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 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_switch_data_segment(comm->src_actor_->get_iface(), buff)) {
188 XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
189 tmpbuff = xbt_malloc(buff_size);
190 memcpy_private(tmpbuff, buff, private_blocks);
193 if (smpi_switch_data_segment(comm->dst_actor_->get_iface(), comm->dst_buff_))
194 XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
196 XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff, comm->dst_buff_);
197 memcpy_private(comm->dst_buff_, tmpbuff, private_blocks);
199 smpi_cleanup_comm_after_copy(comm,buff);
204 void smpi_comm_null_copy_buffer_callback(simgrid::kernel::activity::CommImpl*, void*, size_t)
206 /* nothing done in this version */
210 return MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED;
213 static void smpi_init_papi()
216 // This map holds for each computation unit (such as "default" or "process1" etc.)
217 // the configuration as given by the user (counter data as a pair of (counter_name, counter_counter))
218 // and the (computed) event_set.
220 if (not smpi_cfg_papi_events_file().empty()) {
221 if (PAPI_library_init(PAPI_VER_CURRENT) != PAPI_VER_CURRENT)
222 XBT_ERROR("Could not initialize PAPI library; is it correctly installed and linked?"
223 " Expected version is %u", PAPI_VER_CURRENT);
225 using Tokenizer = boost::tokenizer<boost::char_separator<char>>;
226 boost::char_separator<char> separator_units(";");
227 std::string str = smpi_cfg_papi_events_file();
228 Tokenizer tokens(str, separator_units);
230 // Iterate over all the computational units. This could be processes, hosts, threads, ranks... You name it.
231 // I'm not exactly sure what we will support eventually, so I'll leave it at the general term "units".
232 for (auto const& unit_it : tokens) {
233 boost::char_separator<char> separator_events(":");
234 Tokenizer event_tokens(unit_it, separator_events);
236 int event_set = PAPI_NULL;
237 if (PAPI_create_eventset(&event_set) != PAPI_OK) {
238 // TODO: Should this let the whole simulation die?
239 XBT_CRITICAL("Could not create PAPI event set during init.");
242 // NOTE: We cannot use a map here, as we must obey the order of the counters
243 // This is important for PAPI: We need to map the values of counters back to the event_names (so, when PAPI_read()
245 papi_counter_t counters2values;
247 // Iterate over all counters that were specified for this specific unit.
248 // Note that we need to remove the name of the unit (that could also be the "default" value), which always comes
249 // first. Hence, we start at ++(events.begin())!
250 for (Tokenizer::iterator events_it = ++(event_tokens.begin()); events_it != event_tokens.end(); ++events_it) {
251 int event_code = PAPI_NULL;
252 auto* event_name = const_cast<char*>((*events_it).c_str());
253 if (PAPI_event_name_to_code(event_name, &event_code) != PAPI_OK) {
254 XBT_CRITICAL("Could not find PAPI event '%s'. Skipping.", event_name);
257 if (PAPI_add_event(event_set, event_code) != PAPI_OK) {
258 XBT_ERROR("Could not add PAPI event '%s'. Skipping.", event_name);
261 XBT_DEBUG("Successfully added PAPI event '%s' to the event set.", event_name);
263 counters2values.emplace_back(*events_it, 0LL);
266 std::string unit_name = *(event_tokens.begin());
267 papi_process_data config = {.counter_data = std::move(counters2values), .event_set = event_set};
269 units2papi_setup.insert(std::make_pair(unit_name, std::move(config)));
275 using smpi_entry_point_type = std::function<int(int argc, char* argv[])>;
276 using smpi_c_entry_point_type = int (*)(int argc, char** argv);
277 using smpi_fortran_entry_point_type = void (*)();
279 template <typename F>
280 static int smpi_run_entry_point(const F& entry_point, const std::string& executable_path, std::vector<std::string> args)
282 // copy C strings, we need them writable
283 auto* args4argv = new std::vector<char*>(args.size());
284 std::transform(begin(args), end(args), begin(*args4argv), [](const std::string& s) { return xbt_strdup(s.c_str()); });
286 // set argv[0] to executable_path
287 xbt_free((*args4argv)[0]);
288 (*args4argv)[0] = xbt_strdup(executable_path.c_str());
291 // take a copy of args4argv to keep reference of the allocated strings
292 const std::vector<char*> args2str(*args4argv);
294 int argc = args4argv->size();
295 args4argv->push_back(nullptr);
296 char** argv = args4argv->data();
299 for_rtl_init_ (&argc, argv);
304 _gfortran_set_args(argc, argv);
306 int res = entry_point(argc, argv);
311 for (char* s : args2str)
317 XBT_WARN("SMPI process did not return 0. Return value : %d", res);
318 if (smpi_exit_status == 0)
319 smpi_exit_status = res;
325 // TODO, remove the number of functions involved here
326 static smpi_entry_point_type smpi_resolve_function(void* handle)
328 auto* entry_point_fortran = reinterpret_cast<smpi_fortran_entry_point_type>(dlsym(handle, "user_main_"));
329 if (entry_point_fortran != nullptr) {
330 return [entry_point_fortran](int, char**) {
331 entry_point_fortran();
336 auto* entry_point = reinterpret_cast<smpi_c_entry_point_type>(dlsym(handle, "main"));
337 if (entry_point != nullptr) {
341 return smpi_entry_point_type();
344 static void smpi_copy_file(const std::string& src, const std::string& target, off_t fdin_size)
346 int fdin = open(src.c_str(), O_RDONLY);
347 xbt_assert(fdin >= 0, "Cannot read from %s. Please make sure that the file exists and is executable.", src.c_str());
348 xbt_assert(unlink(target.c_str()) == 0 || errno == ENOENT, "Failed to unlink file %s: %s", target.c_str(),
350 int fdout = open(target.c_str(), O_CREAT | O_RDWR | O_EXCL, S_IRWXU);
351 xbt_assert(fdout >= 0, "Cannot write into %s: %s", target.c_str(), strerror(errno));
353 XBT_DEBUG("Copy %" PRIdMAX " bytes into %s", static_cast<intmax_t>(fdin_size), target.c_str());
355 ssize_t sent_size = sendfile(fdout, fdin, nullptr, fdin_size);
356 if (sent_size == fdin_size) {
361 xbt_assert(sent_size == -1 && errno == ENOSYS,
362 "Error while copying %s: only %zd bytes copied instead of %" PRIdMAX " (errno: %d -- %s)", target.c_str(),
363 sent_size, static_cast<intmax_t>(fdin_size), errno, strerror(errno));
365 // If this point is reached, sendfile() actually is not available. Copy file by hand.
366 std::vector<unsigned char> buf(1024 * 1024 * 4);
367 while (ssize_t got = read(fdin, buf.data(), buf.size())) {
369 xbt_assert(errno == EINTR, "Cannot read from %s", src.c_str());
371 const unsigned char* p = buf.data();
373 while (ssize_t done = write(fdout, p, todo)) {
375 xbt_assert(errno == EINTR, "Cannot write into %s", target.c_str());
387 #if not defined(__APPLE__) && not defined(__HAIKU__)
388 static int visit_libs(struct dl_phdr_info* info, size_t, void* data)
390 auto* libname = static_cast<std::string*>(data);
391 std::string path = info->dlpi_name;
392 if (path.find(*libname) != std::string::npos) {
393 *libname = std::move(path);
400 static void smpi_init_privatization_dlopen(const std::string& executable)
402 // Prepare the copy of the binary (get its size)
403 struct stat fdin_stat;
404 stat(executable.c_str(), &fdin_stat);
405 off_t fdin_size = fdin_stat.st_size;
407 std::string libnames = simgrid::config::get_value<std::string>("smpi/privatize-libs");
408 if (not libnames.empty()) {
410 std::vector<std::string> privatize_libs;
411 boost::split(privatize_libs, libnames, boost::is_any_of(";"));
413 for (auto const& libname : privatize_libs) {
414 // load the library once to add it to the local libs, to get the absolute path
415 void* libhandle = dlopen(libname.c_str(), RTLD_LAZY);
416 xbt_assert(libhandle != nullptr, "Cannot dlopen %s - check your settings in smpi/privatize-libs",
418 // get library name from path
419 std::string fullpath = libname;
420 #if not defined(__APPLE__) && not defined(__HAIKU__)
421 xbt_assert(dl_iterate_phdr(visit_libs, &fullpath) != 0,
422 "Can't find a linked %s - check your settings in smpi/privatize-libs", fullpath.c_str());
423 XBT_DEBUG("Extra lib to privatize '%s' found", fullpath.c_str());
425 xbt_die("smpi/privatize-libs is not (yet) compatible with OSX nor with Haiku");
427 privatize_libs_paths.emplace_back(std::move(fullpath));
432 simgrid::s4u::Engine::get_instance()->register_default([executable, fdin_size](std::vector<std::string> args) {
433 return std::function<void()>([executable, fdin_size, args] {
434 static std::size_t rank = 0;
435 // Copy the dynamic library:
436 simgrid::xbt::Path path(executable);
437 std::string target_executable = simgrid::config::get_value<std::string>("smpi/tmpdir") + "/" +
438 path.get_base_name() + "_" + std::to_string(getpid()) + "_" + std::to_string(rank) + ".so";
440 smpi_copy_file(executable, target_executable, fdin_size);
441 // if smpi/privatize-libs is set, duplicate pointed lib and link each executable copy to a different one.
442 std::vector<std::string> target_libs;
443 for (auto const& libpath : privatize_libs_paths) {
444 // if we were given a full path, strip it
445 size_t index = libpath.find_last_of("/\\");
447 if (index != std::string::npos)
448 libname = libpath.substr(index + 1);
450 if (not libname.empty()) {
451 // load the library to add it to the local libs, to get the absolute path
452 struct stat fdin_stat2;
453 stat(libpath.c_str(), &fdin_stat2);
454 off_t fdin_size2 = fdin_stat2.st_size;
456 // Copy the dynamic library, the new name must be the same length as the old one
457 // just replace the name with 7 digits for the rank and the rest of the name.
458 auto pad = std::min<unsigned>(7, libname.length());
459 std::string target_libname = std::string(pad - std::to_string(rank).length(), '0') + std::to_string(rank) + libname.substr(pad);
460 std::string target_lib = simgrid::config::get_value<std::string>("smpi/tmpdir") + "/" + target_libname;
461 target_libs.push_back(target_lib);
462 XBT_DEBUG("copy lib %s to %s, with size %lld", libpath.c_str(), target_lib.c_str(), (long long)fdin_size2);
463 smpi_copy_file(libpath, target_lib, fdin_size2);
465 std::string sedcommand = "sed -i -e 's/" + libname + "/" + target_libname + "/g' " + target_executable;
466 int status = system(sedcommand.c_str());
467 xbt_assert(status == 0, "error while applying sed command %s \n", sedcommand.c_str());
472 // Load the copy and resolve the entry point:
473 void* handle = dlopen(target_executable.c_str(), RTLD_LAZY | RTLD_LOCAL | WANT_RTLD_DEEPBIND);
474 int saved_errno = errno;
475 if (not simgrid::config::get_value<bool>("smpi/keep-temps")) {
476 unlink(target_executable.c_str());
477 for (const std::string& target_lib : target_libs)
478 unlink(target_lib.c_str());
480 xbt_assert(handle != nullptr,
481 "dlopen failed: %s (errno: %d -- %s).\nError: Did you compile the program with a SMPI-specific "
482 "compiler (spmicc or friends)?",
483 dlerror(), saved_errno, strerror(saved_errno));
485 smpi_entry_point_type entry_point = smpi_resolve_function(handle);
486 xbt_assert(entry_point, "Could not resolve entry point. Does your program contain a main() function?");
487 smpi_run_entry_point(entry_point, executable, args);
492 static void smpi_init_privatization_no_dlopen(const std::string& executable)
494 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP)
495 smpi_prepare_global_memory_segment();
497 // Load the dynamic library and resolve the entry point:
498 void* handle = dlopen(executable.c_str(), RTLD_LAZY | RTLD_LOCAL);
499 xbt_assert(handle != nullptr, "dlopen failed for %s: %s (errno: %d -- %s)", executable.c_str(), dlerror(), errno,
501 smpi_entry_point_type entry_point = smpi_resolve_function(handle);
502 xbt_assert(entry_point, "main not found in %s", executable.c_str());
504 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP)
505 smpi_backup_global_memory_segment();
507 // Execute the same entry point for each simulated process:
508 simgrid::s4u::Engine::get_instance()->register_default([entry_point, executable](std::vector<std::string> args) {
509 return std::function<void()>(
510 [entry_point, executable, args] { smpi_run_entry_point(entry_point, executable, args); });
514 int smpi_main(const char* executable, int argc, char* argv[])
516 if (getenv("SMPI_PRETEND_CC") != nullptr) {
517 /* Hack to ensure that smpicc can pretend to be a simple compiler. Particularly handy to pass it to the
518 * configuration tools */
522 smpi_init_options(true);
523 simgrid::instr::init();
524 SIMIX_global_init(&argc, argv);
526 auto engine = simgrid::s4u::Engine::get_instance();
528 sg_storage_file_system_init();
529 // parse the platform file: get the host list
530 engine->load_platform(argv[1]);
531 simgrid::kernel::activity::CommImpl::set_copy_data_callback(smpi_comm_copy_buffer_callback);
533 if (smpi_cfg_privatization() == SmpiPrivStrategies::DLOPEN)
534 smpi_init_privatization_dlopen(executable);
536 smpi_init_privatization_no_dlopen(executable);
538 simgrid::smpi::colls::set_collectives();
539 simgrid::smpi::colls::smpi_coll_cleanup_callback = nullptr;
543 /* This is a ... heavy way to count the MPI ranks */
545 simgrid::s4u::Actor::on_creation.connect([&rank_counts](const simgrid::s4u::Actor& actor) {
546 if (not actor.is_daemon())
549 engine->load_deployment(argv[2]);
551 SMPI_app_instance_register(smpi_default_instance_name.c_str(), nullptr, rank_counts);
552 MPI_COMM_WORLD = *smpi_deployment_comm_world(smpi_default_instance_name);
554 /* Clean IO before the run */
558 if (MC_is_active()) {
561 simgrid::kernel::EngineImpl::get_instance()->run();
563 xbt_os_walltimer_stop(global_timer);
564 simgrid::smpi::utils::print_time_analysis(xbt_os_timer_elapsed(global_timer));
568 return smpi_exit_status;
571 // Called either directly from the user code, or from the code called by smpirun
574 simgrid::s4u::Actor::on_creation.connect([](simgrid::s4u::Actor& actor) {
575 if (not actor.is_daemon())
576 actor.extension_set<simgrid::smpi::ActorExt>(new simgrid::smpi::ActorExt(&actor));
578 simgrid::s4u::Host::on_creation.connect(
579 [](simgrid::s4u::Host& host) { host.extension_set(new simgrid::smpi::Host(&host)); });
580 for (auto const& host : simgrid::s4u::Engine::get_instance()->get_all_hosts())
581 host->extension_set(new simgrid::smpi::Host(host));
583 if (not MC_is_active()) {
584 global_timer = xbt_os_timer_new();
585 xbt_os_walltimer_start(global_timer);
588 smpi_check_options();
593 smpi_bench_destroy();
594 smpi_shared_destroy();
595 smpi_deployment_cleanup_instances();
597 if (simgrid::smpi::colls::smpi_coll_cleanup_callback != nullptr)
598 simgrid::smpi::colls::smpi_coll_cleanup_callback();
600 MPI_COMM_WORLD = MPI_COMM_NULL;
602 if (not MC_is_active()) {
603 xbt_os_timer_free(global_timer);
606 if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP)
607 smpi_destroy_global_memory_segments();
609 simgrid::smpi::utils::print_memory_analysis();
612 void smpi_mpi_init() {
613 smpi_init_fortran_types();
614 if(smpi_init_sleep > 0)
615 simgrid::s4u::this_actor::sleep_for(smpi_init_sleep);
618 void SMPI_thread_create() {
619 TRACE_smpi_init(simgrid::s4u::this_actor::get_pid(), __func__);
620 smpi_process()->mark_as_initialized();