_->host_->get_cname() : "a finished process", src_buff_, - dst_actor_ ? dst_actor_->host_->get_cname() : "a finished process", dst_buff_, buff_size); + src_actor_ ? src_actor_->get_host()->get_cname() : "a finished process", src_buff_, + dst_actor_ ? dst_actor_->get_host()->get_cname() : "a finished process", dst_buff_, buff_size); /* Copy at most dst_buff_size bytes of the message to receiver's buffer */ if (dst_buff_size_) @@ -214,7 +559,163 @@ void CommImpl::post() /* if there are simcalls associated with the synchro, then answer them */ if (not simcalls_.empty()) { - SIMIX_comm_finish(this); + finish(); + } +} + +void CommImpl::finish() +{ + smx_activity_t synchro = this; + while (not simcalls_.empty()) { + smx_simcall_t simcall = simcalls_.front(); + simcalls_.pop_front(); + + /* If a waitany simcall is waiting for this synchro to finish, then remove it from the other synchros in the waitany + * list. Afterwards, get the position of the actual synchro in the waitany dynar and return it as the result of the + * simcall */ + + if (simcall->call == SIMCALL_NONE) // FIXME: maybe a better way to handle this case + continue; // if process handling comm is killed + if (simcall->call == SIMCALL_COMM_WAITANY) { + SIMIX_waitany_remove_simcall_from_actions(simcall); + if (simcall->timer) { + simcall->timer->remove(); + simcall->timer = nullptr; + } + if (not MC_is_active() && not MC_record_replay_is_active()) + simcall_comm_waitany__set__result(simcall, + xbt_dynar_search(simcall_comm_waitany__get__comms(simcall), &synchro)); + } + + /* If the synchro is still in a rendez-vous point then remove from it */ + if (mbox) + mbox->remove(this); + + XBT_DEBUG("CommImpl::finish(): synchro state = %d", static_cast(state_)); + + /* Check out for errors */ + + if (not simcall->issuer->get_host()->is_on()) { + simcall->issuer->context_->iwannadie = true; + simcall->issuer->exception_ = + std::make_exception_ptr(simgrid::HostFailureException(XBT_THROW_POINT, "Host failed")); + } else { + switch (state_) { + + case SIMIX_DONE: + XBT_DEBUG("Communication %p complete!", this); + copy_data(); + break; + + case SIMIX_SRC_TIMEOUT: + simcall->issuer->exception_ = std::make_exception_ptr( + simgrid::TimeoutError(XBT_THROW_POINT, "Communication timeouted because of the sender")); + break; + + case SIMIX_DST_TIMEOUT: + simcall->issuer->exception_ = std::make_exception_ptr( + simgrid::TimeoutError(XBT_THROW_POINT, "Communication timeouted because of the receiver")); + break; + + case SIMIX_SRC_HOST_FAILURE: + if (simcall->issuer == src_actor_) + simcall->issuer->context_->iwannadie = true; + else + simcall->issuer->exception_ = + std::make_exception_ptr(simgrid::NetworkFailureException(XBT_THROW_POINT, "Remote peer failed")); + break; + + case SIMIX_DST_HOST_FAILURE: + if (simcall->issuer == dst_actor_) + simcall->issuer->context_->iwannadie = true; + else + simcall->issuer->exception_ = + std::make_exception_ptr(simgrid::NetworkFailureException(XBT_THROW_POINT, "Remote peer failed")); + break; + + case SIMIX_LINK_FAILURE: + XBT_DEBUG("Link failure in synchro %p between '%s' and '%s': posting an exception to the issuer: %s (%p) " + "detached:%d", + this, src_actor_ ? src_actor_->get_host()->get_cname() : nullptr, + dst_actor_ ? dst_actor_->get_host()->get_cname() : nullptr, simcall->issuer->get_cname(), + simcall->issuer, detached); + if (src_actor_ == simcall->issuer) { + XBT_DEBUG("I'm source"); + } else if (dst_actor_ == simcall->issuer) { + XBT_DEBUG("I'm dest"); + } else { + XBT_DEBUG("I'm neither source nor dest"); + } + simcall->issuer->throw_exception( + std::make_exception_ptr(simgrid::NetworkFailureException(XBT_THROW_POINT, "Link failure"))); + break; + + case SIMIX_CANCELED: + if (simcall->issuer == dst_actor_) + simcall->issuer->exception_ = std::make_exception_ptr( + simgrid::CancelException(XBT_THROW_POINT, "Communication canceled by the sender")); + else + simcall->issuer->exception_ = std::make_exception_ptr( + simgrid::CancelException(XBT_THROW_POINT, "Communication canceled by the receiver")); + break; + + default: + xbt_die("Unexpected synchro state in CommImpl::finish: %d", static_cast(state_)); + } + } + + /* if there is an exception during a waitany or a testany, indicate the position of the failed communication */ + if (simcall->issuer->exception_ && + (simcall->call == SIMCALL_COMM_WAITANY || simcall->call == SIMCALL_COMM_TESTANY)) { + // First retrieve the rank of our failing synchro + int rank = -1; + if (simcall->call == SIMCALL_COMM_WAITANY) { + rank = xbt_dynar_search(simcall_comm_waitany__get__comms(simcall), &synchro); + } else if (simcall->call == SIMCALL_COMM_TESTANY) { + rank = -1; + auto* comms = simcall_comm_testany__get__comms(simcall); + auto count = simcall_comm_testany__get__count(simcall); + auto element = std::find(comms, comms + count, this); + if (element == comms + count) + rank = -1; + else + rank = element - comms; + } + + // In order to modify the exception we have to rethrow it: + try { + std::rethrow_exception(simcall->issuer->exception_); + } catch (simgrid::TimeoutError& e) { + e.value = rank; + simcall->issuer->exception_ = std::make_exception_ptr(e); + } catch (simgrid::NetworkFailureException& e) { + e.value = rank; + simcall->issuer->exception_ = std::make_exception_ptr(e); + } catch (simgrid::CancelException& e) { + e.value = rank; + simcall->issuer->exception_ = std::make_exception_ptr(e); + } + } + + simcall->issuer->waiting_synchro = nullptr; + simcall->issuer->comms.remove(this); + if (detached) { + if (simcall->issuer == src_actor_) { + if (dst_actor_) + dst_actor_->comms.remove(this); + } else if (simcall->issuer == dst_actor_) { + if (src_actor_) + src_actor_->comms.remove(this); + } else { + dst_actor_->comms.remove(this); + src_actor_->comms.remove(this); + } + } + + if (simcall->issuer->get_host()->is_on()) + SIMIX_simcall_answer(simcall); + else + simcall->issuer->context_->iwannadie = true; } }