X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/b214b15ec055be50978640d04708fb5f64abe53c..bfd968c929b70a92b09b30760def66ab43a239a3:/src/smpi/internals/smpi_global.cpp diff --git a/src/smpi/internals/smpi_global.cpp b/src/smpi/internals/smpi_global.cpp index ab1ebfb125..5cf70257cb 100644 --- a/src/smpi/internals/smpi_global.cpp +++ b/src/smpi/internals/smpi_global.cpp @@ -1,4 +1,4 @@ -/* Copyright (c) 2007-2019. The SimGrid Team. All rights reserved. */ +/* Copyright (c) 2007-2021. 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. */ @@ -14,12 +14,17 @@ #include "src/simix/smx_private.hpp" #include "src/smpi/include/smpi_actor.hpp" #include "xbt/config.hpp" +#include "xbt/file.hpp" #include +#include #include /* split */ #include +#include +#include #include #include /* intmax_t */ +#include /* strerror */ #include #include #include @@ -60,8 +65,7 @@ XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_kernel, smpi, "Logging specific to SMPI (ke #endif #if HAVE_PAPI -std::string papi_default_config_name = "default"; -std::map units2papi_setup; +std::map> units2papi_setup; #endif std::unordered_map location2speedup; @@ -128,17 +132,18 @@ void smpi_comm_set_copy_data_callback(void (*callback) (smx_activity_t, void*, s 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); + saved_callback(comm, buff, size); }; } -static void memcpy_private(void* dest, const void* src, std::vector>& private_blocks) +static void memcpy_private(void* dest, const void* src, const std::vector>& private_blocks) { for (auto const& block : private_blocks) memcpy((uint8_t*)dest+block.first, (uint8_t*)src+block.first, block.second-block.first); } -static void check_blocks(std::vector> &private_blocks, size_t buff_size) { +static void check_blocks(const std::vector>& private_blocks, size_t buff_size) +{ for (auto const& block : private_blocks) xbt_assert(block.first <= block.second && block.second <= buff_size, "Oops, bug in shared malloc."); } @@ -163,7 +168,7 @@ void smpi_comm_copy_buffer_callback(simgrid::kernel::activity::CommImpl* comm, v XBT_DEBUG("Copy the data over"); if(smpi_is_shared(buff, src_private_blocks, &src_offset)) { src_private_blocks = shift_and_frame_private_blocks(src_private_blocks, src_offset, buff_size); - if (src_private_blocks.size()==0){//simple shared malloc ... return. + if (src_private_blocks.empty()) { // simple shared malloc ... return. XBT_VERB("Sender is shared. Let's ignore it."); smpi_cleanup_comm_after_copy(comm, buff); return; @@ -171,11 +176,11 @@ void smpi_comm_copy_buffer_callback(simgrid::kernel::activity::CommImpl* comm, v } else { src_private_blocks.clear(); - src_private_blocks.push_back(std::make_pair(0, buff_size)); + src_private_blocks.emplace_back(0, buff_size); } if (smpi_is_shared((char*)comm->dst_buff_, dst_private_blocks, &dst_offset)) { dst_private_blocks = shift_and_frame_private_blocks(dst_private_blocks, dst_offset, buff_size); - if (dst_private_blocks.size()==0){//simple shared malloc ... return. + if (dst_private_blocks.empty()) { // simple shared malloc ... return. XBT_VERB("Receiver is shared. Let's ignore it."); smpi_cleanup_comm_after_copy(comm, buff); return; @@ -183,7 +188,7 @@ void smpi_comm_copy_buffer_callback(simgrid::kernel::activity::CommImpl* comm, v } else { dst_private_blocks.clear(); - dst_private_blocks.push_back(std::make_pair(0, buff_size)); + dst_private_blocks.emplace_back(0, buff_size); } check_blocks(src_private_blocks, buff_size); check_blocks(dst_private_blocks, buff_size); @@ -194,8 +199,8 @@ void smpi_comm_copy_buffer_callback(simgrid::kernel::activity::CommImpl* comm, v (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)); + smpi_switch_data_segment(comm->src_actor_->get_iface()); + tmpbuff = xbt_malloc(buff_size); memcpy_private(tmpbuff, buff, private_blocks); } @@ -203,7 +208,7 @@ void smpi_comm_copy_buffer_callback(simgrid::kernel::activity::CommImpl* comm, v ((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()); + smpi_switch_data_segment(comm->dst_actor_->get_iface()); } XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff, comm->dst_buff_); memcpy_private(comm->dst_buff_, tmpbuff, private_blocks); @@ -234,7 +239,7 @@ static void smpi_init_papi() XBT_ERROR("Could not initialize PAPI library; is it correctly installed and linked?" " Expected version is %u", PAPI_VER_CURRENT); - typedef boost::tokenizer> Tokenizer; + using Tokenizer = boost::tokenizer>; boost::char_separator separator_units(";"); std::string str = smpi_cfg_papi_events_file(); Tokenizer tokens(str, separator_units); @@ -252,19 +257,16 @@ static void smpi_init_papi() } // NOTE: We cannot use a map here, as we must obey the order of the counters - // This is important for PAPI: We need to map the values of counters back - // to the event_names (so, when PAPI_read() has finished)! + // This is important for PAPI: We need to map the values of counters back to the event_names (so, when PAPI_read() + // has finished)! papi_counter_t counters2values; - // Iterate over all counters that were specified for this specific - // unit. - // Note that we need to remove the name of the unit - // (that could also be the "default" value), which always comes first. - // Hence, we start at ++(events.begin())! + // Iterate over all counters that were specified for this specific unit. + // Note that we need to remove the name of the unit (that could also be the "default" value), which always comes + // first. Hence, we start at ++(events.begin())! for (Tokenizer::iterator events_it = ++(event_tokens.begin()); events_it != event_tokens.end(); ++events_it) { - int event_code = PAPI_NULL; - char* event_name = const_cast((*events_it).c_str()); + auto* event_name = const_cast((*events_it).c_str()); if (PAPI_event_name_to_code(event_name, &event_code) != PAPI_OK) { XBT_CRITICAL("Could not find PAPI event '%s'. Skipping.", event_name); continue; @@ -275,9 +277,7 @@ static void smpi_init_papi() } 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 - std::make_pair(std::string(*events_it), 0LL)); + counters2values.emplace_back(*events_it, 0LL); } std::string unit_name = *(event_tokens.begin()); @@ -289,17 +289,15 @@ static void smpi_init_papi() #endif } - - -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)(); +using smpi_entry_point_type = std::function; +using smpi_c_entry_point_type = int (*)(int argc, char** argv); +using smpi_fortran_entry_point_type = void (*)(); template static int smpi_run_entry_point(const F& entry_point, const std::string& executable_path, std::vector args) { // copy C strings, we need them writable - std::vector* args4argv = new std::vector(args.size()); + auto* 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 @@ -344,7 +342,7 @@ static int smpi_run_entry_point(const F& entry_point, const std::string& executa // 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_"); + auto* entry_point_fortran = reinterpret_cast(dlsym(handle, "user_main_")); if (entry_point_fortran != nullptr) { return [entry_point_fortran](int, char**) { entry_point_fortran(); @@ -352,7 +350,7 @@ static smpi_entry_point_type smpi_resolve_function(void* handle) }; } - smpi_c_entry_point_type entry_point = (smpi_c_entry_point_type)dlsym(handle, "main"); + auto* entry_point = reinterpret_cast(dlsym(handle, "main")); if (entry_point != nullptr) { return entry_point; } @@ -364,12 +362,14 @@ static void smpi_copy_file(const std::string& src, const std::string& target, of { 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_ATTRIB_UNUSED int unlink_status = unlink(target.c_str()); + xbt_assert(unlink_status == 0 || errno == ENOENT, "Failed to unlink file %s: %s", target.c_str(), strerror(errno)); + int fdout = open(target.c_str(), O_CREAT | O_RDWR | O_EXCL, S_IRWXU); + xbt_assert(fdout >= 0, "Cannot write into %s: %s", target.c_str(), strerror(errno)); 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); + ssize_t sent_size = sendfile(fdout, fdin, nullptr, fdin_size); if (sent_size == fdin_size) { close(fdin); close(fdout); @@ -380,15 +380,14 @@ static void smpi_copy_file(const std::string& src, const std::string& target, of } #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)) { + std::vector buf(1024 * 1024 * 4); + while (ssize_t got = read(fdin, buf.data(), buf.size())) { 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)) { + const unsigned char* p = buf.data(); + ssize_t todo = got; + while (ssize_t done = write(fdout, p, todo)) { if (done == -1) { xbt_assert(errno == EINTR, "Cannot write into %s", target.c_str()); } else { @@ -398,7 +397,6 @@ static void smpi_copy_file(const std::string& src, const std::string& target, of } } } - delete[] buf; close(fdin); close(fdout); } @@ -406,13 +404,12 @@ static void smpi_copy_file(const std::string& src, const std::string& target, of #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); + auto* libname = static_cast(data); + std::string path = info->dlpi_name; + if (path.find(*libname) != std::string::npos) { + *libname = std::move(path); return 1; } - return 0; } #endif @@ -433,27 +430,30 @@ static void smpi_init_privatization_dlopen(const std::string& executable) 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); + xbt_assert(libhandle != nullptr, + "Cannot dlopen %s - check your settings in smpi/privatize-libs", libname.c_str()); // get library name from path - char fullpath[512] = {'\0'}; - strncpy(fullpath, libname.c_str(), 511); + std::string fullpath = libname; #if not defined(__APPLE__) && not defined(__HAIKU__) - xbt_assert(0 != dl_iterate_phdr(visit_libs, fullpath), - "Can't find a linked %s - check your settings in smpi/privatize-libs", fullpath); - XBT_DEBUG("Extra lib to privatize '%s' found", fullpath); + XBT_ATTRIB_UNUSED int dl_iterate_res = dl_iterate_phdr(visit_libs, &fullpath); + xbt_assert(dl_iterate_res != 0, "Can't find a linked %s - check your settings in smpi/privatize-libs", + fullpath.c_str()); + XBT_DEBUG("Extra lib to privatize '%s' found", fullpath.c_str()); #else xbt_die("smpi/privatize-libs is not (yet) compatible with OSX nor with Haiku"); #endif - privatize_libs_paths.push_back(fullpath); + privatize_libs_paths.emplace_back(std::move(fullpath)); dlclose(libhandle); } } - simix_global->default_function = [executable, fdin_size](std::vector args) { + simgrid::s4u::Engine::get_instance()->register_default([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"; + simgrid::xbt::Path path(executable); + std::string target_executable = simgrid::config::get_value("smpi/tmpdir") + "/" + + path.get_base_name() + "_" + 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. @@ -473,17 +473,16 @@ static void smpi_init_privatization_dlopen(const std::string& executable) // 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); + auto pad = std::min(7, libname.length()); + std::string target_libname = std::string(pad - std::to_string(rank).length(), '0') + std::to_string(rank) + libname.substr(pad); + std::string target_lib = simgrid::config::get_value("smpi/tmpdir") + "/" + target_libname; 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; - xbt_assert(system(sedcommand.c_str()) == 0, "error while applying sed command %s \n", sedcommand.c_str()); + std::string sedcommand = "sed -i -e 's/" + libname + "/" + target_libname + "/g' " + target_executable; + int status = system(sedcommand.c_str()); + xbt_assert(status == 0, "error while applying sed command %s \n", sedcommand.c_str()); } } @@ -491,19 +490,21 @@ static void smpi_init_privatization_dlopen(const std::string& executable) // 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) { + if (not simgrid::config::get_value("smpi/keep-temps")) { unlink(target_executable.c_str()); for (const std::string& target_lib : target_libs) unlink(target_lib.c_str()); } - xbt_assert(handle != nullptr, "dlopen failed: %s (errno: %d -- %s)", dlerror(), saved_errno, - strerror(saved_errno)); + xbt_assert(handle != nullptr, + "dlopen failed: %s (errno: %d -- %s).\nError: Did you compile the program with a SMPI-specific " + "compiler (spmicc or friends)?", + dlerror(), saved_errno, strerror(saved_errno)); smpi_entry_point_type entry_point = smpi_resolve_function(handle); xbt_assert(entry_point, "Could not resolve entry point"); smpi_run_entry_point(entry_point, executable, args); }); - }; + }); } static void smpi_init_privatization_no_dlopen(const std::string& executable) @@ -522,10 +523,10 @@ static void smpi_init_privatization_no_dlopen(const std::string& executable) smpi_backup_global_memory_segment(); // Execute the same entry point for each simulated process: - simix_global->default_function = [entry_point, executable](std::vector args) { + simgrid::s4u::Engine::get_instance()->register_default([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[]) @@ -538,7 +539,7 @@ int smpi_main(const char* executable, int argc, char* argv[]) SMPI_switch_data_segment = &smpi_switch_data_segment; smpi_init_options(); - TRACE_global_init(); + simgrid::instr::init(); SIMIX_global_init(&argc, argv); auto engine = simgrid::s4u::Engine::get_instance(); @@ -560,7 +561,7 @@ int smpi_main(const char* executable, int argc, char* argv[]) /* This is a ... heavy way to count the MPI ranks */ int rank_counts = 0; - simgrid::s4u::Actor::on_creation.connect([&rank_counts](simgrid::s4u::Actor& actor) { + simgrid::s4u::Actor::on_creation.connect([&rank_counts](const simgrid::s4u::Actor& actor) { if (not actor.is_daemon()) rank_counts++; }); @@ -576,7 +577,6 @@ int smpi_main(const char* executable, int argc, char* argv[]) if (MC_is_active()) { MC_run(); } else { - SIMIX_run(); xbt_os_walltimer_stop(global_timer); @@ -634,8 +634,24 @@ void SMPI_finalize() if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) smpi_destroy_global_memory_segments(); - if (simgrid::smpi::F2C::lookup() != nullptr) - simgrid::smpi::F2C::delete_lookup(); + + if (simgrid::smpi::F2C::lookup() != nullptr && + simgrid::smpi::F2C::lookup()->size() > simgrid::smpi::F2C::get_num_default_handles()) { + XBT_WARN("Probable memory leaks in your code: SMPI detected %zu unfreed MPI handles : " + "display types and addresses (n max) with --cfg=smpi/list-leaks:n.\n" + "Running smpirun with -wrapper \"valgrind --leak-check=full\" can provide more information", + simgrid::smpi::F2C::lookup()->size() - simgrid::smpi::F2C::get_num_default_handles()); + int n = simgrid::config::get_value("smpi/list-leaks"); + for (auto const& p : *simgrid::smpi::F2C::lookup()) { + static int printed = 0; + if (printed >= n) + break; + if (p.first >= simgrid::smpi::F2C::get_num_default_handles()) { + XBT_WARN("Leak %p of type %s", p.second, boost::core::demangle(typeid(*(p.second)).name()).c_str()); + printed++; + } + } + } } void smpi_mpi_init() { @@ -643,3 +659,7 @@ void smpi_mpi_init() { if(smpi_init_sleep > 0) simgrid::s4u::this_actor::sleep_for(smpi_init_sleep); } + +void SMPI_thread_create() { + TRACE_smpi_init(simgrid::s4u::this_actor::get_pid(), __func__); +}