1 /* Copyright (c) 2007-2022. 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/Exception.hpp"
8 #include "simgrid/plugins/file_system.h"
9 #include "simgrid/s4u/Engine.hpp"
10 #include "smpi_coll.hpp"
11 #include "smpi_config.hpp"
12 #include "smpi_f2c.hpp"
13 #include "smpi_host.hpp"
14 #include "src/kernel/EngineImpl.hpp"
15 #include "src/kernel/activity/CommImpl.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");
81 static simgrid::config::Flag<std::string>
82 smpi_hostfile("smpi/hostfile",
83 "Classical MPI hostfile containing list of machines to dispatch "
84 "the processes, one per line",
87 static simgrid::config::Flag<std::string> smpi_replay("smpi/replay",
88 "Replay a trace instead of executing the application", "");
90 static simgrid::config::Flag<int> smpi_np("smpi/np", "Number of processes to be created", 0);
92 static simgrid::config::Flag<int> smpi_map("smpi/map", "Display the mapping between nodes and processes", 0);
94 std::function<void(simgrid::kernel::activity::CommImpl*, void*, size_t)> smpi_comm_copy_data_callback =
95 &smpi_comm_copy_buffer_callback;
97 simgrid::smpi::ActorExt* smpi_process()
99 simgrid::s4u::ActorPtr me = simgrid::s4u::Actor::self();
101 if (me == nullptr) // This happens sometimes (eg, when linking against NS3 because it pulls openMPI...)
104 return me->extension<simgrid::smpi::ActorExt>();
107 simgrid::smpi::ActorExt* smpi_process_remote(simgrid::s4u::ActorPtr actor)
109 if (actor.get() == nullptr)
111 return actor->extension<simgrid::smpi::ActorExt>();
114 MPI_Comm smpi_process_comm_self(){
115 return smpi_process()->comm_self();
118 MPI_Info smpi_process_info_env(){
119 return smpi_process()->info_env();
122 void * smpi_process_get_user_data(){
123 return simgrid::s4u::Actor::self()->get_data<void>();
126 void smpi_process_set_user_data(void *data){
127 simgrid::s4u::Actor::self()->set_data(data);
130 void smpi_comm_set_copy_data_callback(void (*callback) (smx_activity_t, void*, size_t))
132 smpi_comm_copy_data_callback = callback;
135 static void memcpy_private(void* dest, const void* src, const std::vector<std::pair<size_t, size_t>>& private_blocks)
137 for (auto const& [block_begin, block_end] : private_blocks)
138 memcpy((uint8_t*)dest + block_begin, (uint8_t*)src + block_begin, block_end - block_begin);
141 static void check_blocks(const std::vector<std::pair<size_t, size_t>>& private_blocks, size_t buff_size)
143 for (auto const& [block_begin, block_end] : private_blocks)
144 xbt_assert(block_begin <= block_end && block_end <= buff_size, "Oops, bug in shared malloc.");
147 static void smpi_cleanup_comm_after_copy(simgrid::kernel::activity::CommImpl* comm, void* buff){
148 if (comm->is_detached()) {
149 // if this is a detached send, the source buffer was duplicated by SMPI
150 // sender to make the original buffer available to the application ASAP
152 //It seems that the request is used after the call there this should be free somewhere else but where???
153 //xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free
154 comm->src_buff_ = nullptr;
158 void smpi_comm_copy_buffer_callback(simgrid::kernel::activity::CommImpl* comm, void* buff, size_t buff_size)
160 size_t src_offset = 0;
161 size_t dst_offset = 0;
162 std::vector<std::pair<size_t, size_t>> src_private_blocks;
163 std::vector<std::pair<size_t, size_t>> dst_private_blocks;
164 XBT_DEBUG("Copy the data over");
165 if(smpi_is_shared(buff, src_private_blocks, &src_offset)) {
166 src_private_blocks = shift_and_frame_private_blocks(src_private_blocks, src_offset, buff_size);
167 if (src_private_blocks.empty()) { // simple shared malloc ... return.
168 XBT_VERB("Sender is shared. Let's ignore it.");
169 smpi_cleanup_comm_after_copy(comm, buff);
174 src_private_blocks.clear();
175 src_private_blocks.emplace_back(0, buff_size);
177 if (smpi_is_shared((char*)comm->dst_buff_, dst_private_blocks, &dst_offset)) {
178 dst_private_blocks = shift_and_frame_private_blocks(dst_private_blocks, dst_offset, buff_size);
179 if (dst_private_blocks.empty()) { // simple shared malloc ... return.
180 XBT_VERB("Receiver is shared. Let's ignore it.");
181 smpi_cleanup_comm_after_copy(comm, buff);
186 dst_private_blocks.clear();
187 dst_private_blocks.emplace_back(0, buff_size);
189 check_blocks(src_private_blocks, buff_size);
190 check_blocks(dst_private_blocks, buff_size);
191 auto private_blocks = merge_private_blocks(src_private_blocks, dst_private_blocks);
192 check_blocks(private_blocks, buff_size);
194 if (smpi_switch_data_segment(comm->src_actor_->get_iface(), buff)) {
195 XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
196 tmpbuff = xbt_malloc(buff_size);
197 memcpy_private(tmpbuff, buff, private_blocks);
200 if (smpi_switch_data_segment(comm->dst_actor_->get_iface(), comm->dst_buff_))
201 XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
203 XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff, comm->dst_buff_);
204 memcpy_private(comm->dst_buff_, tmpbuff, private_blocks);
206 smpi_cleanup_comm_after_copy(comm,buff);
211 void smpi_comm_null_copy_buffer_callback(simgrid::kernel::activity::CommImpl*, void*, size_t)
213 /* nothing done in this version */
217 return MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED;
220 static void smpi_init_papi()
223 // This map holds for each computation unit (such as "default" or "process1" etc.)
224 // the configuration as given by the user (counter data as a pair of (counter_name, counter_counter))
225 // and the (computed) event_set.
227 if (smpi_cfg_papi_events_file().empty())
230 if (PAPI_library_init(PAPI_VER_CURRENT) != PAPI_VER_CURRENT) {
231 XBT_ERROR("Could not initialize PAPI library; is it correctly installed and linked? Expected version is %u",
236 using Tokenizer = boost::tokenizer<boost::char_separator<char>>;
237 boost::char_separator<char> separator_units(";");
238 std::string str = smpi_cfg_papi_events_file();
239 Tokenizer tokens(str, separator_units);
241 // Iterate over all the computational units. This could be processes, hosts, threads, ranks... You name it.
242 // I'm not exactly sure what we will support eventually, so I'll leave it at the general term "units".
243 for (auto const& unit_it : tokens) {
244 boost::char_separator<char> separator_events(":");
245 Tokenizer event_tokens(unit_it, separator_events);
247 int event_set = PAPI_NULL;
248 if (PAPI_create_eventset(&event_set) != PAPI_OK) {
249 // TODO: Should this let the whole simulation die?
250 XBT_CRITICAL("Could not create PAPI event set during init.");
254 // NOTE: We cannot use a map here, as we must obey the order of the counters
255 // This is important for PAPI: We need to map the values of counters back to the event_names (so, when PAPI_read()
257 papi_counter_t counters2values;
259 // Iterate over all counters that were specified for this specific unit.
260 // Note that we need to remove the name of the unit (that could also be the "default" value), which always comes
261 // first. Hence, we start at ++(events.begin())!
262 for (Tokenizer::iterator events_it = ++(event_tokens.begin()); events_it != event_tokens.end(); ++events_it) {
263 int event_code = PAPI_NULL;
264 auto* event_name = const_cast<char*>((*events_it).c_str());
265 if (PAPI_event_name_to_code(event_name, &event_code) != PAPI_OK) {
266 XBT_CRITICAL("Could not find PAPI event '%s'. Skipping.", event_name);
269 if (PAPI_add_event(event_set, event_code) != PAPI_OK) {
270 XBT_ERROR("Could not add PAPI event '%s'. Skipping.", event_name);
273 XBT_DEBUG("Successfully added PAPI event '%s' to the event set.", event_name);
275 counters2values.emplace_back(*events_it, 0LL);
278 std::string unit_name = *(event_tokens.begin());
279 papi_process_data config = {.counter_data = std::move(counters2values), .event_set = event_set};
281 units2papi_setup.try_emplace(unit_name, std::move(config));
286 using smpi_entry_point_type = std::function<int(int argc, char* argv[])>;
287 using smpi_c_entry_point_type = int (*)(int argc, char** argv);
288 using smpi_fortran_entry_point_type = void (*)();
290 template <typename F>
291 static int smpi_run_entry_point(const F& entry_point, const std::string& executable_path,
292 const std::vector<std::string>& args)
294 // copy C strings, we need them writable
295 std::vector<char*> args4argv(args.size());
296 std::transform(begin(args) + 1, end(args), begin(args4argv) + 1,
297 [](const std::string& s) { return xbt_strdup(s.c_str()); });
299 // set argv[0] to executable_path
300 args4argv[0] = xbt_strdup(executable_path.c_str());
301 // add the final NULL
302 args4argv.push_back(nullptr);
304 // take a copy of args4argv to keep reference of the allocated strings
305 const std::vector<char*> args2str(args4argv);
308 int argc = static_cast<int>(args4argv.size() - 1);
309 char** argv = args4argv.data();
310 int res = entry_point(argc, argv);
312 XBT_WARN("SMPI process did not return 0. Return value : %d", res);
313 if (smpi_exit_status == 0)
314 smpi_exit_status = res;
316 } catch (simgrid::ForcefulKillException const& e) {
317 XBT_DEBUG("Caught a ForcefulKillException: %s", e.what());
320 for (char* s : args2str)
326 static smpi_entry_point_type smpi_resolve_function(void* handle)
328 if (auto* entry_point_fortran = reinterpret_cast<smpi_fortran_entry_point_type>(dlsym(handle, "user_main_"))) {
329 return [entry_point_fortran](int, char**) {
330 entry_point_fortran();
335 if (auto* entry_point = reinterpret_cast<smpi_c_entry_point_type>(dlsym(handle, "main"))) {
339 return smpi_entry_point_type();
342 static void smpi_copy_file(const std::string& src, const std::string& target, off_t fdin_size)
344 int fdin = open(src.c_str(), O_RDONLY);
345 xbt_assert(fdin >= 0, "Cannot read from %s. Please make sure that the file exists and is executable.", src.c_str());
346 xbt_assert(unlink(target.c_str()) == 0 || errno == ENOENT, "Failed to unlink file %s: %s", target.c_str(),
348 int fdout = open(target.c_str(), O_CREAT | O_RDWR | O_EXCL, S_IRWXU);
349 xbt_assert(fdout >= 0, "Cannot write into %s: %s", target.c_str(), strerror(errno));
351 XBT_DEBUG("Copy %" PRIdMAX " bytes into %s", static_cast<intmax_t>(fdin_size), target.c_str());
353 ssize_t sent_size = sendfile(fdout, fdin, nullptr, fdin_size);
354 if (sent_size == fdin_size) {
359 xbt_assert(sent_size == -1 && errno == ENOSYS,
360 "Error while copying %s: only %zd bytes copied instead of %" PRIdMAX " (errno: %d -- %s)", target.c_str(),
361 sent_size, static_cast<intmax_t>(fdin_size), errno, strerror(errno));
363 // If this point is reached, sendfile() actually is not available. Copy file by hand.
364 std::vector<unsigned char> buf(1024 * 1024 * 4);
365 while (ssize_t got = read(fdin, buf.data(), buf.size())) {
367 xbt_assert(errno == EINTR, "Cannot read from %s", src.c_str());
370 const unsigned char* p = buf.data();
372 while (ssize_t done = write(fdout, p, todo)) {
374 xbt_assert(errno == EINTR, "Cannot write into %s", target.c_str());
385 #if not defined(__APPLE__) && not defined(__HAIKU__)
386 static int visit_libs(struct dl_phdr_info* info, size_t, void* data)
388 auto* libname = static_cast<std::string*>(data);
389 std::string path = info->dlpi_name;
390 if (path.find(*libname) == std::string::npos)
393 *libname = std::move(path);
398 static void smpi_init_privatization_dlopen(const std::string& executable)
400 // Prepare the copy of the binary (get its size)
401 struct stat fdin_stat;
402 stat(executable.c_str(), &fdin_stat);
403 off_t fdin_size = fdin_stat.st_size;
405 std::string libnames = simgrid::config::get_value<std::string>("smpi/privatize-libs");
406 if (not libnames.empty()) {
408 std::vector<std::string> privatize_libs;
409 boost::split(privatize_libs, libnames, boost::is_any_of(";"));
411 for (auto const& libname : privatize_libs) {
412 // load the library once to add it to the local libs, to get the absolute path
413 void* libhandle = dlopen(libname.c_str(), RTLD_LAZY);
414 xbt_assert(libhandle != nullptr, "Cannot dlopen %s - check your settings in smpi/privatize-libs",
416 // get library name from path
417 std::string fullpath = libname;
418 #if not defined(__APPLE__) && not defined(__HAIKU__)
419 xbt_assert(dl_iterate_phdr(visit_libs, &fullpath) != 0,
420 "Can't find a linked %s - check your settings in smpi/privatize-libs", fullpath.c_str());
421 XBT_DEBUG("Extra lib to privatize '%s' found", fullpath.c_str());
423 xbt_die("smpi/privatize-libs is not (yet) compatible with OSX nor with Haiku");
425 privatize_libs_paths.emplace_back(std::move(fullpath));
430 simgrid::s4u::Engine::get_instance()->register_default([executable, fdin_size](std::vector<std::string> args) {
431 return simgrid::kernel::actor::ActorCode([executable, fdin_size, args = std::move(args)] {
432 static std::size_t rank = 0;
433 // Copy the dynamic library:
434 simgrid::xbt::Path path(executable);
435 std::string target_executable = simgrid::config::get_value<std::string>("smpi/tmpdir") + "/" +
436 path.get_base_name() + "_" + std::to_string(getpid()) + "_" + std::to_string(rank) + ".so";
438 smpi_copy_file(executable, target_executable, fdin_size);
439 // if smpi/privatize-libs is set, duplicate pointed lib and link each executable copy to a different one.
440 std::vector<std::string> target_libs;
441 for (auto const& libpath : privatize_libs_paths) {
442 // if we were given a full path, strip it
443 size_t index = libpath.find_last_of("/\\");
445 if (index != std::string::npos)
446 libname = libpath.substr(index + 1);
448 if (not libname.empty()) {
449 // load the library to add it to the local libs, to get the absolute path
450 struct stat fdin_stat2;
451 stat(libpath.c_str(), &fdin_stat2);
452 off_t fdin_size2 = fdin_stat2.st_size;
454 // Copy the dynamic library, the new name must be the same length as the old one
455 // just replace the name with 7 digits for the rank and the rest of the name.
456 auto pad = std::min<size_t>(7, libname.length());
457 std::string target_libname = std::string(pad - std::to_string(rank).length(), '0') + std::to_string(rank) + libname.substr(pad);
458 std::string target_lib = simgrid::config::get_value<std::string>("smpi/tmpdir") + "/" + target_libname;
459 target_libs.push_back(target_lib);
460 XBT_DEBUG("copy lib %s to %s, with size %lld", libpath.c_str(), target_lib.c_str(), (long long)fdin_size2);
461 smpi_copy_file(libpath, target_lib, fdin_size2);
463 std::string sedcommand = "sed -i -e 's/" + libname + "/" + target_libname + "/g' " + target_executable;
464 int status = system(sedcommand.c_str());
465 xbt_assert(status == 0, "error while applying sed command %s \n", sedcommand.c_str());
470 // Load the copy and resolve the entry point:
471 void* handle = dlopen(target_executable.c_str(), RTLD_LAZY | RTLD_LOCAL | WANT_RTLD_DEEPBIND);
472 int saved_errno = errno;
473 if (not simgrid::config::get_value<bool>("smpi/keep-temps")) {
474 unlink(target_executable.c_str());
475 for (const std::string& target_lib : target_libs)
476 unlink(target_lib.c_str());
478 xbt_assert(handle != nullptr,
479 "dlopen failed: %s (errno: %d -- %s).\nError: Did you compile the program with a SMPI-specific "
480 "compiler (spmicc or friends)?",
481 dlerror(), saved_errno, strerror(saved_errno));
483 smpi_entry_point_type entry_point = smpi_resolve_function(handle);
484 xbt_assert(entry_point, "Could not resolve entry point. Does your program contain a main() function?");
485 smpi_run_entry_point(entry_point, executable, args);
490 static void smpi_init_privatization_no_dlopen(const std::string& executable)
492 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP)
493 smpi_prepare_global_memory_segment();
495 // Load the dynamic library and resolve the entry point:
496 void* handle = dlopen(executable.c_str(), RTLD_LAZY | RTLD_LOCAL);
497 xbt_assert(handle != nullptr, "dlopen failed for %s: %s (errno: %d -- %s)", executable.c_str(), dlerror(), errno,
499 smpi_entry_point_type entry_point = smpi_resolve_function(handle);
500 xbt_assert(entry_point, "main not found in %s", executable.c_str());
502 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP)
503 smpi_backup_global_memory_segment();
505 // Execute the same entry point for each simulated process:
506 simgrid::s4u::Engine::get_instance()->register_default([entry_point, executable](std::vector<std::string> args) {
507 return simgrid::kernel::actor::ActorCode([entry_point, executable, args = std::move(args)] {
508 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) {
509 simgrid::smpi::ActorExt* ext = smpi_process();
510 /* Now using the segment index of this process */
511 ext->set_privatized_region(smpi_init_global_memory_segment_process());
512 /* Done at the process's creation */
513 smpi_switch_data_segment(simgrid::s4u::Actor::self());
515 smpi_run_entry_point(entry_point, executable, args);
520 int smpi_main(const char* executable, int argc, char* argv[])
522 if (getenv("SMPI_PRETEND_CC") != nullptr) {
523 /* Hack to ensure that smpicc can pretend to be a simple compiler. Particularly handy to pass it to the
524 * configuration tools */
528 smpi_init_options_internal(true);
529 simgrid::s4u::Engine engine(&argc, argv);
531 sg_storage_file_system_init();
532 // parse the platform file: get the host list
533 engine.load_platform(argv[1]);
534 engine.set_default_comm_data_copy_callback(smpi_comm_copy_buffer_callback);
536 if (smpi_cfg_privatization() == SmpiPrivStrategies::DLOPEN)
537 smpi_init_privatization_dlopen(executable);
539 smpi_init_privatization_no_dlopen(executable);
541 simgrid::smpi::colls::set_collectives();
542 simgrid::smpi::colls::smpi_coll_cleanup_callback = nullptr;
544 std::vector<char*> args4argv(argv + 1, argv + argc + 1); // last element is NULL
545 args4argv[0] = xbt_strdup(executable);
546 int real_argc = argc - 1;
547 char** real_argv = args4argv.data();
549 // Setup argc/argv for the Fortran run-time environment
551 for_rtl_init_(&real_argc, real_argv);
553 __io_set_argc(real_argc);
554 __io_set_argv(real_argv);
556 _gfortran_set_args(real_argc, real_argv);
561 const std::vector<const char*> args(real_argv + 1, real_argv + real_argc);
563 smpi_deployment_smpirun(&engine, smpi_hostfile.get(), smpi_np.get(), smpi_replay.get(), smpi_map.get(), args);
565 SMPI_app_instance_register(smpi_default_instance_name.c_str(), nullptr, rank_counts);
566 MPI_COMM_WORLD = *smpi_deployment_comm_world(smpi_default_instance_name);
568 /* Clean IO before the run */
572 engine.get_impl()->run(-1);
574 xbt_os_walltimer_stop(global_timer);
575 simgrid::smpi::utils::print_time_analysis(xbt_os_timer_elapsed(global_timer));
582 xbt_free(args4argv[0]);
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_cb([](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_cb(
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();
612 smpi_cleanup_op_cost_callback();
614 if (simgrid::smpi::colls::smpi_coll_cleanup_callback != nullptr)
615 simgrid::smpi::colls::smpi_coll_cleanup_callback();
617 MPI_COMM_WORLD = MPI_COMM_NULL;
619 if (not MC_is_active()) {
620 xbt_os_timer_free(global_timer);
623 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP)
624 smpi_destroy_global_memory_segments();
626 simgrid::smpi::utils::print_memory_analysis();
629 void smpi_mpi_init() {
630 smpi_init_fortran_types();
631 if(_smpi_init_sleep > 0)
632 simgrid::s4u::this_actor::sleep_for(_smpi_init_sleep);
633 if (not MC_is_active()) {
634 smpi_deployment_startup_barrier(smpi_process()->get_instance_id());
638 void SMPI_thread_create() {
639 TRACE_smpi_init(simgrid::s4u::this_actor::get_pid(), __func__);
640 smpi_process()->mark_as_initialized();
643 void smpi_exit(int res){
645 XBT_WARN("SMPI process did not return 0. Return value : %d", res);
646 smpi_exit_status = res;
648 simgrid::s4u::this_actor::exit();