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/s4u/Engine.hpp"
8 #include "simgrid/plugins/file_system.h"
9 #include "smpi_coll.hpp"
10 #include "smpi_f2c.hpp"
11 #include "smpi_host.hpp"
12 #include "smpi_config.hpp"
13 #include "src/kernel/activity/CommImpl.hpp"
14 #include "src/simix/smx_private.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>
23 #include <boost/core/demangle.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 void (*smpi_comm_copy_data_callback)(simgrid::kernel::activity::CommImpl*, void*,
84 size_t) = &smpi_comm_copy_buffer_callback;
86 simgrid::smpi::ActorExt* smpi_process()
88 simgrid::s4u::ActorPtr me = simgrid::s4u::Actor::self();
90 if (me == nullptr) // This happens sometimes (eg, when linking against NS3 because it pulls openMPI...)
93 return me->extension<simgrid::smpi::ActorExt>();
96 simgrid::smpi::ActorExt* smpi_process_remote(simgrid::s4u::ActorPtr actor)
98 if (actor.get() == nullptr)
100 return actor->extension<simgrid::smpi::ActorExt>();
103 MPI_Comm smpi_process_comm_self(){
104 return smpi_process()->comm_self();
107 MPI_Info smpi_process_info_env(){
108 return smpi_process()->info_env();
111 void * smpi_process_get_user_data(){
112 return simgrid::s4u::Actor::self()->get_data();
115 void smpi_process_set_user_data(void *data){
116 simgrid::s4u::Actor::self()->set_data(data);
119 void smpi_comm_set_copy_data_callback(void (*callback) (smx_activity_t, void*, size_t))
121 static void (*saved_callback)(smx_activity_t, void*, size_t);
122 saved_callback = callback;
123 smpi_comm_copy_data_callback = [](simgrid::kernel::activity::CommImpl* comm, void* buff, size_t size) {
124 saved_callback(comm, buff, size);
128 static void memcpy_private(void* dest, const void* src, const std::vector<std::pair<size_t, size_t>>& private_blocks)
130 for (auto const& block : private_blocks)
131 memcpy((uint8_t*)dest+block.first, (uint8_t*)src+block.first, block.second-block.first);
134 static void check_blocks(const std::vector<std::pair<size_t, size_t>>& private_blocks, size_t buff_size)
136 for (auto const& block : private_blocks)
137 xbt_assert(block.first <= block.second && block.second <= buff_size, "Oops, bug in shared malloc.");
140 static void smpi_cleanup_comm_after_copy(simgrid::kernel::activity::CommImpl* comm, void* buff){
141 if (comm->detached()) {
142 // if this is a detached send, the source buffer was duplicated by SMPI
143 // sender to make the original buffer available to the application ASAP
145 //It seems that the request is used after the call there this should be free somewhere else but where???
146 //xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free
147 comm->src_buff_ = nullptr;
151 void smpi_comm_copy_buffer_callback(simgrid::kernel::activity::CommImpl* comm, void* buff, size_t buff_size)
153 size_t src_offset = 0;
154 size_t dst_offset = 0;
155 std::vector<std::pair<size_t, size_t>> src_private_blocks;
156 std::vector<std::pair<size_t, size_t>> dst_private_blocks;
157 XBT_DEBUG("Copy the data over");
158 if(smpi_is_shared(buff, src_private_blocks, &src_offset)) {
159 src_private_blocks = shift_and_frame_private_blocks(src_private_blocks, src_offset, buff_size);
160 if (src_private_blocks.empty()) { // simple shared malloc ... return.
161 XBT_VERB("Sender is shared. Let's ignore it.");
162 smpi_cleanup_comm_after_copy(comm, buff);
167 src_private_blocks.clear();
168 src_private_blocks.emplace_back(0, buff_size);
170 if (smpi_is_shared((char*)comm->dst_buff_, dst_private_blocks, &dst_offset)) {
171 dst_private_blocks = shift_and_frame_private_blocks(dst_private_blocks, dst_offset, buff_size);
172 if (dst_private_blocks.empty()) { // simple shared malloc ... return.
173 XBT_VERB("Receiver is shared. Let's ignore it.");
174 smpi_cleanup_comm_after_copy(comm, buff);
179 dst_private_blocks.clear();
180 dst_private_blocks.emplace_back(0, buff_size);
182 check_blocks(src_private_blocks, buff_size);
183 check_blocks(dst_private_blocks, buff_size);
184 auto private_blocks = merge_private_blocks(src_private_blocks, dst_private_blocks);
185 check_blocks(private_blocks, buff_size);
187 if ((smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) &&
188 (static_cast<char*>(buff) >= smpi_data_exe_start) &&
189 (static_cast<char*>(buff) < smpi_data_exe_start + smpi_data_exe_size)) {
190 XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
191 smpi_switch_data_segment(comm->src_actor_->get_iface());
192 tmpbuff = xbt_malloc(buff_size);
193 memcpy_private(tmpbuff, buff, private_blocks);
196 if ((smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) &&
197 ((char*)comm->dst_buff_ >= smpi_data_exe_start) &&
198 ((char*)comm->dst_buff_ < smpi_data_exe_start + smpi_data_exe_size)) {
199 XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
200 smpi_switch_data_segment(comm->dst_actor_->get_iface());
202 XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff, comm->dst_buff_);
203 memcpy_private(comm->dst_buff_, tmpbuff, private_blocks);
205 smpi_cleanup_comm_after_copy(comm,buff);
210 void smpi_comm_null_copy_buffer_callback(simgrid::kernel::activity::CommImpl*, void*, size_t)
212 /* nothing done in this version */
216 return MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED;
219 static void smpi_init_papi()
222 // This map holds for each computation unit (such as "default" or "process1" etc.)
223 // the configuration as given by the user (counter data as a pair of (counter_name, counter_counter))
224 // and the (computed) event_set.
226 if (not smpi_cfg_papi_events_file().empty()) {
227 if (PAPI_library_init(PAPI_VER_CURRENT) != PAPI_VER_CURRENT)
228 XBT_ERROR("Could not initialize PAPI library; is it correctly installed and linked?"
229 " Expected version is %u", PAPI_VER_CURRENT);
231 using Tokenizer = boost::tokenizer<boost::char_separator<char>>;
232 boost::char_separator<char> separator_units(";");
233 std::string str = smpi_cfg_papi_events_file();
234 Tokenizer tokens(str, separator_units);
236 // Iterate over all the computational units. This could be processes, hosts, threads, ranks... You name it.
237 // I'm not exactly sure what we will support eventually, so I'll leave it at the general term "units".
238 for (auto const& unit_it : tokens) {
239 boost::char_separator<char> separator_events(":");
240 Tokenizer event_tokens(unit_it, separator_events);
242 int event_set = PAPI_NULL;
243 if (PAPI_create_eventset(&event_set) != PAPI_OK) {
244 // TODO: Should this let the whole simulation die?
245 XBT_CRITICAL("Could not create PAPI event set during init.");
248 // NOTE: We cannot use a map here, as we must obey the order of the counters
249 // This is important for PAPI: We need to map the values of counters back to the event_names (so, when PAPI_read()
251 papi_counter_t counters2values;
253 // Iterate over all counters that were specified for this specific unit.
254 // Note that we need to remove the name of the unit (that could also be the "default" value), which always comes
255 // first. Hence, we start at ++(events.begin())!
256 for (Tokenizer::iterator events_it = ++(event_tokens.begin()); events_it != event_tokens.end(); ++events_it) {
257 int event_code = PAPI_NULL;
258 auto* event_name = const_cast<char*>((*events_it).c_str());
259 if (PAPI_event_name_to_code(event_name, &event_code) != PAPI_OK) {
260 XBT_CRITICAL("Could not find PAPI event '%s'. Skipping.", event_name);
263 if (PAPI_add_event(event_set, event_code) != PAPI_OK) {
264 XBT_ERROR("Could not add PAPI event '%s'. Skipping.", event_name);
267 XBT_DEBUG("Successfully added PAPI event '%s' to the event set.", event_name);
269 counters2values.emplace_back(*events_it, 0LL);
272 std::string unit_name = *(event_tokens.begin());
273 papi_process_data config = {.counter_data = std::move(counters2values), .event_set = event_set};
275 units2papi_setup.insert(std::make_pair(unit_name, std::move(config)));
281 using smpi_entry_point_type = std::function<int(int argc, char* argv[])>;
282 using smpi_c_entry_point_type = int (*)(int argc, char** argv);
283 using smpi_fortran_entry_point_type = void (*)();
285 template <typename F>
286 static int smpi_run_entry_point(const F& entry_point, const std::string& executable_path, std::vector<std::string> args)
288 // copy C strings, we need them writable
289 auto* args4argv = new std::vector<char*>(args.size());
290 std::transform(begin(args), end(args), begin(*args4argv), [](const std::string& s) { return xbt_strdup(s.c_str()); });
292 // set argv[0] to executable_path
293 xbt_free((*args4argv)[0]);
294 (*args4argv)[0] = xbt_strdup(executable_path.c_str());
297 // take a copy of args4argv to keep reference of the allocated strings
298 const std::vector<char*> args2str(*args4argv);
300 int argc = args4argv->size();
301 args4argv->push_back(nullptr);
302 char** argv = args4argv->data();
305 for_rtl_init_ (&argc, argv);
310 _gfortran_set_args(argc, argv);
312 int res = entry_point(argc, argv);
317 for (char* s : args2str)
323 XBT_WARN("SMPI process did not return 0. Return value : %d", res);
324 if (smpi_exit_status == 0)
325 smpi_exit_status = res;
331 // TODO, remove the number of functions involved here
332 static smpi_entry_point_type smpi_resolve_function(void* handle)
334 auto* entry_point_fortran = reinterpret_cast<smpi_fortran_entry_point_type>(dlsym(handle, "user_main_"));
335 if (entry_point_fortran != nullptr) {
336 return [entry_point_fortran](int, char**) {
337 entry_point_fortran();
342 auto* entry_point = reinterpret_cast<smpi_c_entry_point_type>(dlsym(handle, "main"));
343 if (entry_point != nullptr) {
347 return smpi_entry_point_type();
350 static void smpi_copy_file(const std::string& src, const std::string& target, off_t fdin_size)
352 int fdin = open(src.c_str(), O_RDONLY);
353 xbt_assert(fdin >= 0, "Cannot read from %s. Please make sure that the file exists and is executable.", src.c_str());
354 xbt_assert(unlink(target.c_str()) == 0 || errno == ENOENT, "Failed to unlink file %s: %s", target.c_str(),
356 int fdout = open(target.c_str(), O_CREAT | O_RDWR | O_EXCL, S_IRWXU);
357 xbt_assert(fdout >= 0, "Cannot write into %s: %s", target.c_str(), strerror(errno));
359 XBT_DEBUG("Copy %" PRIdMAX " bytes into %s", static_cast<intmax_t>(fdin_size), target.c_str());
361 ssize_t sent_size = sendfile(fdout, fdin, nullptr, fdin_size);
362 if (sent_size == fdin_size) {
367 xbt_assert(sent_size == -1 && errno == ENOSYS,
368 "Error while copying %s: only %zd bytes copied instead of %" PRIdMAX " (errno: %d -- %s)", target.c_str(),
369 sent_size, static_cast<intmax_t>(fdin_size), errno, strerror(errno));
371 // If this point is reached, sendfile() actually is not available. Copy file by hand.
372 std::vector<unsigned char> buf(1024 * 1024 * 4);
373 while (ssize_t got = read(fdin, buf.data(), buf.size())) {
375 xbt_assert(errno == EINTR, "Cannot read from %s", src.c_str());
377 const unsigned char* p = buf.data();
379 while (ssize_t done = write(fdout, p, todo)) {
381 xbt_assert(errno == EINTR, "Cannot write into %s", target.c_str());
393 #if not defined(__APPLE__) && not defined(__HAIKU__)
394 static int visit_libs(struct dl_phdr_info* info, size_t, void* data)
396 auto* libname = static_cast<std::string*>(data);
397 std::string path = info->dlpi_name;
398 if (path.find(*libname) != std::string::npos) {
399 *libname = std::move(path);
406 static void smpi_init_privatization_dlopen(const std::string& executable)
408 // Prepare the copy of the binary (get its size)
409 struct stat fdin_stat;
410 stat(executable.c_str(), &fdin_stat);
411 off_t fdin_size = fdin_stat.st_size;
413 std::string libnames = simgrid::config::get_value<std::string>("smpi/privatize-libs");
414 if (not libnames.empty()) {
416 std::vector<std::string> privatize_libs;
417 boost::split(privatize_libs, libnames, boost::is_any_of(";"));
419 for (auto const& libname : privatize_libs) {
420 // load the library once to add it to the local libs, to get the absolute path
421 void* libhandle = dlopen(libname.c_str(), RTLD_LAZY);
422 xbt_assert(libhandle != nullptr, "Cannot dlopen %s - check your settings in smpi/privatize-libs",
424 // get library name from path
425 std::string fullpath = libname;
426 #if not defined(__APPLE__) && not defined(__HAIKU__)
427 xbt_assert(dl_iterate_phdr(visit_libs, &fullpath) != 0,
428 "Can't find a linked %s - check your settings in smpi/privatize-libs", fullpath.c_str());
429 XBT_DEBUG("Extra lib to privatize '%s' found", fullpath.c_str());
431 xbt_die("smpi/privatize-libs is not (yet) compatible with OSX nor with Haiku");
433 privatize_libs_paths.emplace_back(std::move(fullpath));
438 simgrid::s4u::Engine::get_instance()->register_default([executable, fdin_size](std::vector<std::string> args) {
439 return std::function<void()>([executable, fdin_size, args] {
440 static std::size_t rank = 0;
441 // Copy the dynamic library:
442 simgrid::xbt::Path path(executable);
443 std::string target_executable = simgrid::config::get_value<std::string>("smpi/tmpdir") + "/" +
444 path.get_base_name() + "_" + std::to_string(getpid()) + "_" + std::to_string(rank) + ".so";
446 smpi_copy_file(executable, target_executable, fdin_size);
447 // if smpi/privatize-libs is set, duplicate pointed lib and link each executable copy to a different one.
448 std::vector<std::string> target_libs;
449 for (auto const& libpath : privatize_libs_paths) {
450 // if we were given a full path, strip it
451 size_t index = libpath.find_last_of("/\\");
453 if (index != std::string::npos)
454 libname = libpath.substr(index + 1);
456 if (not libname.empty()) {
457 // load the library to add it to the local libs, to get the absolute path
458 struct stat fdin_stat2;
459 stat(libpath.c_str(), &fdin_stat2);
460 off_t fdin_size2 = fdin_stat2.st_size;
462 // Copy the dynamic library, the new name must be the same length as the old one
463 // just replace the name with 7 digits for the rank and the rest of the name.
464 auto pad = std::min<unsigned>(7, libname.length());
465 std::string target_libname = std::string(pad - std::to_string(rank).length(), '0') + std::to_string(rank) + libname.substr(pad);
466 std::string target_lib = simgrid::config::get_value<std::string>("smpi/tmpdir") + "/" + target_libname;
467 target_libs.push_back(target_lib);
468 XBT_DEBUG("copy lib %s to %s, with size %lld", libpath.c_str(), target_lib.c_str(), (long long)fdin_size2);
469 smpi_copy_file(libpath, target_lib, fdin_size2);
471 std::string sedcommand = "sed -i -e 's/" + libname + "/" + target_libname + "/g' " + target_executable;
472 int status = system(sedcommand.c_str());
473 xbt_assert(status == 0, "error while applying sed command %s \n", sedcommand.c_str());
478 // Load the copy and resolve the entry point:
479 void* handle = dlopen(target_executable.c_str(), RTLD_LAZY | RTLD_LOCAL | WANT_RTLD_DEEPBIND);
480 int saved_errno = errno;
481 if (not simgrid::config::get_value<bool>("smpi/keep-temps")) {
482 unlink(target_executable.c_str());
483 for (const std::string& target_lib : target_libs)
484 unlink(target_lib.c_str());
486 xbt_assert(handle != nullptr,
487 "dlopen failed: %s (errno: %d -- %s).\nError: Did you compile the program with a SMPI-specific "
488 "compiler (spmicc or friends)?",
489 dlerror(), saved_errno, strerror(saved_errno));
491 smpi_entry_point_type entry_point = smpi_resolve_function(handle);
492 xbt_assert(entry_point, "Could not resolve entry point");
493 smpi_run_entry_point(entry_point, executable, args);
498 static void smpi_init_privatization_no_dlopen(const std::string& executable)
500 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP)
501 smpi_prepare_global_memory_segment();
503 // Load the dynamic library and resolve the entry point:
504 void* handle = dlopen(executable.c_str(), RTLD_LAZY | RTLD_LOCAL);
505 xbt_assert(handle != nullptr, "dlopen failed for %s: %s (errno: %d -- %s)", executable.c_str(), dlerror(), errno,
507 smpi_entry_point_type entry_point = smpi_resolve_function(handle);
508 xbt_assert(entry_point, "main not found in %s", executable.c_str());
510 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP)
511 smpi_backup_global_memory_segment();
513 // Execute the same entry point for each simulated process:
514 simgrid::s4u::Engine::get_instance()->register_default([entry_point, executable](std::vector<std::string> args) {
515 return std::function<void()>(
516 [entry_point, executable, args] { 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_switch_data_segment = &smpi_switch_data_segment;
530 simgrid::instr::init();
531 SIMIX_global_init(&argc, argv);
533 auto engine = simgrid::s4u::Engine::get_instance();
535 sg_storage_file_system_init();
536 // parse the platform file: get the host list
537 engine->load_platform(argv[1]);
538 simgrid::kernel::activity::CommImpl::set_copy_data_callback(smpi_comm_copy_buffer_callback);
540 if (smpi_cfg_privatization() == SmpiPrivStrategies::DLOPEN)
541 smpi_init_privatization_dlopen(executable);
543 smpi_init_privatization_no_dlopen(executable);
545 simgrid::smpi::colls::set_collectives();
546 simgrid::smpi::colls::smpi_coll_cleanup_callback = nullptr;
550 /* This is a ... heavy way to count the MPI ranks */
552 simgrid::s4u::Actor::on_creation.connect([&rank_counts](const simgrid::s4u::Actor& actor) {
553 if (not actor.is_daemon())
556 engine->load_deployment(argv[2]);
558 SMPI_app_instance_register(smpi_default_instance_name.c_str(), nullptr, rank_counts);
559 MPI_COMM_WORLD = *smpi_deployment_comm_world(smpi_default_instance_name);
561 /* Clean IO before the run */
565 if (MC_is_active()) {
570 xbt_os_walltimer_stop(global_timer);
571 simgrid::smpi::utils::print_time_analysis(xbt_os_timer_elapsed(global_timer));
575 return smpi_exit_status;
578 // Called either directly from the user code, or from the code called by smpirun
581 simgrid::s4u::Actor::on_creation.connect([](simgrid::s4u::Actor& actor) {
582 if (not actor.is_daemon())
583 actor.extension_set<simgrid::smpi::ActorExt>(new simgrid::smpi::ActorExt(&actor));
585 simgrid::s4u::Host::on_creation.connect(
586 [](simgrid::s4u::Host& host) { host.extension_set(new simgrid::smpi::Host(&host)); });
587 for (auto const& host : simgrid::s4u::Engine::get_instance()->get_all_hosts())
588 host->extension_set(new simgrid::smpi::Host(host));
590 if (not MC_is_active()) {
591 global_timer = xbt_os_timer_new();
592 xbt_os_walltimer_start(global_timer);
595 smpi_check_options();
600 smpi_bench_destroy();
601 smpi_shared_destroy();
602 smpi_deployment_cleanup_instances();
604 if (simgrid::smpi::colls::smpi_coll_cleanup_callback != nullptr)
605 simgrid::smpi::colls::smpi_coll_cleanup_callback();
607 MPI_COMM_WORLD = MPI_COMM_NULL;
609 if (not MC_is_active()) {
610 xbt_os_timer_free(global_timer);
613 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP)
614 smpi_destroy_global_memory_segments();
616 simgrid::smpi::utils::print_memory_analysis();
619 void smpi_mpi_init() {
620 smpi_init_fortran_types();
621 if(smpi_init_sleep > 0)
622 simgrid::s4u::this_actor::sleep_for(smpi_init_sleep);
625 void SMPI_thread_create() {
626 TRACE_smpi_init(simgrid::s4u::this_actor::get_pid(), __func__);
627 smpi_process()->mark_as_initialized();