X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/00b2e701ab8ca3dd350bb8e9d65313986cabe8f0..4d9809aa64447811ca65a1242d4f9e07972c6472:/src/smpi/internals/smpi_global.cpp diff --git a/src/smpi/internals/smpi_global.cpp b/src/smpi/internals/smpi_global.cpp index bb67c34626..dcc7d0fed2 100644 --- a/src/smpi/internals/smpi_global.cpp +++ b/src/smpi/internals/smpi_global.cpp @@ -3,16 +3,20 @@ /* 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 "smpi_host.hpp" #include "mc/mc.h" #include "simgrid/s4u/Engine.hpp" #include "smpi_coll.hpp" -#include "smpi_process.hpp" -#include "src/msg/msg_private.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 @@ -27,7 +31,7 @@ # define MAC_OS_X_VERSION_10_12 101200 # endif # define HAVE_WORKING_MMAP (MAC_OS_X_VERSION_MIN_REQUIRED >= MAC_OS_X_VERSION_10_12) -#elif defined(__FreeBSD__) +#elif defined(__FreeBSD__) || defined(__FreeBSD_kernel__) # define HAVE_WORKING_MMAP 0 #else # define HAVE_WORKING_MMAP 1 @@ -41,12 +45,24 @@ 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 !defined(RTLD_DEEPBIND) || HAVE_SANITIZER_ADDRESS || HAVE_SANITIZER_THREAD +#define WANT_RTLD_DEEPBIND 0 +#else +#define WANT_RTLD_DEEPBIND RTLD_DEEPBIND #endif #if HAVE_PAPI @@ -59,7 +75,7 @@ using simgrid::s4u::Actor; using simgrid::s4u::ActorPtr; std::unordered_map location2speedup; -static std::map process_data; +static std::map process_data; int process_count = 0; static int smpi_exit_status = 0; int smpi_universe_size = 0; @@ -83,8 +99,6 @@ MPI_Errhandler *MPI_ERRHANDLER_NULL = nullptr; // 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_wtime_sleep( - "smpi/wtime", "Minimum time to inject inside a call to MPI_Wtime", 0.0); static simgrid::config::Flag smpi_init_sleep( "smpi/init", "Time to inject inside a call to MPI_Init", 0.0); @@ -95,16 +109,17 @@ int smpi_process_count() return process_count; } -simgrid::smpi::Process* smpi_process() +simgrid::smpi::ActorExt* smpi_process() { ActorPtr me = 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->get_impl()->get_user_data()); - return static_cast(msgExt->data); + + return process_data.at(me); } -simgrid::smpi::Process* smpi_process_remote(ActorPtr actor) +simgrid::smpi::ActorExt* smpi_process_remote(ActorPtr actor) { return process_data.at(actor); } @@ -114,19 +129,15 @@ MPI_Comm smpi_process_comm_self(){ } void smpi_process_init(int *argc, char ***argv){ - simgrid::smpi::Process::init(argc, argv); -} - -int smpi_process_index(){ - return simgrid::s4u::this_actor::get_pid(); + simgrid::smpi::ActorExt::init(); } void * smpi_process_get_user_data(){ - return smpi_process()->get_user_data(); + return Actor::self()->get_impl()->get_user_data(); } void smpi_process_set_user_data(void *data){ - return smpi_process()->set_user_data(data); + Actor::self()->get_impl()->set_user_data(data); } @@ -356,6 +367,8 @@ void smpi_global_destroy() if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) smpi_destroy_global_memory_segments(); smpi_free_static(); + if(simgrid::smpi::F2C::lookup() != nullptr) + simgrid::smpi::F2C::delete_lookup(); } static void smpi_init_options(){ @@ -410,16 +423,44 @@ 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(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); if (smpi_exit_status == 0) @@ -435,7 +476,6 @@ 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); entry_point_fortran(); return 0; }; @@ -456,31 +496,36 @@ static void smpi_copy_file(std::string src, std::string target, off_t fdin_size) 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 %ld bytes into %s", static_cast(fdin_size), target.c_str()); + XBT_DEBUG("Copy %" PRIdMAX " bytes into %s", static_cast(fdin_size), target.c_str()); + bool slow_copy = true; #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.c_str(), sent_size, fdin_size, errno, strerror(errno)); -#else - 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", 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; + if (sent_size == fdin_size) + slow_copy = false; + 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 (slow_copy) { + 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", 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; + } } } } } -#endif close(fdin); close(fdout); } @@ -499,7 +544,7 @@ static int visit_libs(struct dl_phdr_info* info, size_t, void* data) } #endif -static void smpi_init_privatization_dlopen(std::string executable) +static void smpi_init_privatization_dlopen(const std::string& executable) { // Prepare the copy of the binary (get its size) struct stat fdin_stat; @@ -535,14 +580,13 @@ static void smpi_init_privatization_dlopen(std::string executable) simix_global->default_function = [executable, fdin_size](std::vector args) { return std::function([executable, fdin_size, args] { - // 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::string target_lib; + 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("/\\"); @@ -561,8 +605,9 @@ static void smpi_init_privatization_dlopen(std::string executable) unsigned int pad = 7; if (libname.length() < pad) pad = libname.length(); - target_lib = + 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); @@ -575,11 +620,11 @@ static void smpi_init_privatization_dlopen(std::string executable) rank++; // Load the copy and resolve the entry point: - void* handle = dlopen(target_executable.c_str(), RTLD_LAZY | RTLD_LOCAL | RTLD_DEEPBIND); + 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()); - if (not target_lib.empty()) + for (const std::string& target_lib : target_libs) unlink(target_lib.c_str()); } if (handle == nullptr) @@ -587,12 +632,12 @@ static void smpi_init_privatization_dlopen(std::string executable) 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); + smpi_run_entry_point(entry_point, executable, args); }); }; } -static void smpi_init_privatization_no_dlopen(std::string executable) +static void smpi_init_privatization_no_dlopen(const std::string& executable) { if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) smpi_prepare_global_memory_segment(); @@ -607,8 +652,9 @@ static void smpi_init_privatization_no_dlopen(std::string executable) smpi_backup_global_memory_segment(); // 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_global->default_function = [entry_point, executable](std::vector args) { + return std::function( + [entry_point, executable, args] { smpi_run_entry_point(entry_point, executable, args); }); }; } @@ -623,20 +669,18 @@ int smpi_main(const char* executable, int argc, char* argv[]) } TRACE_global_init(); - SIMIX_global_init(&argc, argv); - MSG_init(&argc, argv); // FIXME Remove this MSG call. Once it's removed, we can remove the msg header include as well SMPI_switch_data_segment = &smpi_switch_data_segment; // TODO This will not be executed in the case where smpi_main is not called, // e.g., not for smpi_msg_masterslave. This should be moved to another location - // that is always called -- maybe close to Actor::onCreation? + // that is always called -- maybe close to Actor::on_creation? simgrid::s4u::Host::on_creation.connect( [](simgrid::s4u::Host& host) { host.extension_set(new simgrid::smpi::Host(&host)); }); // 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(); @@ -646,7 +690,7 @@ int smpi_main(const char* executable, int argc, char* argv[]) smpi_init_privatization_no_dlopen(executable); SMPI_init(); - SIMIX_launch_application(argv[2]); + 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); @@ -685,7 +729,7 @@ int smpi_main(const char* executable, int argc, char* argv[]) void SMPI_init(){ simgrid::s4u::Actor::on_creation.connect([](simgrid::s4u::ActorPtr actor) { if (not actor->is_daemon()) { - process_data.insert({actor, new simgrid::smpi::Process(actor, nullptr)}); + process_data.insert({actor, new simgrid::smpi::ActorExt(actor, nullptr)}); } }); simgrid::s4u::Actor::on_destruction.connect([](simgrid::s4u::ActorPtr actor) { @@ -699,7 +743,6 @@ void SMPI_init(){ smpi_init_options(); smpi_global_init(); smpi_check_options(); - simgrid::s4u::on_simulation_end.connect(TRACE_smpi_release); } void SMPI_finalize(){ @@ -707,27 +750,7 @@ 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; -} -