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>
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");
79 static simgrid::config::Flag<double> smpi_init_sleep(
80 "smpi/init", "Time to inject inside a call to MPI_Init", 0.0);
82 void (*smpi_comm_copy_data_callback)(simgrid::kernel::activity::CommImpl*, void*,
83 size_t) = &smpi_comm_copy_buffer_callback;
85 simgrid::smpi::ActorExt* smpi_process()
87 simgrid::s4u::ActorPtr me = simgrid::s4u::Actor::self();
89 if (me == nullptr) // This happens sometimes (eg, when linking against NS3 because it pulls openMPI...)
92 return me->extension<simgrid::smpi::ActorExt>();
95 simgrid::smpi::ActorExt* smpi_process_remote(simgrid::s4u::ActorPtr actor)
97 if (actor.get() == nullptr)
99 return actor->extension<simgrid::smpi::ActorExt>();
102 MPI_Comm smpi_process_comm_self(){
103 return smpi_process()->comm_self();
106 MPI_Info smpi_process_info_env(){
107 return smpi_process()->info_env();
110 void * smpi_process_get_user_data(){
111 return simgrid::s4u::Actor::self()->get_data();
114 void smpi_process_set_user_data(void *data){
115 simgrid::s4u::Actor::self()->set_data(data);
118 void smpi_comm_set_copy_data_callback(void (*callback) (smx_activity_t, void*, size_t))
120 static void (*saved_callback)(smx_activity_t, void*, size_t);
121 saved_callback = callback;
122 smpi_comm_copy_data_callback = [](simgrid::kernel::activity::CommImpl* comm, void* buff, size_t size) {
123 saved_callback(comm, buff, size);
127 static void memcpy_private(void* dest, const void* src, const std::vector<std::pair<size_t, size_t>>& private_blocks)
129 for (auto const& block : private_blocks)
130 memcpy((uint8_t*)dest+block.first, (uint8_t*)src+block.first, block.second-block.first);
133 static void check_blocks(const std::vector<std::pair<size_t, size_t>>& private_blocks, size_t buff_size)
135 for (auto const& block : private_blocks)
136 xbt_assert(block.first <= block.second && block.second <= buff_size, "Oops, bug in shared malloc.");
139 static void smpi_cleanup_comm_after_copy(simgrid::kernel::activity::CommImpl* comm, void* buff){
140 if (comm->detached()) {
141 // if this is a detached send, the source buffer was duplicated by SMPI
142 // sender to make the original buffer available to the application ASAP
144 //It seems that the request is used after the call there this should be free somewhere else but where???
145 //xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free
146 comm->src_buff_ = nullptr;
150 void smpi_comm_copy_buffer_callback(simgrid::kernel::activity::CommImpl* comm, void* buff, size_t buff_size)
152 size_t src_offset = 0;
153 size_t dst_offset = 0;
154 std::vector<std::pair<size_t, size_t>> src_private_blocks;
155 std::vector<std::pair<size_t, size_t>> dst_private_blocks;
156 XBT_DEBUG("Copy the data over");
157 if(smpi_is_shared(buff, src_private_blocks, &src_offset)) {
158 src_private_blocks = shift_and_frame_private_blocks(src_private_blocks, src_offset, buff_size);
159 if (src_private_blocks.empty()) { // simple shared malloc ... return.
160 XBT_VERB("Sender is shared. Let's ignore it.");
161 smpi_cleanup_comm_after_copy(comm, buff);
166 src_private_blocks.clear();
167 src_private_blocks.emplace_back(0, buff_size);
169 if (smpi_is_shared((char*)comm->dst_buff_, dst_private_blocks, &dst_offset)) {
170 dst_private_blocks = shift_and_frame_private_blocks(dst_private_blocks, dst_offset, buff_size);
171 if (dst_private_blocks.empty()) { // simple shared malloc ... return.
172 XBT_VERB("Receiver is shared. Let's ignore it.");
173 smpi_cleanup_comm_after_copy(comm, buff);
178 dst_private_blocks.clear();
179 dst_private_blocks.emplace_back(0, buff_size);
181 check_blocks(src_private_blocks, buff_size);
182 check_blocks(dst_private_blocks, buff_size);
183 auto private_blocks = merge_private_blocks(src_private_blocks, dst_private_blocks);
184 check_blocks(private_blocks, buff_size);
186 if ((smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) &&
187 (static_cast<char*>(buff) >= smpi_data_exe_start) &&
188 (static_cast<char*>(buff) < smpi_data_exe_start + smpi_data_exe_size)) {
189 XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
190 smpi_switch_data_segment(comm->src_actor_->get_iface());
191 tmpbuff = xbt_malloc(buff_size);
192 memcpy_private(tmpbuff, buff, private_blocks);
195 if ((smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) &&
196 ((char*)comm->dst_buff_ >= smpi_data_exe_start) &&
197 ((char*)comm->dst_buff_ < smpi_data_exe_start + smpi_data_exe_size)) {
198 XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
199 smpi_switch_data_segment(comm->dst_actor_->get_iface());
201 XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff, comm->dst_buff_);
202 memcpy_private(comm->dst_buff_, tmpbuff, private_blocks);
204 smpi_cleanup_comm_after_copy(comm,buff);
209 void smpi_comm_null_copy_buffer_callback(simgrid::kernel::activity::CommImpl*, void*, size_t)
211 /* nothing done in this version */
215 return MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED;
218 static void smpi_init_papi()
221 // This map holds for each computation unit (such as "default" or "process1" etc.)
222 // the configuration as given by the user (counter data as a pair of (counter_name, counter_counter))
223 // and the (computed) event_set.
225 if (not smpi_cfg_papi_events_file().empty()) {
226 if (PAPI_library_init(PAPI_VER_CURRENT) != PAPI_VER_CURRENT)
227 XBT_ERROR("Could not initialize PAPI library; is it correctly installed and linked?"
228 " Expected version is %u", PAPI_VER_CURRENT);
230 using Tokenizer = boost::tokenizer<boost::char_separator<char>>;
231 boost::char_separator<char> separator_units(";");
232 std::string str = smpi_cfg_papi_events_file();
233 Tokenizer tokens(str, separator_units);
235 // Iterate over all the computational units. This could be processes, hosts, threads, ranks... You name it.
236 // I'm not exactly sure what we will support eventually, so I'll leave it at the general term "units".
237 for (auto const& unit_it : tokens) {
238 boost::char_separator<char> separator_events(":");
239 Tokenizer event_tokens(unit_it, separator_events);
241 int event_set = PAPI_NULL;
242 if (PAPI_create_eventset(&event_set) != PAPI_OK) {
243 // TODO: Should this let the whole simulation die?
244 XBT_CRITICAL("Could not create PAPI event set during init.");
247 // NOTE: We cannot use a map here, as we must obey the order of the counters
248 // This is important for PAPI: We need to map the values of counters back to the event_names (so, when PAPI_read()
250 papi_counter_t counters2values;
252 // Iterate over all counters that were specified for this specific unit.
253 // Note that we need to remove the name of the unit (that could also be the "default" value), which always comes
254 // first. Hence, we start at ++(events.begin())!
255 for (Tokenizer::iterator events_it = ++(event_tokens.begin()); events_it != event_tokens.end(); ++events_it) {
256 int event_code = PAPI_NULL;
257 auto* event_name = const_cast<char*>((*events_it).c_str());
258 if (PAPI_event_name_to_code(event_name, &event_code) != PAPI_OK) {
259 XBT_CRITICAL("Could not find PAPI event '%s'. Skipping.", event_name);
262 if (PAPI_add_event(event_set, event_code) != PAPI_OK) {
263 XBT_ERROR("Could not add PAPI event '%s'. Skipping.", event_name);
266 XBT_DEBUG("Successfully added PAPI event '%s' to the event set.", event_name);
268 counters2values.emplace_back(*events_it, 0LL);
271 std::string unit_name = *(event_tokens.begin());
272 papi_process_data config = {.counter_data = std::move(counters2values), .event_set = event_set};
274 units2papi_setup.insert(std::make_pair(unit_name, std::move(config)));
280 using smpi_entry_point_type = std::function<int(int argc, char* argv[])>;
281 using smpi_c_entry_point_type = int (*)(int argc, char** argv);
282 using smpi_fortran_entry_point_type = void (*)();
284 template <typename F>
285 static int smpi_run_entry_point(const F& entry_point, const std::string& executable_path, std::vector<std::string> args)
287 // copy C strings, we need them writable
288 auto* args4argv = new std::vector<char*>(args.size());
289 std::transform(begin(args), end(args), begin(*args4argv), [](const std::string& s) { return xbt_strdup(s.c_str()); });
291 // set argv[0] to executable_path
292 xbt_free((*args4argv)[0]);
293 (*args4argv)[0] = xbt_strdup(executable_path.c_str());
296 // take a copy of args4argv to keep reference of the allocated strings
297 const std::vector<char*> args2str(*args4argv);
299 int argc = args4argv->size();
300 args4argv->push_back(nullptr);
301 char** argv = args4argv->data();
304 for_rtl_init_ (&argc, argv);
309 _gfortran_set_args(argc, argv);
311 int res = entry_point(argc, argv);
316 for (char* s : args2str)
322 XBT_WARN("SMPI process did not return 0. Return value : %d", res);
323 if (smpi_exit_status == 0)
324 smpi_exit_status = res;
330 // TODO, remove the number of functions involved here
331 static smpi_entry_point_type smpi_resolve_function(void* handle)
333 auto* entry_point_fortran = reinterpret_cast<smpi_fortran_entry_point_type>(dlsym(handle, "user_main_"));
334 if (entry_point_fortran != nullptr) {
335 return [entry_point_fortran](int, char**) {
336 entry_point_fortran();
341 auto* entry_point = reinterpret_cast<smpi_c_entry_point_type>(dlsym(handle, "main"));
342 if (entry_point != nullptr) {
346 return smpi_entry_point_type();
349 static void smpi_copy_file(const std::string& src, const std::string& target, off_t fdin_size)
351 int fdin = open(src.c_str(), O_RDONLY);
352 xbt_assert(fdin >= 0, "Cannot read from %s. Please make sure that the file exists and is executable.", src.c_str());
353 xbt_assert(unlink(target.c_str()) == 0 || errno == ENOENT, "Failed to unlink file %s: %s", target.c_str(),
355 int fdout = open(target.c_str(), O_CREAT | O_RDWR | O_EXCL, S_IRWXU);
356 xbt_assert(fdout >= 0, "Cannot write into %s: %s", target.c_str(), strerror(errno));
358 XBT_DEBUG("Copy %" PRIdMAX " bytes into %s", static_cast<intmax_t>(fdin_size), target.c_str());
360 ssize_t sent_size = sendfile(fdout, fdin, nullptr, fdin_size);
361 if (sent_size == fdin_size) {
366 xbt_assert(sent_size == -1 && errno == ENOSYS,
367 "Error while copying %s: only %zd bytes copied instead of %" PRIdMAX " (errno: %d -- %s)", target.c_str(),
368 sent_size, static_cast<intmax_t>(fdin_size), errno, strerror(errno));
370 // If this point is reached, sendfile() actually is not available. Copy file by hand.
371 std::vector<unsigned char> buf(1024 * 1024 * 4);
372 while (ssize_t got = read(fdin, buf.data(), buf.size())) {
374 xbt_assert(errno == EINTR, "Cannot read from %s", src.c_str());
376 const unsigned char* p = buf.data();
378 while (ssize_t done = write(fdout, p, todo)) {
380 xbt_assert(errno == EINTR, "Cannot write into %s", target.c_str());
392 #if not defined(__APPLE__) && not defined(__HAIKU__)
393 static int visit_libs(struct dl_phdr_info* info, size_t, void* data)
395 auto* libname = static_cast<std::string*>(data);
396 std::string path = info->dlpi_name;
397 if (path.find(*libname) != std::string::npos) {
398 *libname = std::move(path);
405 static void smpi_init_privatization_dlopen(const std::string& executable)
407 // Prepare the copy of the binary (get its size)
408 struct stat fdin_stat;
409 stat(executable.c_str(), &fdin_stat);
410 off_t fdin_size = fdin_stat.st_size;
412 std::string libnames = simgrid::config::get_value<std::string>("smpi/privatize-libs");
413 if (not libnames.empty()) {
415 std::vector<std::string> privatize_libs;
416 boost::split(privatize_libs, libnames, boost::is_any_of(";"));
418 for (auto const& libname : privatize_libs) {
419 // load the library once to add it to the local libs, to get the absolute path
420 void* libhandle = dlopen(libname.c_str(), RTLD_LAZY);
421 xbt_assert(libhandle != nullptr, "Cannot dlopen %s - check your settings in smpi/privatize-libs",
423 // get library name from path
424 std::string fullpath = libname;
425 #if not defined(__APPLE__) && not defined(__HAIKU__)
426 xbt_assert(dl_iterate_phdr(visit_libs, &fullpath) != 0,
427 "Can't find a linked %s - check your settings in smpi/privatize-libs", fullpath.c_str());
428 XBT_DEBUG("Extra lib to privatize '%s' found", fullpath.c_str());
430 xbt_die("smpi/privatize-libs is not (yet) compatible with OSX nor with Haiku");
432 privatize_libs_paths.emplace_back(std::move(fullpath));
437 simgrid::s4u::Engine::get_instance()->register_default([executable, fdin_size](std::vector<std::string> args) {
438 return std::function<void()>([executable, fdin_size, args] {
439 static std::size_t rank = 0;
440 // Copy the dynamic library:
441 simgrid::xbt::Path path(executable);
442 std::string target_executable = simgrid::config::get_value<std::string>("smpi/tmpdir") + "/" +
443 path.get_base_name() + "_" + std::to_string(getpid()) + "_" + std::to_string(rank) + ".so";
445 smpi_copy_file(executable, target_executable, fdin_size);
446 // if smpi/privatize-libs is set, duplicate pointed lib and link each executable copy to a different one.
447 std::vector<std::string> target_libs;
448 for (auto const& libpath : privatize_libs_paths) {
449 // if we were given a full path, strip it
450 size_t index = libpath.find_last_of("/\\");
452 if (index != std::string::npos)
453 libname = libpath.substr(index + 1);
455 if (not libname.empty()) {
456 // load the library to add it to the local libs, to get the absolute path
457 struct stat fdin_stat2;
458 stat(libpath.c_str(), &fdin_stat2);
459 off_t fdin_size2 = fdin_stat2.st_size;
461 // Copy the dynamic library, the new name must be the same length as the old one
462 // just replace the name with 7 digits for the rank and the rest of the name.
463 auto pad = std::min<unsigned>(7, libname.length());
464 std::string target_libname = std::string(pad - std::to_string(rank).length(), '0') + std::to_string(rank) + libname.substr(pad);
465 std::string target_lib = simgrid::config::get_value<std::string>("smpi/tmpdir") + "/" + target_libname;
466 target_libs.push_back(target_lib);
467 XBT_DEBUG("copy lib %s to %s, with size %lld", libpath.c_str(), target_lib.c_str(), (long long)fdin_size2);
468 smpi_copy_file(libpath, target_lib, fdin_size2);
470 std::string sedcommand = "sed -i -e 's/" + libname + "/" + target_libname + "/g' " + target_executable;
471 int status = system(sedcommand.c_str());
472 xbt_assert(status == 0, "error while applying sed command %s \n", sedcommand.c_str());
477 // Load the copy and resolve the entry point:
478 void* handle = dlopen(target_executable.c_str(), RTLD_LAZY | RTLD_LOCAL | WANT_RTLD_DEEPBIND);
479 int saved_errno = errno;
480 if (not simgrid::config::get_value<bool>("smpi/keep-temps")) {
481 unlink(target_executable.c_str());
482 for (const std::string& target_lib : target_libs)
483 unlink(target_lib.c_str());
485 xbt_assert(handle != nullptr,
486 "dlopen failed: %s (errno: %d -- %s).\nError: Did you compile the program with a SMPI-specific "
487 "compiler (spmicc or friends)?",
488 dlerror(), saved_errno, strerror(saved_errno));
490 smpi_entry_point_type entry_point = smpi_resolve_function(handle);
491 xbt_assert(entry_point, "Could not resolve entry point. Does your program contain a main() function?");
492 smpi_run_entry_point(entry_point, executable, args);
497 static void smpi_init_privatization_no_dlopen(const std::string& executable)
499 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP)
500 smpi_prepare_global_memory_segment();
502 // Load the dynamic library and resolve the entry point:
503 void* handle = dlopen(executable.c_str(), RTLD_LAZY | RTLD_LOCAL);
504 xbt_assert(handle != nullptr, "dlopen failed for %s: %s (errno: %d -- %s)", executable.c_str(), dlerror(), errno,
506 smpi_entry_point_type entry_point = smpi_resolve_function(handle);
507 xbt_assert(entry_point, "main not found in %s", executable.c_str());
509 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP)
510 smpi_backup_global_memory_segment();
512 // Execute the same entry point for each simulated process:
513 simgrid::s4u::Engine::get_instance()->register_default([entry_point, executable](std::vector<std::string> args) {
514 return std::function<void()>(
515 [entry_point, executable, args] { smpi_run_entry_point(entry_point, executable, args); });
519 int smpi_main(const char* executable, int argc, char* argv[])
521 if (getenv("SMPI_PRETEND_CC") != nullptr) {
522 /* Hack to ensure that smpicc can pretend to be a simple compiler. Particularly handy to pass it to the
523 * configuration tools */
527 SMPI_switch_data_segment = &smpi_switch_data_segment;
529 simgrid::instr::init();
530 SIMIX_global_init(&argc, argv);
532 auto engine = simgrid::s4u::Engine::get_instance();
534 sg_storage_file_system_init();
535 // parse the platform file: get the host list
536 engine->load_platform(argv[1]);
537 simgrid::kernel::activity::CommImpl::set_copy_data_callback(smpi_comm_copy_buffer_callback);
539 if (smpi_cfg_privatization() == SmpiPrivStrategies::DLOPEN)
540 smpi_init_privatization_dlopen(executable);
542 smpi_init_privatization_no_dlopen(executable);
544 simgrid::smpi::colls::set_collectives();
545 simgrid::smpi::colls::smpi_coll_cleanup_callback = nullptr;
549 /* This is a ... heavy way to count the MPI ranks */
551 simgrid::s4u::Actor::on_creation.connect([&rank_counts](const simgrid::s4u::Actor& actor) {
552 if (not actor.is_daemon())
555 engine->load_deployment(argv[2]);
557 SMPI_app_instance_register(smpi_default_instance_name.c_str(), nullptr, rank_counts);
558 MPI_COMM_WORLD = *smpi_deployment_comm_world(smpi_default_instance_name);
560 /* Clean IO before the run */
564 if (MC_is_active()) {
569 xbt_os_walltimer_stop(global_timer);
570 simgrid::smpi::utils::print_time_analysis(xbt_os_timer_elapsed(global_timer));
574 return smpi_exit_status;
577 // Called either directly from the user code, or from the code called by smpirun
580 simgrid::s4u::Actor::on_creation.connect([](simgrid::s4u::Actor& actor) {
581 if (not actor.is_daemon())
582 actor.extension_set<simgrid::smpi::ActorExt>(new simgrid::smpi::ActorExt(&actor));
584 simgrid::s4u::Host::on_creation.connect(
585 [](simgrid::s4u::Host& host) { host.extension_set(new simgrid::smpi::Host(&host)); });
586 for (auto const& host : simgrid::s4u::Engine::get_instance()->get_all_hosts())
587 host->extension_set(new simgrid::smpi::Host(host));
589 if (not MC_is_active()) {
590 global_timer = xbt_os_timer_new();
591 xbt_os_walltimer_start(global_timer);
594 smpi_check_options();
599 smpi_bench_destroy();
600 smpi_shared_destroy();
601 smpi_deployment_cleanup_instances();
603 if (simgrid::smpi::colls::smpi_coll_cleanup_callback != nullptr)
604 simgrid::smpi::colls::smpi_coll_cleanup_callback();
606 MPI_COMM_WORLD = MPI_COMM_NULL;
608 if (not MC_is_active()) {
609 xbt_os_timer_free(global_timer);
612 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP)
613 smpi_destroy_global_memory_segments();
615 simgrid::smpi::utils::print_memory_analysis();
618 void smpi_mpi_init() {
619 smpi_init_fortran_types();
620 if(smpi_init_sleep > 0)
621 simgrid::s4u::this_actor::sleep_for(smpi_init_sleep);
624 void SMPI_thread_create() {
625 TRACE_smpi_init(simgrid::s4u::this_actor::get_pid(), __func__);
626 smpi_process()->mark_as_initialized();