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/smpi/include/smpi_actor.hpp"
16 #include "xbt/config.hpp"
17 #include "xbt/file.hpp"
21 #include <boost/algorithm/string.hpp> /* split */
22 #include <boost/tokenizer.hpp>
25 #include <cstdint> /* intmax_t */
26 #include <cstring> /* strerror */
33 #include <sys/sendfile.h>
40 #if not defined(__APPLE__) && not defined(__HAIKU__)
44 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_kernel, smpi, "Logging specific to SMPI (kernel)");
47 extern "C" void for_rtl_init_ (int *, char **);
48 extern "C" void for_rtl_finish_ ();
50 extern "C" void __io_set_argc(int);
51 extern "C" void __io_set_argv(char **);
53 extern "C" void _gfortran_set_args(int, char **);
56 /* RTLD_DEEPBIND is a bad idea of GNU ld that obviously does not exist on other platforms
57 * See https://www.akkadia.org/drepper/dsohowto.pdf
58 * and https://lists.freebsd.org/pipermail/freebsd-current/2016-March/060284.html
60 #if !RTLD_DEEPBIND || HAVE_SANITIZER_ADDRESS || HAVE_SANITIZER_THREAD
61 #define WANT_RTLD_DEEPBIND 0
63 #define WANT_RTLD_DEEPBIND RTLD_DEEPBIND
67 std::map</* computation unit name */ std::string, papi_process_data, std::less<>> units2papi_setup;
70 std::unordered_map<std::string, double> location2speedup;
72 static int smpi_exit_status = 0;
73 xbt_os_timer_t global_timer;
74 static std::vector<std::string> privatize_libs_paths;
76 // No instance gets manually created; check also the smpirun.in script as
77 // this default name is used there as well (when the <actor> tag is generated).
78 static const std::string smpi_default_instance_name("smpirun");
80 static simgrid::config::Flag<std::string>
81 smpi_hostfile("smpi/hostfile",
82 "Classical MPI hostfile containing list of machines to dispatch "
83 "the processes, one per line",
86 static simgrid::config::Flag<std::string> smpi_replay("smpi/replay",
87 "Replay a trace instead of executing the application", "");
89 static simgrid::config::Flag<int> smpi_np("smpi/np", "Number of processes to be created", 0);
91 static simgrid::config::Flag<int> smpi_map("smpi/map", "Display the mapping between nodes and processes", 0);
93 void (*smpi_comm_copy_data_callback)(simgrid::kernel::activity::CommImpl*, void*,
94 size_t) = &smpi_comm_copy_buffer_callback;
96 simgrid::smpi::ActorExt* smpi_process()
98 simgrid::s4u::ActorPtr me = simgrid::s4u::Actor::self();
100 if (me == nullptr) // This happens sometimes (eg, when linking against NS3 because it pulls openMPI...)
103 return me->extension<simgrid::smpi::ActorExt>();
106 simgrid::smpi::ActorExt* smpi_process_remote(simgrid::s4u::ActorPtr actor)
108 if (actor.get() == nullptr)
110 return actor->extension<simgrid::smpi::ActorExt>();
113 MPI_Comm smpi_process_comm_self(){
114 return smpi_process()->comm_self();
117 MPI_Info smpi_process_info_env(){
118 return smpi_process()->info_env();
121 void * smpi_process_get_user_data(){
122 return simgrid::s4u::Actor::self()->get_data();
125 void smpi_process_set_user_data(void *data){
126 simgrid::s4u::Actor::self()->set_data(data);
129 void smpi_comm_set_copy_data_callback(void (*callback) (smx_activity_t, void*, size_t))
131 static void (*saved_callback)(smx_activity_t, void*, size_t);
132 saved_callback = callback;
133 smpi_comm_copy_data_callback = [](simgrid::kernel::activity::CommImpl* comm, void* buff, size_t size) {
134 saved_callback(comm, buff, size);
138 static void memcpy_private(void* dest, const void* src, const std::vector<std::pair<size_t, size_t>>& private_blocks)
140 for (auto const& block : private_blocks)
141 memcpy((uint8_t*)dest+block.first, (uint8_t*)src+block.first, block.second-block.first);
144 static void check_blocks(const std::vector<std::pair<size_t, size_t>>& private_blocks, size_t buff_size)
146 for (auto const& block : private_blocks)
147 xbt_assert(block.first <= block.second && block.second <= buff_size, "Oops, bug in shared malloc.");
150 static void smpi_cleanup_comm_after_copy(simgrid::kernel::activity::CommImpl* comm, void* buff){
151 if (comm->detached()) {
152 // if this is a detached send, the source buffer was duplicated by SMPI
153 // sender to make the original buffer available to the application ASAP
155 //It seems that the request is used after the call there this should be free somewhere else but where???
156 //xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free
157 comm->src_buff_ = nullptr;
161 void smpi_comm_copy_buffer_callback(simgrid::kernel::activity::CommImpl* comm, void* buff, size_t buff_size)
163 size_t src_offset = 0;
164 size_t dst_offset = 0;
165 std::vector<std::pair<size_t, size_t>> src_private_blocks;
166 std::vector<std::pair<size_t, size_t>> dst_private_blocks;
167 XBT_DEBUG("Copy the data over");
168 if(smpi_is_shared(buff, src_private_blocks, &src_offset)) {
169 src_private_blocks = shift_and_frame_private_blocks(src_private_blocks, src_offset, buff_size);
170 if (src_private_blocks.empty()) { // simple shared malloc ... return.
171 XBT_VERB("Sender is shared. Let's ignore it.");
172 smpi_cleanup_comm_after_copy(comm, buff);
177 src_private_blocks.clear();
178 src_private_blocks.emplace_back(0, buff_size);
180 if (smpi_is_shared((char*)comm->dst_buff_, dst_private_blocks, &dst_offset)) {
181 dst_private_blocks = shift_and_frame_private_blocks(dst_private_blocks, dst_offset, buff_size);
182 if (dst_private_blocks.empty()) { // simple shared malloc ... return.
183 XBT_VERB("Receiver is shared. Let's ignore it.");
184 smpi_cleanup_comm_after_copy(comm, buff);
189 dst_private_blocks.clear();
190 dst_private_blocks.emplace_back(0, buff_size);
192 check_blocks(src_private_blocks, buff_size);
193 check_blocks(dst_private_blocks, buff_size);
194 auto private_blocks = merge_private_blocks(src_private_blocks, dst_private_blocks);
195 check_blocks(private_blocks, buff_size);
197 if (smpi_switch_data_segment(comm->src_actor_->get_iface(), buff)) {
198 XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
199 tmpbuff = xbt_malloc(buff_size);
200 memcpy_private(tmpbuff, buff, private_blocks);
203 if (smpi_switch_data_segment(comm->dst_actor_->get_iface(), comm->dst_buff_))
204 XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
206 XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff, comm->dst_buff_);
207 memcpy_private(comm->dst_buff_, tmpbuff, private_blocks);
209 smpi_cleanup_comm_after_copy(comm,buff);
214 void smpi_comm_null_copy_buffer_callback(simgrid::kernel::activity::CommImpl*, void*, size_t)
216 /* nothing done in this version */
220 return MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED;
223 static void smpi_init_papi()
226 // This map holds for each computation unit (such as "default" or "process1" etc.)
227 // the configuration as given by the user (counter data as a pair of (counter_name, counter_counter))
228 // and the (computed) event_set.
230 if (smpi_cfg_papi_events_file().empty())
233 if (PAPI_library_init(PAPI_VER_CURRENT) != PAPI_VER_CURRENT) {
234 XBT_ERROR("Could not initialize PAPI library; is it correctly installed and linked? Expected version is %u",
239 using Tokenizer = boost::tokenizer<boost::char_separator<char>>;
240 boost::char_separator<char> separator_units(";");
241 std::string str = smpi_cfg_papi_events_file();
242 Tokenizer tokens(str, separator_units);
244 // Iterate over all the computational units. This could be processes, hosts, threads, ranks... You name it.
245 // I'm not exactly sure what we will support eventually, so I'll leave it at the general term "units".
246 for (auto const& unit_it : tokens) {
247 boost::char_separator<char> separator_events(":");
248 Tokenizer event_tokens(unit_it, separator_events);
250 int event_set = PAPI_NULL;
251 if (PAPI_create_eventset(&event_set) != PAPI_OK) {
252 // TODO: Should this let the whole simulation die?
253 XBT_CRITICAL("Could not create PAPI event set during init.");
257 // NOTE: We cannot use a map here, as we must obey the order of the counters
258 // This is important for PAPI: We need to map the values of counters back to the event_names (so, when PAPI_read()
260 papi_counter_t counters2values;
262 // Iterate over all counters that were specified for this specific unit.
263 // Note that we need to remove the name of the unit (that could also be the "default" value), which always comes
264 // first. Hence, we start at ++(events.begin())!
265 for (Tokenizer::iterator events_it = ++(event_tokens.begin()); events_it != event_tokens.end(); ++events_it) {
266 int event_code = PAPI_NULL;
267 auto* event_name = const_cast<char*>((*events_it).c_str());
268 if (PAPI_event_name_to_code(event_name, &event_code) != PAPI_OK) {
269 XBT_CRITICAL("Could not find PAPI event '%s'. Skipping.", event_name);
272 if (PAPI_add_event(event_set, event_code) != PAPI_OK) {
273 XBT_ERROR("Could not add PAPI event '%s'. Skipping.", event_name);
276 XBT_DEBUG("Successfully added PAPI event '%s' to the event set.", event_name);
278 counters2values.emplace_back(*events_it, 0LL);
281 std::string unit_name = *(event_tokens.begin());
282 papi_process_data config = {.counter_data = std::move(counters2values), .event_set = event_set};
284 units2papi_setup.insert(std::make_pair(unit_name, std::move(config)));
289 using smpi_entry_point_type = std::function<int(int argc, char* argv[])>;
290 using smpi_c_entry_point_type = int (*)(int argc, char** argv);
291 using smpi_fortran_entry_point_type = void (*)();
293 template <typename F>
294 static int smpi_run_entry_point(const F& entry_point, const std::string& executable_path,
295 const std::vector<std::string>& args)
297 // copy C strings, we need them writable
298 auto* args4argv = new std::vector<char*>(args.size());
299 std::transform(begin(args), end(args), begin(*args4argv), [](const std::string& s) { return xbt_strdup(s.c_str()); });
301 // set argv[0] to executable_path
302 xbt_free((*args4argv)[0]);
303 (*args4argv)[0] = xbt_strdup(executable_path.c_str());
306 // take a copy of args4argv to keep reference of the allocated strings
307 const std::vector<char*> args2str(*args4argv);
309 int argc = static_cast<int>(args4argv->size());
310 args4argv->push_back(nullptr);
311 char** argv = args4argv->data();
314 for_rtl_init_ (&argc, argv);
319 _gfortran_set_args(argc, argv);
321 int res = entry_point(argc, argv);
326 for (char* s : args2str)
332 XBT_WARN("SMPI process did not return 0. Return value : %d", res);
333 if (smpi_exit_status == 0)
334 smpi_exit_status = res;
340 // TODO, remove the number of functions involved here
341 static smpi_entry_point_type smpi_resolve_function(void* handle)
343 auto* entry_point_fortran = reinterpret_cast<smpi_fortran_entry_point_type>(dlsym(handle, "user_main_"));
344 if (entry_point_fortran != nullptr) {
345 return [entry_point_fortran](int, char**) {
346 entry_point_fortran();
351 auto* entry_point = reinterpret_cast<smpi_c_entry_point_type>(dlsym(handle, "main"));
352 if (entry_point != nullptr) {
356 return smpi_entry_point_type();
359 static void smpi_copy_file(const std::string& src, const std::string& target, off_t fdin_size)
361 int fdin = open(src.c_str(), O_RDONLY);
362 xbt_assert(fdin >= 0, "Cannot read from %s. Please make sure that the file exists and is executable.", src.c_str());
363 xbt_assert(unlink(target.c_str()) == 0 || errno == ENOENT, "Failed to unlink file %s: %s", target.c_str(),
365 int fdout = open(target.c_str(), O_CREAT | O_RDWR | O_EXCL, S_IRWXU);
366 xbt_assert(fdout >= 0, "Cannot write into %s: %s", target.c_str(), strerror(errno));
368 XBT_DEBUG("Copy %" PRIdMAX " bytes into %s", static_cast<intmax_t>(fdin_size), target.c_str());
370 ssize_t sent_size = sendfile(fdout, fdin, nullptr, fdin_size);
371 if (sent_size == fdin_size) {
376 xbt_assert(sent_size == -1 && errno == ENOSYS,
377 "Error while copying %s: only %zd bytes copied instead of %" PRIdMAX " (errno: %d -- %s)", target.c_str(),
378 sent_size, static_cast<intmax_t>(fdin_size), errno, strerror(errno));
380 // If this point is reached, sendfile() actually is not available. Copy file by hand.
381 std::vector<unsigned char> buf(1024 * 1024 * 4);
382 while (ssize_t got = read(fdin, buf.data(), buf.size())) {
384 xbt_assert(errno == EINTR, "Cannot read from %s", src.c_str());
387 const unsigned char* p = buf.data();
389 while (ssize_t done = write(fdout, p, todo)) {
391 xbt_assert(errno == EINTR, "Cannot write into %s", target.c_str());
402 #if not defined(__APPLE__) && not defined(__HAIKU__)
403 static int visit_libs(struct dl_phdr_info* info, size_t, void* data)
405 auto* libname = static_cast<std::string*>(data);
406 std::string path = info->dlpi_name;
407 if (path.find(*libname) != std::string::npos) {
408 *libname = std::move(path);
415 static void smpi_init_privatization_dlopen(const std::string& executable)
417 // Prepare the copy of the binary (get its size)
418 struct stat fdin_stat;
419 stat(executable.c_str(), &fdin_stat);
420 off_t fdin_size = fdin_stat.st_size;
422 std::string libnames = simgrid::config::get_value<std::string>("smpi/privatize-libs");
423 if (not libnames.empty()) {
425 std::vector<std::string> privatize_libs;
426 boost::split(privatize_libs, libnames, boost::is_any_of(";"));
428 for (auto const& libname : privatize_libs) {
429 // load the library once to add it to the local libs, to get the absolute path
430 void* libhandle = dlopen(libname.c_str(), RTLD_LAZY);
431 xbt_assert(libhandle != nullptr, "Cannot dlopen %s - check your settings in smpi/privatize-libs",
433 // get library name from path
434 std::string fullpath = libname;
435 #if not defined(__APPLE__) && not defined(__HAIKU__)
436 xbt_assert(dl_iterate_phdr(visit_libs, &fullpath) != 0,
437 "Can't find a linked %s - check your settings in smpi/privatize-libs", fullpath.c_str());
438 XBT_DEBUG("Extra lib to privatize '%s' found", fullpath.c_str());
440 xbt_die("smpi/privatize-libs is not (yet) compatible with OSX nor with Haiku");
442 privatize_libs_paths.emplace_back(std::move(fullpath));
447 simgrid::s4u::Engine::get_instance()->register_default([executable, fdin_size](std::vector<std::string> args) {
448 return simgrid::kernel::actor::ActorCode([executable, fdin_size, args = std::move(args)] {
449 static std::size_t rank = 0;
450 // Copy the dynamic library:
451 simgrid::xbt::Path path(executable);
452 std::string target_executable = simgrid::config::get_value<std::string>("smpi/tmpdir") + "/" +
453 path.get_base_name() + "_" + std::to_string(getpid()) + "_" + std::to_string(rank) + ".so";
455 smpi_copy_file(executable, target_executable, fdin_size);
456 // if smpi/privatize-libs is set, duplicate pointed lib and link each executable copy to a different one.
457 std::vector<std::string> target_libs;
458 for (auto const& libpath : privatize_libs_paths) {
459 // if we were given a full path, strip it
460 size_t index = libpath.find_last_of("/\\");
462 if (index != std::string::npos)
463 libname = libpath.substr(index + 1);
465 if (not libname.empty()) {
466 // load the library to add it to the local libs, to get the absolute path
467 struct stat fdin_stat2;
468 stat(libpath.c_str(), &fdin_stat2);
469 off_t fdin_size2 = fdin_stat2.st_size;
471 // Copy the dynamic library, the new name must be the same length as the old one
472 // just replace the name with 7 digits for the rank and the rest of the name.
473 auto pad = std::min<size_t>(7, libname.length());
474 std::string target_libname = std::string(pad - std::to_string(rank).length(), '0') + std::to_string(rank) + libname.substr(pad);
475 std::string target_lib = simgrid::config::get_value<std::string>("smpi/tmpdir") + "/" + target_libname;
476 target_libs.push_back(target_lib);
477 XBT_DEBUG("copy lib %s to %s, with size %lld", libpath.c_str(), target_lib.c_str(), (long long)fdin_size2);
478 smpi_copy_file(libpath, target_lib, fdin_size2);
480 std::string sedcommand = "sed -i -e 's/" + libname + "/" + target_libname + "/g' " + target_executable;
481 int status = system(sedcommand.c_str());
482 xbt_assert(status == 0, "error while applying sed command %s \n", sedcommand.c_str());
487 // Load the copy and resolve the entry point:
488 void* handle = dlopen(target_executable.c_str(), RTLD_LAZY | RTLD_LOCAL | WANT_RTLD_DEEPBIND);
489 int saved_errno = errno;
490 if (not simgrid::config::get_value<bool>("smpi/keep-temps")) {
491 unlink(target_executable.c_str());
492 for (const std::string& target_lib : target_libs)
493 unlink(target_lib.c_str());
495 xbt_assert(handle != nullptr,
496 "dlopen failed: %s (errno: %d -- %s).\nError: Did you compile the program with a SMPI-specific "
497 "compiler (spmicc or friends)?",
498 dlerror(), saved_errno, strerror(saved_errno));
500 smpi_entry_point_type entry_point = smpi_resolve_function(handle);
501 xbt_assert(entry_point, "Could not resolve entry point. Does your program contain a main() function?");
502 smpi_run_entry_point(entry_point, executable, args);
507 static void smpi_init_privatization_no_dlopen(const std::string& executable)
509 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP)
510 smpi_prepare_global_memory_segment();
512 // Load the dynamic library and resolve the entry point:
513 void* handle = dlopen(executable.c_str(), RTLD_LAZY | RTLD_LOCAL);
514 xbt_assert(handle != nullptr, "dlopen failed for %s: %s (errno: %d -- %s)", executable.c_str(), dlerror(), errno,
516 smpi_entry_point_type entry_point = smpi_resolve_function(handle);
517 xbt_assert(entry_point, "main not found in %s", executable.c_str());
519 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP)
520 smpi_backup_global_memory_segment();
522 // Execute the same entry point for each simulated process:
523 simgrid::s4u::Engine::get_instance()->register_default([entry_point, executable](std::vector<std::string> args) {
524 return simgrid::kernel::actor::ActorCode([entry_point, executable, args = std::move(args)] {
525 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) {
526 simgrid::smpi::ActorExt* ext = smpi_process();
527 /* Now using the segment index of this process */
528 ext->set_privatized_region(smpi_init_global_memory_segment_process());
529 /* Done at the process's creation */
530 smpi_switch_data_segment(simgrid::s4u::Actor::self());
532 smpi_run_entry_point(entry_point, executable, args);
537 int smpi_main(const char* executable, int argc, char* argv[])
539 if (getenv("SMPI_PRETEND_CC") != nullptr) {
540 /* Hack to ensure that smpicc can pretend to be a simple compiler. Particularly handy to pass it to the
541 * configuration tools */
545 smpi_init_options_internal(true);
546 simgrid::s4u::Engine engine(&argc, argv);
548 sg_storage_file_system_init();
549 // parse the platform file: get the host list
550 engine.load_platform(argv[1]);
551 engine.set_default_comm_data_copy_callback(smpi_comm_copy_buffer_callback);
553 if (smpi_cfg_privatization() == SmpiPrivStrategies::DLOPEN)
554 smpi_init_privatization_dlopen(executable);
556 smpi_init_privatization_no_dlopen(executable);
558 simgrid::smpi::colls::set_collectives();
559 simgrid::smpi::colls::smpi_coll_cleanup_callback = nullptr;
563 const std::vector<const char*> args(argv + 2, argv + argc);
565 smpi_deployment_smpirun(&engine, smpi_hostfile.get(), smpi_np.get(), smpi_replay.get(), smpi_map.get(), args);
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 engine.get_impl()->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();
639 void smpi_exit(int res){
641 XBT_WARN("SMPI process did not return 0. Return value : %d", res);
642 smpi_exit_status = res;
644 simgrid::s4u::this_actor::exit();
645 while(1);//necessary for the noreturn attribute