X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/a3d08dd00246eb26ced68c5b0e046096706bbe23..3d65d2eaad96faf1c97960216cf8afe08308f9e5:/src/smpi/internals/smpi_global.cpp diff --git a/src/smpi/internals/smpi_global.cpp b/src/smpi/internals/smpi_global.cpp index 1ccc8bfbae..f3f49b38ca 100644 --- a/src/smpi/internals/smpi_global.cpp +++ b/src/smpi/internals/smpi_global.cpp @@ -1,31 +1,45 @@ -/* Copyright (c) 2007-2017. The SimGrid Team. All rights reserved. */ +/* Copyright (c) 2007-2019. 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 "mc/mc.h" -#include "simgrid/s4u/Mailbox.hpp" -#include "simgrid/s4u/Host.hpp" -#include "src/msg/msg_private.h" -#include "src/simix/smx_private.h" -#include "src/surf/surf_interface.hpp" -#include "SmpiHost.hpp" -#include "xbt/config.hpp" -#include "private.h" -#include "private.hpp" +#include "simgrid/s4u/Engine.hpp" +#include "simgrid/plugins/file_system.h" #include "smpi_coll.hpp" -#include "smpi_comm.hpp" -#include "smpi_group.hpp" -#include "smpi_info.hpp" -#include "smpi_process.hpp" +#include "smpi_f2c.hpp" +#include "smpi_host.hpp" +#include "src/kernel/activity/CommImpl.hpp" +#include "src/simix/smx_private.hpp" +#include "src/smpi/include/smpi_actor.hpp" +#include "xbt/config.hpp" +#include #include /* DBL_MAX */ +#include +#include /* intmax_t */ #include #include #include #include -#if HAVE_SENDFILE +#if not defined(__APPLE__) && not defined(__HAIKU__) +#include +#endif + +#if defined(__APPLE__) +# include +# ifndef MAC_OS_X_VERSION_10_12 +# define MAC_OS_X_VERSION_10_12 101200 +# endif +constexpr bool HAVE_WORKING_MMAP = (MAC_OS_X_VERSION_MIN_REQUIRED >= MAC_OS_X_VERSION_10_12); +#elif defined(__FreeBSD__) || defined(__FreeBSD_kernel__) || defined(__sun) || defined(__HAIKU__) +constexpr bool HAVE_WORKING_MMAP = false; +#else +constexpr bool HAVE_WORKING_MMAP = true; +#endif + +#if SG_HAVE_SENDFILE #include #endif @@ -33,103 +47,114 @@ XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_kernel, smpi, "Logging specific to SMPI (ke #include #include /* trim_right / trim_left */ -#ifndef RTLD_DEEPBIND +#if SMPI_IFORT + extern "C" void for_rtl_init_ (int *, char **); + extern "C" void for_rtl_finish_ (); +#elif SMPI_FLANG + extern "C" void __io_set_argc(int); + extern "C" void __io_set_argv(char **); +#elif SMPI_GFORTRAN + extern "C" void _gfortran_set_args(int, char **); +#endif + /* RTLD_DEEPBIND is a bad idea of GNU ld that obviously does not exist on other platforms * See https://www.akkadia.org/drepper/dsohowto.pdf * and https://lists.freebsd.org/pipermail/freebsd-current/2016-March/060284.html */ -#define RTLD_DEEPBIND 0 +#if !RTLD_DEEPBIND || HAVE_SANITIZER_ADDRESS || HAVE_SANITIZER_THREAD +#define WANT_RTLD_DEEPBIND 0 +#else +#define WANT_RTLD_DEEPBIND RTLD_DEEPBIND #endif #if HAVE_PAPI #include "papi.h" -const char* papi_default_config_name = "default"; - -struct papi_process_data { - papi_counter_t counter_data; - int event_set; -}; - +std::string papi_default_config_name = "default"; +std::map units2papi_setup; #endif + std::unordered_map location2speedup; -simgrid::smpi::Process **process_data = nullptr; +static std::map process_data; int process_count = 0; +static int smpi_exit_status = 0; int smpi_universe_size = 0; -int* index_to_process_data = nullptr; extern double smpi_total_benched_time; xbt_os_timer_t global_timer; +static std::vector privatize_libs_paths; +/** + * Setting MPI_COMM_WORLD to MPI_COMM_UNINITIALIZED (it's a variable) + * is important because the implementation of MPI_Comm checks + * "this == MPI_COMM_UNINITIALIZED"? If yes, it uses smpi_process()->comm_world() + * instead of "this". + * This is basically how we only have one global variable but all processes have + * different communicators (the one their SMPI instance uses). + * + * See smpi_comm.cpp and the functions therein for details. + */ MPI_Comm MPI_COMM_WORLD = MPI_COMM_UNINITIALIZED; MPI_Errhandler *MPI_ERRORS_RETURN = nullptr; MPI_Errhandler *MPI_ERRORS_ARE_FATAL = nullptr; MPI_Errhandler *MPI_ERRHANDLER_NULL = nullptr; -static simgrid::config::Flag smpi_wtime_sleep( - "smpi/wtime", "Minimum time to inject inside a call to MPI_Wtime", 0.0); +// No instance gets manually created; check also the smpirun.in script as +// this default name is used there as well (when the tag is generated). +static const std::string smpi_default_instance_name("smpirun"); static simgrid::config::Flag smpi_init_sleep( "smpi/init", "Time to inject inside a call to MPI_Init", 0.0); -void (*smpi_comm_copy_data_callback) (smx_activity_t, void*, size_t) = &smpi_comm_copy_buffer_callback; +void (*smpi_comm_copy_data_callback)(simgrid::kernel::activity::CommImpl*, void*, + size_t) = &smpi_comm_copy_buffer_callback; int smpi_process_count() { return process_count; } -simgrid::smpi::Process* smpi_process() +simgrid::smpi::ActorExt* smpi_process() { - smx_actor_t me = SIMIX_process_self(); + simgrid::s4u::ActorPtr me = simgrid::s4u::Actor::self(); + if (me == nullptr) // This happens sometimes (eg, when linking against NS3 because it pulls openMPI...) return nullptr; - simgrid::msg::ActorExt* msgExt = static_cast(me->userdata); - return static_cast(msgExt->data); + + return process_data.at(me.get()); } -simgrid::smpi::Process* smpi_process_remote(int index) +simgrid::smpi::ActorExt* smpi_process_remote(simgrid::s4u::ActorPtr actor) { - return process_data[index_to_process_data[index]]; + return process_data.at(actor.get()); } MPI_Comm smpi_process_comm_self(){ return smpi_process()->comm_self(); } -void smpi_process_init(int *argc, char ***argv){ - simgrid::smpi::Process::init(argc, argv); +MPI_Info smpi_process_info_env(){ + return smpi_process()->info_env(); } -int smpi_process_index(){ - return smpi_process()->index(); +void smpi_process_init(int*, char***) +{ + simgrid::smpi::ActorExt::init(); } void * smpi_process_get_user_data(){ - return smpi_process()->get_user_data(); + return simgrid::s4u::Actor::self()->get_impl()->get_user_data(); } void smpi_process_set_user_data(void *data){ - return smpi_process()->set_user_data(data); -} - - -int smpi_global_size() -{ - char *value = getenv("SMPI_GLOBAL_SIZE"); - xbt_assert(value,"Please set env var SMPI_GLOBAL_SIZE to the expected number of processes."); - - return xbt_str_parse_int(value, "SMPI_GLOBAL_SIZE contains a non-numerical value: %s"); + simgrid::s4u::Actor::self()->get_impl()->set_user_data(data); } void smpi_comm_set_copy_data_callback(void (*callback) (smx_activity_t, void*, size_t)) { - smpi_comm_copy_data_callback = callback; + static void (*saved_callback)(smx_activity_t, void*, size_t); + saved_callback = callback; + smpi_comm_copy_data_callback = [](simgrid::kernel::activity::CommImpl* comm, void* buff, size_t size) { + saved_callback(smx_activity_t(comm), buff, size); + }; } -static void print(std::vector> vec) { - std::fprintf(stderr, "{"); - for (auto const& elt : vec) { - std::fprintf(stderr, "(0x%zx, 0x%zx),", elt.first, elt.second); - } - std::fprintf(stderr, "}\n"); -} static void memcpy_private(void* dest, const void* src, std::vector>& private_blocks) { for (auto const& block : private_blocks) @@ -141,10 +166,8 @@ static void check_blocks(std::vector> &private_blocks, xbt_assert(block.first <= block.second && block.second <= buff_size, "Oops, bug in shared malloc."); } -void smpi_comm_copy_buffer_callback(smx_activity_t synchro, void *buff, size_t buff_size) +void smpi_comm_copy_buffer_callback(simgrid::kernel::activity::CommImpl* comm, void* buff, size_t buff_size) { - simgrid::kernel::activity::CommImplPtr comm = - boost::dynamic_pointer_cast(synchro); int src_shared = 0; int dst_shared = 0; size_t src_offset = 0; @@ -160,8 +183,8 @@ void smpi_comm_copy_buffer_callback(smx_activity_t synchro, void *buff, size_t b src_private_blocks.clear(); src_private_blocks.push_back(std::make_pair(0, buff_size)); } - if((dst_shared=smpi_is_shared((char*)comm->dst_buff, dst_private_blocks, &dst_offset))) { - XBT_DEBUG("Receiver %p is shared. Let's ignore it.", (char*)comm->dst_buff); + if ((dst_shared = smpi_is_shared((char*)comm->dst_buff_, dst_private_blocks, &dst_offset))) { + XBT_DEBUG("Receiver %p is shared. Let's ignore it.", (char*)comm->dst_buff_); dst_private_blocks = shift_and_frame_private_blocks(dst_private_blocks, dst_offset, buff_size); } else { @@ -173,77 +196,75 @@ void smpi_comm_copy_buffer_callback(smx_activity_t synchro, void *buff, size_t b auto private_blocks = merge_private_blocks(src_private_blocks, dst_private_blocks); check_blocks(private_blocks, buff_size); void* tmpbuff=buff; - if((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) && (static_cast(buff) >= smpi_start_data_exe) - && (static_cast(buff) < smpi_start_data_exe + smpi_size_data_exe ) - ){ - XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !"); - - smpi_switch_data_segment( - static_cast((static_cast(comm->src_proc->userdata)->data)) - ->index()); - tmpbuff = static_cast(xbt_malloc(buff_size)); - memcpy_private(tmpbuff, buff, private_blocks); + if ((smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) && + (static_cast(buff) >= smpi_data_exe_start) && + (static_cast(buff) < smpi_data_exe_start + smpi_data_exe_size)) { + XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !"); + smpi_switch_data_segment(comm->src_actor_->iface()); + tmpbuff = static_cast(xbt_malloc(buff_size)); + memcpy_private(tmpbuff, buff, private_blocks); } - if((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) && ((char*)comm->dst_buff >= smpi_start_data_exe) - && ((char*)comm->dst_buff < smpi_start_data_exe + smpi_size_data_exe )){ - XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment"); - smpi_switch_data_segment( - static_cast((static_cast(comm->dst_proc->userdata)->data)) - ->index()); + if ((smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) && + ((char*)comm->dst_buff_ >= smpi_data_exe_start) && + ((char*)comm->dst_buff_ < smpi_data_exe_start + smpi_data_exe_size)) { + XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment"); + smpi_switch_data_segment(comm->dst_actor_->iface()); } - XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff,comm->dst_buff); - memcpy_private(comm->dst_buff, tmpbuff, private_blocks); + XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff, comm->dst_buff_); + memcpy_private(comm->dst_buff_, tmpbuff, private_blocks); - if (comm->detached) { + 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); //It seems that the request is used after the call there this should be free somewhere else but where??? //xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free - comm->src_buff = nullptr; + comm->src_buff_ = nullptr; } if (tmpbuff != buff) xbt_free(tmpbuff); } -void smpi_comm_null_copy_buffer_callback(smx_activity_t comm, void *buff, size_t buff_size) +void smpi_comm_null_copy_buffer_callback(simgrid::kernel::activity::CommImpl*, void*, size_t) { /* nothing done in this version */ } -static void smpi_check_options(){ +static void smpi_check_options() +{ //check correctness of MPI parameters - xbt_assert(xbt_cfg_get_int("smpi/async-small-thresh") <= xbt_cfg_get_int("smpi/send-is-detached-thresh")); + xbt_assert(simgrid::config::get_value("smpi/async-small-thresh") <= + simgrid::config::get_value("smpi/send-is-detached-thresh")); - if (xbt_cfg_is_default_value("smpi/host-speed")) { - XBT_INFO("You did not set the power of the host running the simulation. " - "The timings will certainly not be accurate. " - "Use the option \"--cfg=smpi/host-speed:\" to set its value." - "Check http://simgrid.org/simgrid/latest/doc/options.html#options_smpi_bench for more information."); - } + if (simgrid::config::is_default("smpi/host-speed")) { + XBT_INFO("You did not set the power of the host running the simulation. " + "The timings will certainly not be accurate. " + "Use the option \"--cfg=smpi/host-speed:\" to set its value. " + "Check " + "https://simgrid.org/doc/latest/Configuring_SimGrid.html#automatic-benchmarking-of-smpi-code for more " + "information."); + } - xbt_assert(xbt_cfg_get_double("smpi/cpu-threshold") >=0, - "The 'smpi/cpu-threshold' option cannot have negative values [anymore]. If you want to discard " - "the simulation of any computation, please use 'smpi/simulate-computation:no' instead."); + xbt_assert(simgrid::config::get_value("smpi/cpu-threshold") >= 0, + "The 'smpi/cpu-threshold' option cannot have negative values [anymore]. If you want to discard " + "the simulation of any computation, please use 'smpi/simulate-computation:no' instead."); } int smpi_enabled() { - return process_data != nullptr; + return not process_data.empty(); } void smpi_global_init() { - MPI_Group group; - if (not MC_is_active()) { global_timer = xbt_os_timer_new(); xbt_os_walltimer_start(global_timer); } - if (xbt_cfg_get_string("smpi/comp-adjustment-file")[0] != '\0') { - std::string filename {xbt_cfg_get_string("smpi/comp-adjustment-file")}; + std::string filename = simgrid::config::get_value("smpi/comp-adjustment-file"); + if (not filename.empty()) { std::ifstream fstream(filename); if (not fstream.is_open()) { xbt_die("Could not open file %s. Does it exist?", filename.c_str()); @@ -267,17 +288,15 @@ void smpi_global_init() // This map holds for each computation unit (such as "default" or "process1" etc.) // the configuration as given by the user (counter data as a pair of (counter_name, counter_counter)) // and the (computed) event_set. - std::map units2papi_setup; - if (xbt_cfg_get_string("smpi/papi-events")[0] != '\0') { + if (not simgrid::config::get_value("smpi/papi-events").empty()) { if (PAPI_library_init(PAPI_VER_CURRENT) != PAPI_VER_CURRENT) XBT_ERROR("Could not initialize PAPI library; is it correctly installed and linked?" - " Expected version is %i", - PAPI_VER_CURRENT); + " Expected version is %u", PAPI_VER_CURRENT); typedef boost::tokenizer> Tokenizer; boost::char_separator separator_units(";"); - std::string str = std::string(xbt_cfg_get_string("smpi/papi-events")); + std::string str = simgrid::config::get_value("smpi/papi-events"); Tokenizer tokens(str, separator_units); // Iterate over all the computational units. This could be processes, hosts, threads, ranks... You name it. @@ -306,17 +325,15 @@ void smpi_global_init() int event_code = PAPI_NULL; char* event_name = const_cast((*events_it).c_str()); - if (PAPI_event_name_to_code(event_name, &event_code) == PAPI_OK) { - if (PAPI_add_event(event_set, event_code) != PAPI_OK) { - XBT_ERROR("Could not add PAPI event '%s'. Skipping.", event_name); - continue; - } else { - XBT_DEBUG("Successfully added PAPI event '%s' to the event set.", event_name); - } - } else { + if (PAPI_event_name_to_code(event_name, &event_code) != PAPI_OK) { XBT_CRITICAL("Could not find PAPI event '%s'. Skipping.", event_name); continue; } + if (PAPI_add_event(event_set, event_code) != PAPI_OK) { + XBT_ERROR("Could not add PAPI event '%s'. Skipping.", event_name); + continue; + } + XBT_DEBUG("Successfully added PAPI event '%s' to the event set.", event_name); counters2values.push_back( // We cannot just pass *events_it, as this is of type const basic_string @@ -330,64 +347,16 @@ void smpi_global_init() } } #endif - - int smpirun = 0; - msg_bar_t finalization_barrier = nullptr; - if (process_count == 0){ - process_count = SIMIX_process_count(); - smpirun=1; - finalization_barrier = MSG_barrier_init(process_count); - } - smpi_universe_size = process_count; - process_data = new simgrid::smpi::Process*[process_count]; - for (int i = 0; i < process_count; i++) { - process_data[i] = new simgrid::smpi::Process(i, finalization_barrier); - } - //if the process was launched through smpirun script we generate a global mpi_comm_world - //if not, we let MPI_COMM_NULL, and the comm world will be private to each mpi instance - if (smpirun) { - group = new simgrid::smpi::Group(process_count); - MPI_COMM_WORLD = new simgrid::smpi::Comm(group, nullptr); - MPI_Attr_put(MPI_COMM_WORLD, MPI_UNIVERSE_SIZE, reinterpret_cast(process_count)); - - for (int i = 0; i < process_count; i++) - group->set_mapping(i, i); - } } void smpi_global_destroy() { - int count = smpi_process_count(); - smpi_bench_destroy(); smpi_shared_destroy(); - if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){ - delete MPI_COMM_WORLD->group(); - MSG_barrier_destroy(process_data[0]->finalization_barrier()); - }else{ - smpi_deployment_cleanup_instances(); - } - for (int i = 0; i < count; i++) { - if(process_data[i]->comm_self()!=MPI_COMM_NULL){ - simgrid::smpi::Comm::destroy(process_data[i]->comm_self()); - } - if(process_data[i]->comm_intra()!=MPI_COMM_NULL){ - simgrid::smpi::Comm::destroy(process_data[i]->comm_intra()); - } - xbt_os_timer_free(process_data[i]->timer()); - xbt_mutex_destroy(process_data[i]->mailboxes_mutex()); - delete process_data[i]; - } - delete[] process_data; - process_data = nullptr; - - if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){ - MPI_COMM_WORLD->cleanup_smp(); - MPI_COMM_WORLD->cleanup_attr(); - if(simgrid::smpi::Colls::smpi_coll_cleanup_callback!=nullptr) - simgrid::smpi::Colls::smpi_coll_cleanup_callback(); - delete MPI_COMM_WORLD; - } + smpi_deployment_cleanup_instances(); + + if (simgrid::smpi::Colls::smpi_coll_cleanup_callback != nullptr) + simgrid::smpi::Colls::smpi_coll_cleanup_callback(); MPI_COMM_WORLD = MPI_COMM_NULL; @@ -395,122 +364,115 @@ void smpi_global_destroy() xbt_os_timer_free(global_timer); } - xbt_free(index_to_process_data); - if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) + if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) smpi_destroy_global_memory_segments(); - smpi_free_static(); -} - -extern "C" { - -static void smpi_init_logs(){ - - /* Connect log categories. See xbt/log.c */ - - XBT_LOG_CONNECT(smpi); /* Keep this line as soon as possible in this function: xbt_log_appender_file.c depends on it - DO NOT connect this in XBT or so, or it will be useless to xbt_log_appender_file.c */ - XBT_LOG_CONNECT(instr_smpi); - XBT_LOG_CONNECT(smpi_bench); - XBT_LOG_CONNECT(smpi_coll); - XBT_LOG_CONNECT(smpi_colls); - XBT_LOG_CONNECT(smpi_comm); - XBT_LOG_CONNECT(smpi_datatype); - XBT_LOG_CONNECT(smpi_dvfs); - XBT_LOG_CONNECT(smpi_group); - XBT_LOG_CONNECT(smpi_host); - XBT_LOG_CONNECT(smpi_kernel); - XBT_LOG_CONNECT(smpi_mpi); - XBT_LOG_CONNECT(smpi_memory); - XBT_LOG_CONNECT(smpi_op); - XBT_LOG_CONNECT(smpi_pmpi); - XBT_LOG_CONNECT(smpi_process); - XBT_LOG_CONNECT(smpi_request); - XBT_LOG_CONNECT(smpi_replay); - XBT_LOG_CONNECT(smpi_rma); - XBT_LOG_CONNECT(smpi_shared); - XBT_LOG_CONNECT(smpi_utils); -} + if(simgrid::smpi::F2C::lookup() != nullptr) + simgrid::smpi::F2C::delete_lookup(); } static void smpi_init_options(){ - //return if already called - if (smpi_cpu_threshold > -1) - return; - simgrid::smpi::Colls::set_collectives(); - simgrid::smpi::Colls::smpi_coll_cleanup_callback=nullptr; - smpi_cpu_threshold = xbt_cfg_get_double("smpi/cpu-threshold"); - smpi_host_speed = xbt_cfg_get_double("smpi/host-speed"); - const char* smpi_privatize_option = xbt_cfg_get_string("smpi/privatization"); - if (std::strcmp(smpi_privatize_option, "no") == 0) - smpi_privatize_global_variables = SMPI_PRIVATIZE_NONE; - else if (std::strcmp(smpi_privatize_option, "yes") == 0) - smpi_privatize_global_variables = SMPI_PRIVATIZE_DEFAULT; - else if (std::strcmp(smpi_privatize_option, "mmap") == 0) - smpi_privatize_global_variables = SMPI_PRIVATIZE_MMAP; - else if (std::strcmp(smpi_privatize_option, "dlopen") == 0) - smpi_privatize_global_variables = SMPI_PRIVATIZE_DLOPEN; - - // Some compatibility stuff: - else if (std::strcmp(smpi_privatize_option, "1") == 0) - smpi_privatize_global_variables = SMPI_PRIVATIZE_DEFAULT; - else if (std::strcmp(smpi_privatize_option, "0") == 0) - smpi_privatize_global_variables = SMPI_PRIVATIZE_NONE; - - else - xbt_die("Invalid value for smpi/privatization: '%s'", smpi_privatize_option); - -#if defined(__FreeBSD__) - if (smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) { - XBT_INFO("Mixing mmap privatization is broken on FreeBSD, switching to dlopen privatization instead."); - smpi_privatize_global_variables = SMPI_PRIVATIZE_DLOPEN; - } -#endif + // return if already called + if (smpi_cpu_threshold > -1) + return; + simgrid::smpi::Colls::set_collectives(); + simgrid::smpi::Colls::smpi_coll_cleanup_callback = nullptr; + smpi_cpu_threshold = simgrid::config::get_value("smpi/cpu-threshold"); + smpi_host_speed = simgrid::config::get_value("smpi/host-speed"); + xbt_assert(smpi_host_speed > 0.0, "You're trying to set the host_speed to a non-positive value (given: %f)", smpi_host_speed); + std::string smpi_privatize_option = simgrid::config::get_value("smpi/privatization"); + if (smpi_privatize_option == "no" || smpi_privatize_option == "0") + smpi_privatize_global_variables = SmpiPrivStrategies::NONE; + else if (smpi_privatize_option == "yes" || smpi_privatize_option == "1") + smpi_privatize_global_variables = SmpiPrivStrategies::DEFAULT; + else if (smpi_privatize_option == "mmap") + smpi_privatize_global_variables = SmpiPrivStrategies::MMAP; + else if (smpi_privatize_option == "dlopen") + smpi_privatize_global_variables = SmpiPrivStrategies::DLOPEN; + else + xbt_die("Invalid value for smpi/privatization: '%s'", smpi_privatize_option.c_str()); + + if (not SMPI_switch_data_segment) { + XBT_DEBUG("Running without smpi_main(); disable smpi/privatization."); + smpi_privatize_global_variables = SmpiPrivStrategies::NONE; + } + if (not HAVE_WORKING_MMAP && smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) { + XBT_INFO("mmap privatization is broken on this platform, switching to dlopen privatization instead."); + smpi_privatize_global_variables = SmpiPrivStrategies::DLOPEN; + } - if (smpi_cpu_threshold < 0) - smpi_cpu_threshold = DBL_MAX; - - char* val = xbt_cfg_get_string("smpi/shared-malloc"); - if (not strcasecmp(val, "yes") || not strcmp(val, "1") || not strcasecmp(val, "on") || - not strcasecmp(val, "global")) { - smpi_cfg_shared_malloc = shmalloc_global; - } else if (not strcasecmp(val, "local")) { - smpi_cfg_shared_malloc = shmalloc_local; - } else if (not strcasecmp(val, "no") || not strcmp(val, "0") || not strcasecmp(val, "off")) { - smpi_cfg_shared_malloc = shmalloc_none; - } else { - xbt_die("Invalid value '%s' for option smpi/shared-malloc. Possible values: 'on' or 'global', 'local', 'off'", - val); - } + if (smpi_cpu_threshold < 0) + smpi_cpu_threshold = DBL_MAX; + + std::string val = simgrid::config::get_value("smpi/shared-malloc"); + if ((val == "yes") || (val == "1") || (val == "on") || (val == "global")) { + smpi_cfg_shared_malloc = SharedMallocType::GLOBAL; + } else if (val == "local") { + smpi_cfg_shared_malloc = SharedMallocType::LOCAL; + } else if ((val == "no") || (val == "0") || (val == "off")) { + smpi_cfg_shared_malloc = SharedMallocType::NONE; + } else { + xbt_die("Invalid value '%s' for option smpi/shared-malloc. Possible values: 'on' or 'global', 'local', 'off'", + val.c_str()); + } } typedef std::function smpi_entry_point_type; typedef int (* smpi_c_entry_point_type)(int argc, char **argv); typedef void (*smpi_fortran_entry_point_type)(); -static int smpi_run_entry_point(smpi_entry_point_type entry_point, std::vector args) +static int smpi_run_entry_point(const smpi_entry_point_type& entry_point, const std::string& executable_path, + std::vector args) { - char noarg[] = {'\0'}; - const int argc = args.size(); - std::unique_ptr argv(new char*[argc + 1]); - for (int i = 0; i != argc; ++i) - argv[i] = args[i].empty() ? noarg : &args[i].front(); - argv[argc] = nullptr; - - int res = entry_point(argc, argv.get()); + // copy C strings, we need them writable + std::vector* args4argv = new std::vector(args.size()); + std::transform(begin(args), end(args), begin(*args4argv), [](const std::string& s) { return xbt_strdup(s.c_str()); }); + + // set argv[0] to executable_path + xbt_free((*args4argv)[0]); + (*args4argv)[0] = xbt_strdup(executable_path.c_str()); + +#if !SMPI_IFORT + // take a copy of args4argv to keep reference of the allocated strings + const std::vector args2str(*args4argv); +#endif + int argc = args4argv->size(); + args4argv->push_back(nullptr); + char** argv = args4argv->data(); + + simgrid::smpi::ActorExt::init(); +#if SMPI_IFORT + for_rtl_init_ (&argc, argv); +#elif SMPI_FLANG + __io_set_argc(argc); + __io_set_argv(argv); +#elif SMPI_GFORTRAN + _gfortran_set_args(argc, argv); +#endif + int res = entry_point(argc, argv); + +#if SMPI_IFORT + for_rtl_finish_ (); +#else + for (char* s : args2str) + xbt_free(s); + delete args4argv; +#endif + if (res != 0){ XBT_WARN("SMPI process did not return 0. Return value : %d", res); - smpi_process()->set_return_value(res); + if (smpi_exit_status == 0) + smpi_exit_status = res; } return 0; } + // TODO, remove the number of functions involved here static smpi_entry_point_type smpi_resolve_function(void* handle) { smpi_fortran_entry_point_type entry_point_fortran = (smpi_fortran_entry_point_type)dlsym(handle, "user_main_"); if (entry_point_fortran != nullptr) { - return [entry_point_fortran](int argc, char** argv) { - smpi_process_init(&argc, &argv); + return [entry_point_fortran](int, char**) { entry_point_fortran(); return 0; }; @@ -524,10 +486,179 @@ static smpi_entry_point_type smpi_resolve_function(void* handle) return smpi_entry_point_type(); } -int smpi_main(const char* executable, int argc, char *argv[]) +static void smpi_copy_file(const std::string& src, const std::string& target, off_t fdin_size) { - srand(SMPI_RAND_SEED); + int fdin = open(src.c_str(), O_RDONLY); + xbt_assert(fdin >= 0, "Cannot read from %s. Please make sure that the file exists and is executable.", src.c_str()); + int fdout = open(target.c_str(), O_CREAT | O_RDWR, S_IRWXU); + xbt_assert(fdout >= 0, "Cannot write into %s", target.c_str()); + + XBT_DEBUG("Copy %" PRIdMAX " bytes into %s", static_cast(fdin_size), target.c_str()); +#if SG_HAVE_SENDFILE + ssize_t sent_size = sendfile(fdout, fdin, NULL, fdin_size); + if (sent_size == fdin_size) { + close(fdin); + close(fdout); + return; + } else if (sent_size != -1 || errno != ENOSYS) { + xbt_die("Error while copying %s: only %zd bytes copied instead of %" PRIdMAX " (errno: %d -- %s)", target.c_str(), + sent_size, static_cast(fdin_size), errno, strerror(errno)); + } +#endif + // If this point is reached, sendfile() actually is not available. Copy file by hand. + const int bufsize = 1024 * 1024 * 4; + char* buf = new char[bufsize]; + while (int got = read(fdin, buf, bufsize)) { + if (got == -1) { + xbt_assert(errno == EINTR, "Cannot read from %s", src.c_str()); + } else { + char* p = buf; + int todo = got; + while (int done = write(fdout, p, todo)) { + if (done == -1) { + xbt_assert(errno == EINTR, "Cannot write into %s", target.c_str()); + } else { + p += done; + todo -= done; + } + } + } + } + delete[] buf; + close(fdin); + close(fdout); +} + +#if not defined(__APPLE__) && not defined(__HAIKU__) +static int visit_libs(struct dl_phdr_info* info, size_t, void* data) +{ + char* libname = (char*)(data); + const char *path = info->dlpi_name; + if(strstr(path, libname)){ + strncpy(libname, path, 512); + return 1; + } + + return 0; +} +#endif + +static void smpi_init_privatization_dlopen(const std::string& executable) +{ + // Prepare the copy of the binary (get its size) + struct stat fdin_stat; + stat(executable.c_str(), &fdin_stat); + off_t fdin_size = fdin_stat.st_size; + + std::string libnames = simgrid::config::get_value("smpi/privatize-libs"); + if (not libnames.empty()) { + // split option + std::vector privatize_libs; + boost::split(privatize_libs, libnames, boost::is_any_of(";")); + + for (auto const& libname : privatize_libs) { + // load the library once to add it to the local libs, to get the absolute path + void* libhandle = dlopen(libname.c_str(), RTLD_LAZY); + // get library name from path + char fullpath[512] = {'\0'}; + strncpy(fullpath, libname.c_str(), 511); +#if not defined(__APPLE__) && not defined(__HAIKU__) + int ret = dl_iterate_phdr(visit_libs, fullpath); + if (ret == 0) + xbt_die("Can't find a linked %s - check the setting you gave to smpi/privatize-libs", fullpath); + else + XBT_DEBUG("Extra lib to privatize found : %s", fullpath); +#else + xbt_die("smpi/privatize-libs is not (yet) compatible with OSX"); +#endif + privatize_libs_paths.push_back(fullpath); + dlclose(libhandle); + } + } + simix_global->default_function = [executable, fdin_size](std::vector args) { + return std::function([executable, fdin_size, args] { + static std::size_t rank = 0; + // Copy the dynamic library: + std::string target_executable = + executable + "_" + std::to_string(getpid()) + "_" + std::to_string(rank) + ".so"; + + smpi_copy_file(executable, target_executable, fdin_size); + // if smpi/privatize-libs is set, duplicate pointed lib and link each executable copy to a different one. + std::vector target_libs; + for (auto const& libpath : privatize_libs_paths) { + // if we were given a full path, strip it + size_t index = libpath.find_last_of("/\\"); + std::string libname; + if (index != std::string::npos) + libname = libpath.substr(index + 1); + + if (not libname.empty()) { + // load the library to add it to the local libs, to get the absolute path + struct stat fdin_stat2; + stat(libpath.c_str(), &fdin_stat2); + off_t fdin_size2 = fdin_stat2.st_size; + + // Copy the dynamic library, the new name must be the same length as the old one + // just replace the name with 7 digits for the rank and the rest of the name. + unsigned int pad = 7; + if (libname.length() < pad) + pad = libname.length(); + std::string target_lib = + std::string(pad - std::to_string(rank).length(), '0') + std::to_string(rank) + libname.substr(pad); + target_libs.push_back(target_lib); + XBT_DEBUG("copy lib %s to %s, with size %lld", libpath.c_str(), target_lib.c_str(), (long long)fdin_size2); + smpi_copy_file(libpath, target_lib, fdin_size2); + + std::string sedcommand = "sed -i -e 's/" + libname + "/" + target_lib + "/g' " + target_executable; + int ret = system(sedcommand.c_str()); + if (ret != 0) + xbt_die("error while applying sed command %s \n", sedcommand.c_str()); + } + } + + rank++; + // Load the copy and resolve the entry point: + void* handle = dlopen(target_executable.c_str(), RTLD_LAZY | RTLD_LOCAL | WANT_RTLD_DEEPBIND); + int saved_errno = errno; + if (simgrid::config::get_value("smpi/keep-temps") == false) { + unlink(target_executable.c_str()); + for (const std::string& target_lib : target_libs) + unlink(target_lib.c_str()); + } + if (handle == nullptr) + xbt_die("dlopen failed: %s (errno: %d -- %s)", dlerror(), saved_errno, strerror(saved_errno)); + smpi_entry_point_type entry_point = smpi_resolve_function(handle); + if (not entry_point) + xbt_die("Could not resolve entry point"); + smpi_run_entry_point(entry_point, executable, args); + }); + }; +} + +static void smpi_init_privatization_no_dlopen(const std::string& executable) +{ + if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) + smpi_prepare_global_memory_segment(); + // Load the dynamic library and resolve the entry point: + void* handle = dlopen(executable.c_str(), RTLD_LAZY | RTLD_LOCAL); + if (handle == nullptr) + xbt_die("dlopen failed for %s: %s (errno: %d -- %s)", executable.c_str(), dlerror(), errno, strerror(errno)); + smpi_entry_point_type entry_point = smpi_resolve_function(handle); + if (not entry_point) + xbt_die("main not found in %s", executable.c_str()); + if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) + smpi_backup_global_memory_segment(); + + // Execute the same entry point for each simulated process: + simix_global->default_function = [entry_point, executable](std::vector args) { + return std::function( + [entry_point, executable, args] { smpi_run_entry_point(entry_point, executable, args); }); + }; +} + +int smpi_main(const char* executable, int argc, char* argv[]) +{ if (getenv("SMPI_PRETEND_CC") != nullptr) { /* Hack to ensure that smpicc can pretend to be a simple compiler. Particularly handy to pass it to the * configuration tools */ @@ -535,113 +666,27 @@ int smpi_main(const char* executable, int argc, char *argv[]) } TRACE_global_init(); - SIMIX_global_init(&argc, argv); - MSG_init(&argc,argv); SMPI_switch_data_segment = &smpi_switch_data_segment; - - simgrid::s4u::Host::onCreation.connect([](simgrid::s4u::Host& host) { - host.extension_set(new simgrid::smpi::SmpiHost(&host)); - }); - + sg_storage_file_system_init(); // parse the platform file: get the host list - SIMIX_create_environment(argv[1]); + simgrid::s4u::Engine::get_instance()->load_platform(argv[1]); SIMIX_comm_set_copy_data_callback(smpi_comm_copy_buffer_callback); smpi_init_options(); - - if (smpi_privatize_global_variables == SMPI_PRIVATIZE_DLOPEN) { - - std::string executable_copy = executable; - - // Prepare the copy of the binary (get its size) - struct stat fdin_stat; - stat(executable_copy.c_str(), &fdin_stat); - off_t fdin_size = fdin_stat.st_size; - static std::size_t rank = 0; - - simix_global->default_function = [executable_copy, fdin_size](std::vector args) { - return std::function([executable_copy, fdin_size, args] { - - // Copy the dynamic library: - std::string target_executable = executable_copy - + "_" + std::to_string(getpid()) - + "_" + std::to_string(rank++) + ".so"; - - int fdin = open(executable_copy.c_str(), O_RDONLY); - xbt_assert(fdin >= 0, "Cannot read from %s", executable_copy.c_str()); - int fdout = open(target_executable.c_str(), O_CREAT | O_RDWR, S_IRWXU); - xbt_assert(fdout >= 0, "Cannot write into %s", target_executable.c_str()); - -#if HAVE_SENDFILE - ssize_t sent_size = sendfile(fdout, fdin, NULL, fdin_size); - xbt_assert(sent_size == fdin_size, - "Error while copying %s: only %zd bytes copied instead of %ld (errno: %d -- %s)", - target_executable.c_str(), sent_size, fdin_size, errno, strerror(errno)); -#else - XBT_VERB("Copy %d bytes into %s", static_cast(fdin_size), target_executable.c_str()); - const int bufsize = 1024 * 1024 * 4; - char buf[bufsize]; - while (int got = read(fdin, buf, bufsize)) { - if (got == -1) { - xbt_assert(errno == EINTR, "Cannot read from %s", executable_copy.c_str()); - } else { - char* p = buf; - int todo = got; - while (int done = write(fdout, p, todo)) { - if (done == -1) { - xbt_assert(errno == EINTR, "Cannot write into %s", target_executable.c_str()); - } else { - p += done; - todo -= done; - } - } - } - } -#endif - close(fdin); - close(fdout); - - // Load the copy and resolve the entry point: - void* handle = dlopen(target_executable.c_str(), RTLD_LAZY | RTLD_LOCAL | RTLD_DEEPBIND); - int saved_errno = errno; - if (xbt_cfg_get_boolean("smpi/keep-temps") == false) - unlink(target_executable.c_str()); - if (handle == nullptr) - xbt_die("dlopen failed: %s (errno: %d -- %s)", dlerror(), saved_errno, strerror(saved_errno)); - smpi_entry_point_type entry_point = smpi_resolve_function(handle); - if (not entry_point) - xbt_die("Could not resolve entry point"); - - smpi_run_entry_point(entry_point, args); - }); - }; - - } - else { - - // Load the dynamic library and resolve the entry point: - void* handle = dlopen(executable, RTLD_LAZY | RTLD_LOCAL | RTLD_DEEPBIND); - if (handle == nullptr) - xbt_die("dlopen failed for %s: %s (errno: %d -- %s)", executable, dlerror(), errno, strerror(errno)); - smpi_entry_point_type entry_point = smpi_resolve_function(handle); - if (not entry_point) - xbt_die("main not found in %s", executable); - // TODO, register the executable for SMPI privatization - - // Execute the same entry point for each simulated process: - simix_global->default_function = [entry_point](std::vector args) { - return std::function([entry_point, args] { - smpi_run_entry_point(entry_point, args); - }); - }; - - } - - SIMIX_launch_application(argv[2]); + if (smpi_privatize_global_variables == SmpiPrivStrategies::DLOPEN) + smpi_init_privatization_dlopen(executable); + else + smpi_init_privatization_no_dlopen(executable); SMPI_init(); + simgrid::s4u::Engine::get_instance()->load_deployment(argv[2]); + SMPI_app_instance_register(smpi_default_instance_name.c_str(), nullptr, + process_data.size()); // This call has a side effect on process_count... + MPI_COMM_WORLD = *smpi_deployment_comm_world(smpi_default_instance_name); + smpi_universe_size = process_count; + /* Clean IO before the run */ fflush(stdout); @@ -654,7 +699,7 @@ int smpi_main(const char* executable, int argc, char *argv[]) SIMIX_run(); xbt_os_walltimer_stop(global_timer); - if (xbt_cfg_get_boolean("smpi/display-timing")){ + if (simgrid::config::get_value("smpi/display-timing")) { double global_time = xbt_os_timer_elapsed(global_timer); XBT_INFO("Simulated time: %g seconds. \n\n" "The simulation took %g seconds (after parsing and platform setup)\n" @@ -666,31 +711,31 @@ int smpi_main(const char* executable, int argc, char *argv[]) "You may want to use sampling functions or trace replay to reduce this."); } } - int count = smpi_process_count(); - int ret = 0; - for (int i = 0; i < count; i++) { - if(process_data[i]->return_value()!=0){ - ret=process_data[i]->return_value();//return first non 0 value - break; - } - } smpi_global_destroy(); - TRACE_end(); - - return ret; + return smpi_exit_status; } // Called either directly from the user code, or from the code called by smpirun void SMPI_init(){ - smpi_init_logs(); + simgrid::s4u::Actor::on_creation.connect([](simgrid::s4u::Actor& actor) { + if (not actor.is_daemon()) { + process_data.insert({&actor, new simgrid::smpi::ActorExt(&actor, nullptr)}); + } + }); + simgrid::s4u::Actor::on_destruction.connect([](simgrid::s4u::Actor const& actor) { + auto it = process_data.find(&actor); + if (it != process_data.end()) { + delete it->second; + process_data.erase(it); + } + }); + simgrid::s4u::Host::on_creation.connect( + [](simgrid::s4u::Host& host) { host.extension_set(new simgrid::smpi::Host(&host)); }); + smpi_init_options(); smpi_global_init(); smpi_check_options(); - TRACE_smpi_alloc(); - simgrid::surf::surfExitCallbacks.connect(TRACE_smpi_release); - if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) - smpi_initialize_global_memory_segments(); } void SMPI_finalize(){ @@ -698,27 +743,7 @@ void SMPI_finalize(){ } void smpi_mpi_init() { + smpi_init_fortran_types(); if(smpi_init_sleep > 0) simcall_process_sleep(smpi_init_sleep); } - -double smpi_mpi_wtime(){ - double time; - if (smpi_process()->initialized() != 0 && smpi_process()->finalized() == 0 && smpi_process()->sampling() == 0) { - smpi_bench_end(); - time = SIMIX_get_clock(); - // to avoid deadlocks if used as a break condition, such as - // while (MPI_Wtime(...) < time_limit) { - // .... - // } - // because the time will not normally advance when only calls to MPI_Wtime - // are made -> deadlock (MPI_Wtime never reaches the time limit) - if(smpi_wtime_sleep > 0) - simcall_process_sleep(smpi_wtime_sleep); - smpi_bench_begin(); - } else { - time = SIMIX_get_clock(); - } - return time; -} -