1 /* Copyright (c) 2007-2021. The SimGrid Team. All rights reserved. */
3 /* This program is free software; you can redistribute it and/or modify it
4 * under the terms of the license (GNU LGPL) which comes with this package. */
7 #include "simgrid/plugins/file_system.h"
8 #include "simgrid/s4u/Engine.hpp"
9 #include "smpi_coll.hpp"
10 #include "smpi_config.hpp"
11 #include "smpi_f2c.hpp"
12 #include "smpi_host.hpp"
13 #include "src/kernel/EngineImpl.hpp"
14 #include "src/kernel/activity/CommImpl.hpp"
15 #include "src/simix/smx_private.hpp"
16 #include "src/smpi/include/smpi_actor.hpp"
17 #include "xbt/config.hpp"
18 #include "xbt/file.hpp"
22 #include <boost/algorithm/string.hpp> /* split */
23 #include <boost/tokenizer.hpp>
26 #include <cstdint> /* intmax_t */
27 #include <cstring> /* strerror */
34 #include <sys/sendfile.h>
41 #if not defined(__APPLE__) && not defined(__HAIKU__)
45 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_kernel, smpi, "Logging specific to SMPI (kernel)");
48 extern "C" void for_rtl_init_ (int *, char **);
49 extern "C" void for_rtl_finish_ ();
51 extern "C" void __io_set_argc(int);
52 extern "C" void __io_set_argv(char **);
54 extern "C" void _gfortran_set_args(int, char **);
57 /* RTLD_DEEPBIND is a bad idea of GNU ld that obviously does not exist on other platforms
58 * See https://www.akkadia.org/drepper/dsohowto.pdf
59 * and https://lists.freebsd.org/pipermail/freebsd-current/2016-March/060284.html
61 #if !RTLD_DEEPBIND || HAVE_SANITIZER_ADDRESS || HAVE_SANITIZER_THREAD
62 #define WANT_RTLD_DEEPBIND 0
64 #define WANT_RTLD_DEEPBIND RTLD_DEEPBIND
68 std::map</* computation unit name */ std::string, papi_process_data, std::less<>> units2papi_setup;
71 std::unordered_map<std::string, double> location2speedup;
73 static int smpi_exit_status = 0;
74 xbt_os_timer_t global_timer;
75 static std::vector<std::string> privatize_libs_paths;
77 // No instance gets manually created; check also the smpirun.in script as
78 // this default name is used there as well (when the <actor> tag is generated).
79 static const std::string smpi_default_instance_name("smpirun");
80 static simgrid::config::Flag<double> smpi_init_sleep(
81 "smpi/init", "Time to inject inside a call to MPI_Init", 0.0);
83 static simgrid::config::Flag<std::string>
84 smpi_hostfile("smpi/hostfile",
85 "Classical MPI hostfile containing list of machines to dispatch "
86 "the processes, one per line",
89 static simgrid::config::Flag<std::string> smpi_replay("smpi/replay",
90 "Replay a trace instead of executing the application", "");
92 static simgrid::config::Flag<int> smpi_np("smpi/np", "Number of processes to be created", 0);
94 static simgrid::config::Flag<int> smpi_map("smpi/map", "Display the mapping between nodes and processes", 0);
96 void (*smpi_comm_copy_data_callback)(simgrid::kernel::activity::CommImpl*, void*,
97 size_t) = &smpi_comm_copy_buffer_callback;
99 simgrid::smpi::ActorExt* smpi_process()
101 simgrid::s4u::ActorPtr me = simgrid::s4u::Actor::self();
103 if (me == nullptr) // This happens sometimes (eg, when linking against NS3 because it pulls openMPI...)
106 return me->extension<simgrid::smpi::ActorExt>();
109 simgrid::smpi::ActorExt* smpi_process_remote(simgrid::s4u::ActorPtr actor)
111 if (actor.get() == nullptr)
113 return actor->extension<simgrid::smpi::ActorExt>();
116 MPI_Comm smpi_process_comm_self(){
117 return smpi_process()->comm_self();
120 MPI_Info smpi_process_info_env(){
121 return smpi_process()->info_env();
124 void * smpi_process_get_user_data(){
125 return simgrid::s4u::Actor::self()->get_data();
128 void smpi_process_set_user_data(void *data){
129 simgrid::s4u::Actor::self()->set_data(data);
132 void smpi_comm_set_copy_data_callback(void (*callback) (smx_activity_t, void*, size_t))
134 static void (*saved_callback)(smx_activity_t, void*, size_t);
135 saved_callback = callback;
136 smpi_comm_copy_data_callback = [](simgrid::kernel::activity::CommImpl* comm, void* buff, size_t size) {
137 saved_callback(comm, buff, size);
141 static void memcpy_private(void* dest, const void* src, const std::vector<std::pair<size_t, size_t>>& private_blocks)
143 for (auto const& block : private_blocks)
144 memcpy((uint8_t*)dest+block.first, (uint8_t*)src+block.first, block.second-block.first);
147 static void check_blocks(const std::vector<std::pair<size_t, size_t>>& private_blocks, size_t buff_size)
149 for (auto const& block : private_blocks)
150 xbt_assert(block.first <= block.second && block.second <= buff_size, "Oops, bug in shared malloc.");
153 static void smpi_cleanup_comm_after_copy(simgrid::kernel::activity::CommImpl* comm, void* buff){
154 if (comm->detached()) {
155 // if this is a detached send, the source buffer was duplicated by SMPI
156 // sender to make the original buffer available to the application ASAP
158 //It seems that the request is used after the call there this should be free somewhere else but where???
159 //xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free
160 comm->src_buff_ = nullptr;
164 void smpi_comm_copy_buffer_callback(simgrid::kernel::activity::CommImpl* comm, void* buff, size_t buff_size)
166 size_t src_offset = 0;
167 size_t dst_offset = 0;
168 std::vector<std::pair<size_t, size_t>> src_private_blocks;
169 std::vector<std::pair<size_t, size_t>> dst_private_blocks;
170 XBT_DEBUG("Copy the data over");
171 if(smpi_is_shared(buff, src_private_blocks, &src_offset)) {
172 src_private_blocks = shift_and_frame_private_blocks(src_private_blocks, src_offset, buff_size);
173 if (src_private_blocks.empty()) { // simple shared malloc ... return.
174 XBT_VERB("Sender is shared. Let's ignore it.");
175 smpi_cleanup_comm_after_copy(comm, buff);
180 src_private_blocks.clear();
181 src_private_blocks.emplace_back(0, buff_size);
183 if (smpi_is_shared((char*)comm->dst_buff_, dst_private_blocks, &dst_offset)) {
184 dst_private_blocks = shift_and_frame_private_blocks(dst_private_blocks, dst_offset, buff_size);
185 if (dst_private_blocks.empty()) { // simple shared malloc ... return.
186 XBT_VERB("Receiver is shared. Let's ignore it.");
187 smpi_cleanup_comm_after_copy(comm, buff);
192 dst_private_blocks.clear();
193 dst_private_blocks.emplace_back(0, buff_size);
195 check_blocks(src_private_blocks, buff_size);
196 check_blocks(dst_private_blocks, buff_size);
197 auto private_blocks = merge_private_blocks(src_private_blocks, dst_private_blocks);
198 check_blocks(private_blocks, buff_size);
200 if (smpi_switch_data_segment(comm->src_actor_->get_iface(), buff)) {
201 XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
202 tmpbuff = xbt_malloc(buff_size);
203 memcpy_private(tmpbuff, buff, private_blocks);
206 if (smpi_switch_data_segment(comm->dst_actor_->get_iface(), comm->dst_buff_))
207 XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
209 XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff, comm->dst_buff_);
210 memcpy_private(comm->dst_buff_, tmpbuff, private_blocks);
212 smpi_cleanup_comm_after_copy(comm,buff);
217 void smpi_comm_null_copy_buffer_callback(simgrid::kernel::activity::CommImpl*, void*, size_t)
219 /* nothing done in this version */
223 return MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED;
226 static void smpi_init_papi()
229 // This map holds for each computation unit (such as "default" or "process1" etc.)
230 // the configuration as given by the user (counter data as a pair of (counter_name, counter_counter))
231 // and the (computed) event_set.
233 if (not smpi_cfg_papi_events_file().empty()) {
234 if (PAPI_library_init(PAPI_VER_CURRENT) != PAPI_VER_CURRENT)
235 XBT_ERROR("Could not initialize PAPI library; is it correctly installed and linked?"
236 " Expected version is %u", PAPI_VER_CURRENT);
238 using Tokenizer = boost::tokenizer<boost::char_separator<char>>;
239 boost::char_separator<char> separator_units(";");
240 std::string str = smpi_cfg_papi_events_file();
241 Tokenizer tokens(str, separator_units);
243 // Iterate over all the computational units. This could be processes, hosts, threads, ranks... You name it.
244 // I'm not exactly sure what we will support eventually, so I'll leave it at the general term "units".
245 for (auto const& unit_it : tokens) {
246 boost::char_separator<char> separator_events(":");
247 Tokenizer event_tokens(unit_it, separator_events);
249 int event_set = PAPI_NULL;
250 if (PAPI_create_eventset(&event_set) != PAPI_OK) {
251 // TODO: Should this let the whole simulation die?
252 XBT_CRITICAL("Could not create PAPI event set during init.");
255 // NOTE: We cannot use a map here, as we must obey the order of the counters
256 // This is important for PAPI: We need to map the values of counters back to the event_names (so, when PAPI_read()
258 papi_counter_t counters2values;
260 // Iterate over all counters that were specified for this specific unit.
261 // Note that we need to remove the name of the unit (that could also be the "default" value), which always comes
262 // first. Hence, we start at ++(events.begin())!
263 for (Tokenizer::iterator events_it = ++(event_tokens.begin()); events_it != event_tokens.end(); ++events_it) {
264 int event_code = PAPI_NULL;
265 auto* event_name = const_cast<char*>((*events_it).c_str());
266 if (PAPI_event_name_to_code(event_name, &event_code) != PAPI_OK) {
267 XBT_CRITICAL("Could not find PAPI event '%s'. Skipping.", event_name);
270 if (PAPI_add_event(event_set, event_code) != PAPI_OK) {
271 XBT_ERROR("Could not add PAPI event '%s'. Skipping.", event_name);
274 XBT_DEBUG("Successfully added PAPI event '%s' to the event set.", event_name);
276 counters2values.emplace_back(*events_it, 0LL);
279 std::string unit_name = *(event_tokens.begin());
280 papi_process_data config = {.counter_data = std::move(counters2values), .event_set = event_set};
282 units2papi_setup.insert(std::make_pair(unit_name, std::move(config)));
288 using smpi_entry_point_type = std::function<int(int argc, char* argv[])>;
289 using smpi_c_entry_point_type = int (*)(int argc, char** argv);
290 using smpi_fortran_entry_point_type = void (*)();
292 template <typename F>
293 static int smpi_run_entry_point(const F& entry_point, const std::string& executable_path, std::vector<std::string> args)
295 // copy C strings, we need them writable
296 auto* args4argv = new std::vector<char*>(args.size());
297 std::transform(begin(args), end(args), begin(*args4argv), [](const std::string& s) { return xbt_strdup(s.c_str()); });
299 // set argv[0] to executable_path
300 xbt_free((*args4argv)[0]);
301 (*args4argv)[0] = xbt_strdup(executable_path.c_str());
304 // take a copy of args4argv to keep reference of the allocated strings
305 const std::vector<char*> args2str(*args4argv);
307 int argc = args4argv->size();
308 args4argv->push_back(nullptr);
309 char** argv = args4argv->data();
312 for_rtl_init_ (&argc, argv);
317 _gfortran_set_args(argc, argv);
319 int res = entry_point(argc, argv);
324 for (char* s : args2str)
330 XBT_WARN("SMPI process did not return 0. Return value : %d", res);
331 if (smpi_exit_status == 0)
332 smpi_exit_status = res;
338 // TODO, remove the number of functions involved here
339 static smpi_entry_point_type smpi_resolve_function(void* handle)
341 auto* entry_point_fortran = reinterpret_cast<smpi_fortran_entry_point_type>(dlsym(handle, "user_main_"));
342 if (entry_point_fortran != nullptr) {
343 return [entry_point_fortran](int, char**) {
344 entry_point_fortran();
349 auto* entry_point = reinterpret_cast<smpi_c_entry_point_type>(dlsym(handle, "main"));
350 if (entry_point != nullptr) {
354 return smpi_entry_point_type();
357 static void smpi_copy_file(const std::string& src, const std::string& target, off_t fdin_size)
359 int fdin = open(src.c_str(), O_RDONLY);
360 xbt_assert(fdin >= 0, "Cannot read from %s. Please make sure that the file exists and is executable.", src.c_str());
361 xbt_assert(unlink(target.c_str()) == 0 || errno == ENOENT, "Failed to unlink file %s: %s", target.c_str(),
363 int fdout = open(target.c_str(), O_CREAT | O_RDWR | O_EXCL, S_IRWXU);
364 xbt_assert(fdout >= 0, "Cannot write into %s: %s", target.c_str(), strerror(errno));
366 XBT_DEBUG("Copy %" PRIdMAX " bytes into %s", static_cast<intmax_t>(fdin_size), target.c_str());
368 ssize_t sent_size = sendfile(fdout, fdin, nullptr, fdin_size);
369 if (sent_size == fdin_size) {
374 xbt_assert(sent_size == -1 && errno == ENOSYS,
375 "Error while copying %s: only %zd bytes copied instead of %" PRIdMAX " (errno: %d -- %s)", target.c_str(),
376 sent_size, static_cast<intmax_t>(fdin_size), errno, strerror(errno));
378 // If this point is reached, sendfile() actually is not available. Copy file by hand.
379 std::vector<unsigned char> buf(1024 * 1024 * 4);
380 while (ssize_t got = read(fdin, buf.data(), buf.size())) {
382 xbt_assert(errno == EINTR, "Cannot read from %s", src.c_str());
384 const unsigned char* p = buf.data();
386 while (ssize_t done = write(fdout, p, todo)) {
388 xbt_assert(errno == EINTR, "Cannot write into %s", target.c_str());
400 #if not defined(__APPLE__) && not defined(__HAIKU__)
401 static int visit_libs(struct dl_phdr_info* info, size_t, void* data)
403 auto* libname = static_cast<std::string*>(data);
404 std::string path = info->dlpi_name;
405 if (path.find(*libname) != std::string::npos) {
406 *libname = std::move(path);
413 static void smpi_init_privatization_dlopen(const std::string& executable)
415 // Prepare the copy of the binary (get its size)
416 struct stat fdin_stat;
417 stat(executable.c_str(), &fdin_stat);
418 off_t fdin_size = fdin_stat.st_size;
420 std::string libnames = simgrid::config::get_value<std::string>("smpi/privatize-libs");
421 if (not libnames.empty()) {
423 std::vector<std::string> privatize_libs;
424 boost::split(privatize_libs, libnames, boost::is_any_of(";"));
426 for (auto const& libname : privatize_libs) {
427 // load the library once to add it to the local libs, to get the absolute path
428 void* libhandle = dlopen(libname.c_str(), RTLD_LAZY);
429 xbt_assert(libhandle != nullptr, "Cannot dlopen %s - check your settings in smpi/privatize-libs",
431 // get library name from path
432 std::string fullpath = libname;
433 #if not defined(__APPLE__) && not defined(__HAIKU__)
434 xbt_assert(dl_iterate_phdr(visit_libs, &fullpath) != 0,
435 "Can't find a linked %s - check your settings in smpi/privatize-libs", fullpath.c_str());
436 XBT_DEBUG("Extra lib to privatize '%s' found", fullpath.c_str());
438 xbt_die("smpi/privatize-libs is not (yet) compatible with OSX nor with Haiku");
440 privatize_libs_paths.emplace_back(std::move(fullpath));
445 simgrid::s4u::Engine::get_instance()->register_default([executable, fdin_size](std::vector<std::string> args) {
446 return std::function<void()>([executable, fdin_size, args] {
447 static std::size_t rank = 0;
448 // Copy the dynamic library:
449 simgrid::xbt::Path path(executable);
450 std::string target_executable = simgrid::config::get_value<std::string>("smpi/tmpdir") + "/" +
451 path.get_base_name() + "_" + std::to_string(getpid()) + "_" + std::to_string(rank) + ".so";
453 smpi_copy_file(executable, target_executable, fdin_size);
454 // if smpi/privatize-libs is set, duplicate pointed lib and link each executable copy to a different one.
455 std::vector<std::string> target_libs;
456 for (auto const& libpath : privatize_libs_paths) {
457 // if we were given a full path, strip it
458 size_t index = libpath.find_last_of("/\\");
460 if (index != std::string::npos)
461 libname = libpath.substr(index + 1);
463 if (not libname.empty()) {
464 // load the library to add it to the local libs, to get the absolute path
465 struct stat fdin_stat2;
466 stat(libpath.c_str(), &fdin_stat2);
467 off_t fdin_size2 = fdin_stat2.st_size;
469 // Copy the dynamic library, the new name must be the same length as the old one
470 // just replace the name with 7 digits for the rank and the rest of the name.
471 auto pad = std::min<unsigned>(7, libname.length());
472 std::string target_libname = std::string(pad - std::to_string(rank).length(), '0') + std::to_string(rank) + libname.substr(pad);
473 std::string target_lib = simgrid::config::get_value<std::string>("smpi/tmpdir") + "/" + target_libname;
474 target_libs.push_back(target_lib);
475 XBT_DEBUG("copy lib %s to %s, with size %lld", libpath.c_str(), target_lib.c_str(), (long long)fdin_size2);
476 smpi_copy_file(libpath, target_lib, fdin_size2);
478 std::string sedcommand = "sed -i -e 's/" + libname + "/" + target_libname + "/g' " + target_executable;
479 int status = system(sedcommand.c_str());
480 xbt_assert(status == 0, "error while applying sed command %s \n", sedcommand.c_str());
485 // Load the copy and resolve the entry point:
486 void* handle = dlopen(target_executable.c_str(), RTLD_LAZY | RTLD_LOCAL | WANT_RTLD_DEEPBIND);
487 int saved_errno = errno;
488 if (not simgrid::config::get_value<bool>("smpi/keep-temps")) {
489 unlink(target_executable.c_str());
490 for (const std::string& target_lib : target_libs)
491 unlink(target_lib.c_str());
493 xbt_assert(handle != nullptr,
494 "dlopen failed: %s (errno: %d -- %s).\nError: Did you compile the program with a SMPI-specific "
495 "compiler (spmicc or friends)?",
496 dlerror(), saved_errno, strerror(saved_errno));
498 smpi_entry_point_type entry_point = smpi_resolve_function(handle);
499 xbt_assert(entry_point, "Could not resolve entry point. Does your program contain a main() function?");
500 smpi_run_entry_point(entry_point, executable, args);
505 static void smpi_init_privatization_no_dlopen(const std::string& executable)
507 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP)
508 smpi_prepare_global_memory_segment();
510 // Load the dynamic library and resolve the entry point:
511 void* handle = dlopen(executable.c_str(), RTLD_LAZY | RTLD_LOCAL);
512 xbt_assert(handle != nullptr, "dlopen failed for %s: %s (errno: %d -- %s)", executable.c_str(), dlerror(), errno,
514 smpi_entry_point_type entry_point = smpi_resolve_function(handle);
515 xbt_assert(entry_point, "main not found in %s", executable.c_str());
517 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP)
518 smpi_backup_global_memory_segment();
520 // Execute the same entry point for each simulated process:
521 simgrid::s4u::Engine::get_instance()->register_default([entry_point, executable](std::vector<std::string> args) {
522 return std::function<void()>([entry_point, executable, args] {
523 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) {
524 simgrid::smpi::ActorExt* ext = smpi_process();
525 /* Now using the segment index of this process */
526 ext->set_privatized_region(smpi_init_global_memory_segment_process());
527 /* Done at the process's creation */
528 smpi_switch_data_segment(simgrid::s4u::Actor::self());
530 smpi_run_entry_point(entry_point, executable, args);
535 int smpi_main(const char* executable, int argc, char* argv[])
537 if (getenv("SMPI_PRETEND_CC") != nullptr) {
538 /* Hack to ensure that smpicc can pretend to be a simple compiler. Particularly handy to pass it to the
539 * configuration tools */
543 smpi_init_options_internal(true);
544 simgrid::instr::init();
545 SIMIX_global_init(&argc, argv);
547 auto engine = simgrid::s4u::Engine::get_instance();
549 sg_storage_file_system_init();
550 // parse the platform file: get the host list
551 engine->load_platform(argv[1]);
552 simgrid::kernel::activity::CommImpl::set_copy_data_callback(smpi_comm_copy_buffer_callback);
554 if (smpi_cfg_privatization() == SmpiPrivStrategies::DLOPEN)
555 smpi_init_privatization_dlopen(executable);
557 smpi_init_privatization_no_dlopen(executable);
559 simgrid::smpi::colls::set_collectives();
560 simgrid::smpi::colls::smpi_coll_cleanup_callback = nullptr;
564 int rank_counts = smpi_deployment_smpirun(engine, smpi_hostfile.get(), smpi_np.get(), smpi_replay.get(),
565 smpi_map.get(), argc - 2, argv + 2);
567 SMPI_app_instance_register(smpi_default_instance_name.c_str(), nullptr, rank_counts);
568 MPI_COMM_WORLD = *smpi_deployment_comm_world(smpi_default_instance_name);
570 /* Clean IO before the run */
574 if (MC_is_active()) {
577 simgrid::kernel::EngineImpl::get_instance()->run();
579 xbt_os_walltimer_stop(global_timer);
580 simgrid::smpi::utils::print_time_analysis(xbt_os_timer_elapsed(global_timer));
584 return smpi_exit_status;
587 // Called either directly from the user code, or from the code called by smpirun
589 smpi_init_options_internal(false);
590 simgrid::s4u::Actor::on_creation.connect([](simgrid::s4u::Actor& actor) {
591 if (not actor.is_daemon())
592 actor.extension_set<simgrid::smpi::ActorExt>(new simgrid::smpi::ActorExt(&actor));
594 simgrid::s4u::Host::on_creation.connect(
595 [](simgrid::s4u::Host& host) { host.extension_set(new simgrid::smpi::Host(&host)); });
596 for (auto const& host : simgrid::s4u::Engine::get_instance()->get_all_hosts())
597 host->extension_set(new simgrid::smpi::Host(host));
599 if (not MC_is_active()) {
600 global_timer = xbt_os_timer_new();
601 xbt_os_walltimer_start(global_timer);
604 smpi_check_options();
609 smpi_bench_destroy();
610 smpi_shared_destroy();
611 smpi_deployment_cleanup_instances();
613 if (simgrid::smpi::colls::smpi_coll_cleanup_callback != nullptr)
614 simgrid::smpi::colls::smpi_coll_cleanup_callback();
616 MPI_COMM_WORLD = MPI_COMM_NULL;
618 if (not MC_is_active()) {
619 xbt_os_timer_free(global_timer);
622 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP)
623 smpi_destroy_global_memory_segments();
625 simgrid::smpi::utils::print_memory_analysis();
628 void smpi_mpi_init() {
629 smpi_init_fortran_types();
630 if(smpi_init_sleep > 0)
631 simgrid::s4u::this_actor::sleep_for(smpi_init_sleep);
634 void SMPI_thread_create() {
635 TRACE_smpi_init(simgrid::s4u::this_actor::get_pid(), __func__);
636 smpi_process()->mark_as_initialized();