-/* Copyright (c) 2007-2018. The SimGrid Team. All rights reserved. */
+/* Copyright (c) 2007-2019. The SimGrid Team. All rights reserved. */
/* This program is free software; you can redistribute it and/or modify it
* under the terms of the license (GNU LGPL) which comes with this package. */
#include "src/smpi/include/smpi_actor.hpp"
#include "xbt/config.hpp"
+#include <algorithm>
#include <cfloat> /* DBL_MAX */
+#include <cinttypes>
+#include <cstdint> /* intmax_t */
#include <dlfcn.h>
#include <fcntl.h>
#include <fstream>
# ifndef MAC_OS_X_VERSION_10_12
# define MAC_OS_X_VERSION_10_12 101200
# endif
-# define HAVE_WORKING_MMAP (MAC_OS_X_VERSION_MIN_REQUIRED >= MAC_OS_X_VERSION_10_12)
-#elif defined(__FreeBSD__)
-# define HAVE_WORKING_MMAP 0
+constexpr bool HAVE_WORKING_MMAP = (MAC_OS_X_VERSION_MIN_REQUIRED >= MAC_OS_X_VERSION_10_12);
+#elif defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
+constexpr bool HAVE_WORKING_MMAP = false;
#else
-# define HAVE_WORKING_MMAP 1
+constexpr bool HAVE_WORKING_MMAP = true;
#endif
-#if HAVE_SENDFILE
+#if SG_HAVE_SENDFILE
#include <sys/sendfile.h>
#endif
#if SMPI_IFORT
extern "C" void for_rtl_init_ (int *, char **);
+ extern "C" void for_rtl_finish_ ();
#elif SMPI_FLANG
extern "C" void __io_set_argc(int);
extern "C" void __io_set_argv(char **);
extern "C" void _gfortran_set_args(int, char **);
#endif
-#ifndef RTLD_DEEPBIND
/* RTLD_DEEPBIND is a bad idea of GNU ld that obviously does not exist on other platforms
* See https://www.akkadia.org/drepper/dsohowto.pdf
* and https://lists.freebsd.org/pipermail/freebsd-current/2016-March/060284.html
*/
-#define RTLD_DEEPBIND 0
+#if !RTLD_DEEPBIND || HAVE_SANITIZER_ADDRESS || HAVE_SANITIZER_THREAD
+#define WANT_RTLD_DEEPBIND 0
+#else
+#define WANT_RTLD_DEEPBIND RTLD_DEEPBIND
#endif
#if HAVE_PAPI
std::map</* computation unit name */ std::string, papi_process_data> units2papi_setup;
#endif
-using simgrid::s4u::Actor;
-using simgrid::s4u::ActorPtr;
std::unordered_map<std::string, double> location2speedup;
-static std::map</*process_id*/ ActorPtr, simgrid::smpi::ActorExt*> process_data;
+static std::map</*process_id*/ simgrid::s4u::Actor const*, simgrid::smpi::ActorExt*> process_data;
int process_count = 0;
static int smpi_exit_status = 0;
int smpi_universe_size = 0;
// No instance gets manually created; check also the smpirun.in script as
// this default name is used there as well (when the <actor> tag is generated).
static const std::string smpi_default_instance_name("smpirun");
-static simgrid::config::Flag<double> smpi_wtime_sleep(
- "smpi/wtime", "Minimum time to inject inside a call to MPI_Wtime", 0.0);
static simgrid::config::Flag<double> smpi_init_sleep(
"smpi/init", "Time to inject inside a call to MPI_Init", 0.0);
-void (*smpi_comm_copy_data_callback) (smx_activity_t, void*, size_t) = &smpi_comm_copy_buffer_callback;
+void (*smpi_comm_copy_data_callback)(simgrid::kernel::activity::CommImpl*, void*,
+ size_t) = &smpi_comm_copy_buffer_callback;
int smpi_process_count()
{
simgrid::smpi::ActorExt* smpi_process()
{
- ActorPtr me = Actor::self();
+ simgrid::s4u::ActorPtr me = simgrid::s4u::Actor::self();
if (me == nullptr) // This happens sometimes (eg, when linking against NS3 because it pulls openMPI...)
return nullptr;
- return process_data.at(me);
+ return process_data.at(me.get());
}
-simgrid::smpi::ActorExt* smpi_process_remote(ActorPtr actor)
+simgrid::smpi::ActorExt* smpi_process_remote(simgrid::s4u::ActorPtr actor)
{
- return process_data.at(actor);
+ return process_data.at(actor.get());
}
MPI_Comm smpi_process_comm_self(){
}
void smpi_process_init(int *argc, char ***argv){
- simgrid::smpi::ActorExt::init(argc, argv);
-}
-
-int smpi_process_index(){
- return simgrid::s4u::this_actor::get_pid();
+ simgrid::smpi::ActorExt::init();
}
void * smpi_process_get_user_data(){
- return Actor::self()->get_impl()->get_user_data();
+ return simgrid::s4u::Actor::self()->get_impl()->get_user_data();
}
void smpi_process_set_user_data(void *data){
- Actor::self()->get_impl()->set_user_data(data);
-}
-
-
-int smpi_global_size()
-{
- char *value = getenv("SMPI_GLOBAL_SIZE");
- xbt_assert(value,"Please set env var SMPI_GLOBAL_SIZE to the expected number of processes.");
-
- return xbt_str_parse_int(value, "SMPI_GLOBAL_SIZE contains a non-numerical value: %s");
+ simgrid::s4u::Actor::self()->get_impl()->set_user_data(data);
}
void smpi_comm_set_copy_data_callback(void (*callback) (smx_activity_t, void*, size_t))
{
- smpi_comm_copy_data_callback = callback;
+ static void (*saved_callback)(smx_activity_t, void*, size_t);
+ saved_callback = callback;
+ smpi_comm_copy_data_callback = [](simgrid::kernel::activity::CommImpl* comm, void* buff, size_t size) {
+ saved_callback(smx_activity_t(comm), buff, size);
+ };
}
static void memcpy_private(void* dest, const void* src, std::vector<std::pair<size_t, size_t>>& private_blocks)
xbt_assert(block.first <= block.second && block.second <= buff_size, "Oops, bug in shared malloc.");
}
-void smpi_comm_copy_buffer_callback(smx_activity_t synchro, void *buff, size_t buff_size)
+void smpi_comm_copy_buffer_callback(simgrid::kernel::activity::CommImpl* comm, void* buff, size_t buff_size)
{
- simgrid::kernel::activity::CommImplPtr comm =
- boost::dynamic_pointer_cast<simgrid::kernel::activity::CommImpl>(synchro);
int src_shared = 0;
int dst_shared = 0;
size_t src_offset = 0;
src_private_blocks.clear();
src_private_blocks.push_back(std::make_pair(0, buff_size));
}
- if((dst_shared=smpi_is_shared((char*)comm->dst_buff, dst_private_blocks, &dst_offset))) {
- XBT_DEBUG("Receiver %p is shared. Let's ignore it.", (char*)comm->dst_buff);
+ if ((dst_shared = smpi_is_shared((char*)comm->dst_buff_, dst_private_blocks, &dst_offset))) {
+ XBT_DEBUG("Receiver %p is shared. Let's ignore it.", (char*)comm->dst_buff_);
dst_private_blocks = shift_and_frame_private_blocks(dst_private_blocks, dst_offset, buff_size);
}
else {
(static_cast<char*>(buff) >= smpi_data_exe_start) &&
(static_cast<char*>(buff) < smpi_data_exe_start + smpi_data_exe_size)) {
XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
- smpi_switch_data_segment(comm->src_proc->iface());
+ smpi_switch_data_segment(comm->src_actor_->iface());
tmpbuff = static_cast<void*>(xbt_malloc(buff_size));
memcpy_private(tmpbuff, buff, private_blocks);
}
- if ((smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) && ((char*)comm->dst_buff >= smpi_data_exe_start) &&
- ((char*)comm->dst_buff < smpi_data_exe_start + smpi_data_exe_size)) {
+ if ((smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) &&
+ ((char*)comm->dst_buff_ >= smpi_data_exe_start) &&
+ ((char*)comm->dst_buff_ < smpi_data_exe_start + smpi_data_exe_size)) {
XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
- smpi_switch_data_segment(comm->dst_proc->iface());
+ smpi_switch_data_segment(comm->dst_actor_->iface());
}
- XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff,comm->dst_buff);
- memcpy_private(comm->dst_buff, tmpbuff, private_blocks);
+ XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff, comm->dst_buff_);
+ memcpy_private(comm->dst_buff_, tmpbuff, private_blocks);
- if (comm->detached) {
+ if (comm->detached_) {
// if this is a detached send, the source buffer was duplicated by SMPI
// sender to make the original buffer available to the application ASAP
xbt_free(buff);
//It seems that the request is used after the call there this should be free somewhere else but where???
//xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free
- comm->src_buff = nullptr;
+ comm->src_buff_ = nullptr;
}
if (tmpbuff != buff)
xbt_free(tmpbuff);
}
-void smpi_comm_null_copy_buffer_callback(smx_activity_t comm, void *buff, size_t buff_size)
+void smpi_comm_null_copy_buffer_callback(simgrid::kernel::activity::CommImpl*, void*, size_t)
{
/* nothing done in this version */
}
int event_code = PAPI_NULL;
char* event_name = const_cast<char*>((*events_it).c_str());
- if (PAPI_event_name_to_code(event_name, &event_code) == PAPI_OK) {
- if (PAPI_add_event(event_set, event_code) != PAPI_OK) {
- XBT_ERROR("Could not add PAPI event '%s'. Skipping.", event_name);
- continue;
- } else {
- XBT_DEBUG("Successfully added PAPI event '%s' to the event set.", event_name);
- }
- } else {
+ if (PAPI_event_name_to_code(event_name, &event_code) != PAPI_OK) {
XBT_CRITICAL("Could not find PAPI event '%s'. Skipping.", event_name);
continue;
}
+ if (PAPI_add_event(event_set, event_code) != PAPI_OK) {
+ XBT_ERROR("Could not add PAPI event '%s'. Skipping.", event_name);
+ continue;
+ }
+ XBT_DEBUG("Successfully added PAPI event '%s' to the event set.", event_name);
counters2values.push_back(
// We cannot just pass *events_it, as this is of type const basic_string
if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP)
smpi_destroy_global_memory_segments();
- smpi_free_static();
if(simgrid::smpi::F2C::lookup() != nullptr)
simgrid::smpi::F2C::delete_lookup();
}
XBT_DEBUG("Running without smpi_main(); disable smpi/privatization.");
smpi_privatize_global_variables = SmpiPrivStrategies::NONE;
}
-#if !HAVE_WORKING_MMAP
- if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) {
+ if (not HAVE_WORKING_MMAP && smpi_privatize_global_variables == SmpiPrivStrategies::MMAP) {
XBT_INFO("mmap privatization is broken on this platform, switching to dlopen privatization instead.");
smpi_privatize_global_variables = SmpiPrivStrategies::DLOPEN;
}
-#endif
if (smpi_cpu_threshold < 0)
smpi_cpu_threshold = DBL_MAX;
typedef int (* smpi_c_entry_point_type)(int argc, char **argv);
typedef void (*smpi_fortran_entry_point_type)();
-static int smpi_run_entry_point(smpi_entry_point_type entry_point, std::vector<std::string> args)
+static int smpi_run_entry_point(const smpi_entry_point_type& entry_point, const std::string& executable_path,
+ std::vector<std::string> args)
{
- char noarg[] = {'\0'};
- int argc = args.size();
- std::unique_ptr<char*[]> argv(new char*[argc + 1]);
- for (int i = 0; i != argc; ++i)
- argv[i] = args[i].empty() ? noarg : &args[i].front();
- argv[argc] = nullptr;
- char ** argvptr=argv.get();
- simgrid::smpi::ActorExt::init(&argc, &argvptr);
+ // copy C strings, we need them writable
+ std::vector<char*>* args4argv = new std::vector<char*>(args.size());
+ std::transform(begin(args), end(args), begin(*args4argv), [](const std::string& s) { return xbt_strdup(s.c_str()); });
+
+ // set argv[0] to executable_path
+ xbt_free((*args4argv)[0]);
+ (*args4argv)[0] = xbt_strdup(executable_path.c_str());
+
+#if !SMPI_IFORT
+ // take a copy of args4argv to keep reference of the allocated strings
+ const std::vector<char*> args2str(*args4argv);
+#endif
+ int argc = args4argv->size();
+ args4argv->push_back(nullptr);
+ char** argv = args4argv->data();
+
+ simgrid::smpi::ActorExt::init();
#if SMPI_IFORT
- for_rtl_init_ (&argc, argvptr);
+ for_rtl_init_ (&argc, argv);
#elif SMPI_FLANG
__io_set_argc(argc);
- __io_set_argv(argvptr);
+ __io_set_argv(argv);
#elif SMPI_GFORTRAN
- _gfortran_set_args(argc, argvptr);
+ _gfortran_set_args(argc, argv);
#endif
- int res = entry_point(argc, argvptr);
+ int res = entry_point(argc, argv);
+
+#if SMPI_IFORT
+ for_rtl_finish_ ();
+#else
+ for (char* s : args2str)
+ xbt_free(s);
+ delete args4argv;
+#endif
+
if (res != 0){
XBT_WARN("SMPI process did not return 0. Return value : %d", res);
if (smpi_exit_status == 0)
return smpi_entry_point_type();
}
-static void smpi_copy_file(std::string src, std::string target, off_t fdin_size)
+static void smpi_copy_file(const std::string& src, const std::string& target, off_t fdin_size)
{
int fdin = open(src.c_str(), O_RDONLY);
xbt_assert(fdin >= 0, "Cannot read from %s. Please make sure that the file exists and is executable.", src.c_str());
int fdout = open(target.c_str(), O_CREAT | O_RDWR, S_IRWXU);
xbt_assert(fdout >= 0, "Cannot write into %s", target.c_str());
- XBT_DEBUG("Copy %ld bytes into %s", static_cast<long>(fdin_size), target.c_str());
-#if HAVE_SENDFILE
+ XBT_DEBUG("Copy %" PRIdMAX " bytes into %s", static_cast<intmax_t>(fdin_size), target.c_str());
+#if SG_HAVE_SENDFILE
ssize_t sent_size = sendfile(fdout, fdin, NULL, fdin_size);
- xbt_assert(sent_size == fdin_size, "Error while copying %s: only %zd bytes copied instead of %ld (errno: %d -- %s)",
- target.c_str(), sent_size, fdin_size, errno, strerror(errno));
-#else
+ if (sent_size == fdin_size) {
+ close(fdin);
+ close(fdout);
+ return;
+ } else if (sent_size != -1 || errno != ENOSYS) {
+ xbt_die("Error while copying %s: only %zd bytes copied instead of %" PRIdMAX " (errno: %d -- %s)", target.c_str(),
+ sent_size, static_cast<intmax_t>(fdin_size), errno, strerror(errno));
+ }
+#endif
+ // If this point is reached, sendfile() actually is not available. Copy file by hand.
const int bufsize = 1024 * 1024 * 4;
char buf[bufsize];
while (int got = read(fdin, buf, bufsize)) {
}
}
}
-#endif
close(fdin);
close(fdout);
}
}
#endif
-static void smpi_init_privatization_dlopen(std::string executable)
+static void smpi_init_privatization_dlopen(const std::string& executable)
{
// Prepare the copy of the binary (get its size)
struct stat fdin_stat;
stat(executable.c_str(), &fdin_stat);
off_t fdin_size = fdin_stat.st_size;
- static std::size_t rank = 0;
std::string libnames = simgrid::config::get_value<std::string>("smpi/privatize-libs");
if (not libnames.empty()) {
void* libhandle = dlopen(libname.c_str(), RTLD_LAZY);
// get library name from path
char fullpath[512] = {'\0'};
- strcpy(fullpath, libname.c_str());
+ strncpy(fullpath, libname.c_str(), 511);
#if not defined(__APPLE__)
int ret = dl_iterate_phdr(visit_libs, fullpath);
if (ret == 0)
simix_global->default_function = [executable, fdin_size](std::vector<std::string> args) {
return std::function<void()>([executable, fdin_size, args] {
-
+ static std::size_t rank = 0;
// Copy the dynamic library:
std::string target_executable =
executable + "_" + std::to_string(getpid()) + "_" + std::to_string(rank) + ".so";
smpi_copy_file(executable, target_executable, fdin_size);
// if smpi/privatize-libs is set, duplicate pointed lib and link each executable copy to a different one.
- std::string target_lib;
+ std::vector<std::string> target_libs;
for (auto const& libpath : privatize_libs_paths) {
// if we were given a full path, strip it
size_t index = libpath.find_last_of("/\\");
unsigned int pad = 7;
if (libname.length() < pad)
pad = libname.length();
- target_lib =
+ std::string target_lib =
std::string(pad - std::to_string(rank).length(), '0') + std::to_string(rank) + libname.substr(pad);
+ target_libs.push_back(target_lib);
XBT_DEBUG("copy lib %s to %s, with size %lld", libpath.c_str(), target_lib.c_str(), (long long)fdin_size2);
smpi_copy_file(libpath, target_lib, fdin_size2);
rank++;
// Load the copy and resolve the entry point:
- void* handle = dlopen(target_executable.c_str(), RTLD_LAZY | RTLD_LOCAL | RTLD_DEEPBIND);
+ void* handle = dlopen(target_executable.c_str(), RTLD_LAZY | RTLD_LOCAL | WANT_RTLD_DEEPBIND);
int saved_errno = errno;
if (simgrid::config::get_value<bool>("smpi/keep-temps") == false) {
unlink(target_executable.c_str());
- if (not target_lib.empty())
+ for (const std::string& target_lib : target_libs)
unlink(target_lib.c_str());
}
if (handle == nullptr)
smpi_entry_point_type entry_point = smpi_resolve_function(handle);
if (not entry_point)
xbt_die("Could not resolve entry point");
- smpi_run_entry_point(entry_point, args);
+ smpi_run_entry_point(entry_point, executable, args);
});
};
}
-static void smpi_init_privatization_no_dlopen(std::string executable)
+static void smpi_init_privatization_no_dlopen(const std::string& executable)
{
if (smpi_privatize_global_variables == SmpiPrivStrategies::MMAP)
smpi_prepare_global_memory_segment();
smpi_backup_global_memory_segment();
// Execute the same entry point for each simulated process:
- simix_global->default_function = [entry_point](std::vector<std::string> args) {
- return std::function<void()>([entry_point, args] { smpi_run_entry_point(entry_point, args); });
+ simix_global->default_function = [entry_point, executable](std::vector<std::string> args) {
+ return std::function<void()>(
+ [entry_point, executable, args] { smpi_run_entry_point(entry_point, executable, args); });
};
}
int smpi_main(const char* executable, int argc, char* argv[])
{
- srand(SMPI_RAND_SEED);
-
if (getenv("SMPI_PRETEND_CC") != nullptr) {
/* Hack to ensure that smpicc can pretend to be a simple compiler. Particularly handy to pass it to the
* configuration tools */
SMPI_switch_data_segment = &smpi_switch_data_segment;
- // TODO This will not be executed in the case where smpi_main is not called,
- // e.g., not for smpi_msg_masterslave. This should be moved to another location
- // that is always called -- maybe close to Actor::on_creation?
- simgrid::s4u::Host::on_creation.connect(
- [](simgrid::s4u::Host& host) { host.extension_set(new simgrid::smpi::Host(&host)); });
-
// parse the platform file: get the host list
simgrid::s4u::Engine::get_instance()->load_platform(argv[1]);
SIMIX_comm_set_copy_data_callback(smpi_comm_copy_buffer_callback);
// Called either directly from the user code, or from the code called by smpirun
void SMPI_init(){
- simgrid::s4u::Actor::on_creation.connect([](simgrid::s4u::ActorPtr actor) {
- if (not actor->is_daemon()) {
- process_data.insert({actor, new simgrid::smpi::ActorExt(actor, nullptr)});
+ simgrid::s4u::Actor::on_creation.connect([](simgrid::s4u::Actor& actor) {
+ if (not actor.is_daemon()) {
+ process_data.insert({&actor, new simgrid::smpi::ActorExt(&actor, nullptr)});
}
});
- simgrid::s4u::Actor::on_destruction.connect([](simgrid::s4u::ActorPtr actor) {
- auto it = process_data.find(actor);
+ simgrid::s4u::Actor::on_destruction.connect([](simgrid::s4u::Actor const& actor) {
+ auto it = process_data.find(&actor);
if (it != process_data.end()) {
delete it->second;
process_data.erase(it);
}
});
+ simgrid::s4u::Host::on_creation.connect(
+ [](simgrid::s4u::Host& host) { host.extension_set(new simgrid::smpi::Host(&host)); });
smpi_init_options();
smpi_global_init();
if(smpi_init_sleep > 0)
simcall_process_sleep(smpi_init_sleep);
}
-
-double smpi_mpi_wtime(){
- double time;
- if (smpi_process()->initialized() != 0 && smpi_process()->finalized() == 0 && smpi_process()->sampling() == 0) {
- smpi_bench_end();
- time = SIMIX_get_clock();
- // to avoid deadlocks if used as a break condition, such as
- // while (MPI_Wtime(...) < time_limit) {
- // ....
- // }
- // because the time will not normally advance when only calls to MPI_Wtime
- // are made -> deadlock (MPI_Wtime never reaches the time limit)
- if(smpi_wtime_sleep > 0)
- simcall_process_sleep(smpi_wtime_sleep);
- smpi_bench_begin();
- } else {
- time = SIMIX_get_clock();
- }
- return time;
-}
-