1 /* Copyright (c) 2015-2023. 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. */
6 #include "src/mc/remote/AppSide.hpp"
7 #include "simgrid/s4u/Host.hpp"
8 #include "src/internal_config.h"
9 #include "src/kernel/EngineImpl.hpp"
10 #include "src/kernel/actor/ActorImpl.hpp"
11 #include "src/kernel/actor/SimcallObserver.hpp"
12 #include "src/mc/mc_base.hpp"
13 #include "src/mc/mc_config.hpp"
14 #include "src/mc/mc_environ.h"
15 #if SIMGRID_HAVE_STATEFUL_MC
16 #include "src/mc/sosp/RemoteProcessMemory.hpp"
19 #include "src/smpi/include/private.hpp"
21 #include "src/sthread/sthread.h"
22 #include "src/xbt/coverage.h"
24 #include <simgrid/modelchecker.h>
28 #include <cstdio> // setvbuf
32 #include <sys/ptrace.h>
33 #include <sys/socket.h>
34 #include <sys/types.h>
38 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(mc_client, mc, "MC client logic");
39 XBT_LOG_EXTERNAL_CATEGORY(mc_global);
41 namespace simgrid::mc {
43 std::unique_ptr<AppSide> AppSide::instance_;
45 AppSide* AppSide::get()
47 // Only initialize the MC world once
48 if (instance_ != nullptr)
49 return instance_.get();
51 if (std::getenv(MC_ENV_SOCKET_FD) == nullptr) // We are not in MC mode: don't initialize the MC world
54 XBT_DEBUG("Initialize the MC world. %s=%s", MC_ENV_NEED_PTRACE, std::getenv(MC_ENV_NEED_PTRACE));
56 simgrid::mc::set_model_checking_mode(ModelCheckingMode::APP_SIDE);
58 setvbuf(stdout, nullptr, _IOLBF, 0);
60 // Fetch socket from MC_ENV_SOCKET_FD:
61 const char* fd_env = std::getenv(MC_ENV_SOCKET_FD);
62 int fd = xbt_str_parse_int(fd_env, "Not a number in variable '" MC_ENV_SOCKET_FD "'");
63 XBT_DEBUG("Model-checked application found socket FD %i", fd);
65 instance_ = std::make_unique<simgrid::mc::AppSide>(fd);
67 // Wait for the model-checker:
68 if (getenv(MC_ENV_NEED_PTRACE) != nullptr) {
71 ptrace(PTRACE_TRACEME, 0, nullptr, nullptr);
73 ptrace(PT_TRACE_ME, 0, nullptr, 0);
75 xbt_die("no ptrace equivalent coded for this platform, please don't use the liveness checker here.");
78 xbt_assert(errno == 0 && raise(SIGSTOP) == 0, "Could not wait for the model-checker (errno = %d: %s)", errno,
82 instance_->handle_messages();
83 return instance_.get();
86 void AppSide::handle_deadlock_check(const s_mc_message_t*) const
88 const auto* engine = kernel::EngineImpl::get_instance();
89 const auto& actor_list = engine->get_actor_list();
90 bool deadlock = not actor_list.empty() && std::none_of(begin(actor_list), end(actor_list), [](const auto& kv) {
91 return mc::actor_is_enabled(kv.second);
95 XBT_CINFO(mc_global, "**************************");
96 XBT_CINFO(mc_global, "*** DEADLOCK DETECTED ***");
97 XBT_CINFO(mc_global, "**************************");
98 engine->display_all_actor_status();
101 s_mc_message_int_t answer = {};
102 answer.type = MessageType::DEADLOCK_CHECK_REPLY;
103 answer.value = deadlock;
104 xbt_assert(channel_.send(answer) == 0, "Could not send response: %s", strerror(errno));
106 void AppSide::handle_simcall_execute(const s_mc_message_simcall_execute_t* message) const
108 kernel::actor::ActorImpl* actor = kernel::EngineImpl::get_instance()->get_actor_by_pid(message->aid_);
109 xbt_assert(actor != nullptr, "Invalid pid %ld", message->aid_);
111 // The client may send some messages to the server while processing the transition
112 actor->simcall_handle(message->times_considered_);
113 // Say the server that the transition is over and that it should proceed
114 xbt_assert(channel_.send(MessageType::WAITING) == 0, "Could not send MESSAGE_WAITING to model-checker: %s",
117 // Finish the RPC from the server: return a serialized observer, to build a Transition on Checker side
118 s_mc_message_simcall_execute_answer_t answer = {};
119 answer.type = MessageType::SIMCALL_EXECUTE_REPLY;
120 std::stringstream stream;
121 if (actor->simcall_.observer_ != nullptr) {
122 actor->simcall_.observer_->serialize(stream);
124 stream << (short)mc::Transition::Type::UNKNOWN;
126 std::string str = stream.str();
127 xbt_assert(str.size() + 1 <= answer.buffer.size(),
128 "The serialized simcall is too large for the buffer. Please fix the code.");
129 strncpy(answer.buffer.data(), str.c_str(), answer.buffer.size() - 1);
130 answer.buffer.back() = '\0';
132 XBT_DEBUG("send SIMCALL_EXECUTE_ANSWER(%s) ~> '%s'", actor->get_cname(), str.c_str());
133 xbt_assert(channel_.send(answer) == 0, "Could not send response: %s", strerror(errno));
136 void AppSide::handle_finalize(const s_mc_message_int_t* msg) const
138 bool terminate_asap = msg->value;
139 XBT_DEBUG("Finalize (terminate = %d)", (int)terminate_asap);
140 if (not terminate_asap) {
141 if (XBT_LOG_ISENABLED(mc_client, xbt_log_priority_debug))
142 kernel::EngineImpl::get_instance()->display_all_actor_status();
144 XBT_DEBUG("Smpi_enabled: %d", SMPI_is_inited());
145 if (SMPI_is_inited())
149 coverage_checkpoint();
150 xbt_assert(channel_.send(MessageType::FINALIZE_REPLY) == 0, "Could not answer to FINALIZE: %s", strerror(errno));
155 void AppSide::handle_fork(const s_mc_message_int_t* msg)
159 /* Reap any zombie child, saving its status for later use in AppSide::handle_wait_child() */
160 while ((pid = waitpid(-1, &status, WNOHANG)) > 0)
161 child_statuses_[pid] = status;
164 xbt_assert(pid >= 0, "Could not fork application sub-process: %s.", strerror(errno));
166 if (pid == 0) { // Child
167 int sock = socket(AF_UNIX,
169 SOCK_STREAM, /* Mac OSX does not have AF_UNIX + SOCK_SEQPACKET, even if that's faster*/
175 struct sockaddr_un addr = {};
176 addr.sun_family = AF_UNIX;
177 snprintf(addr.sun_path, 64, "/tmp/simgrid-mc-%" PRIu64, msg->value);
178 auto addr_size = offsetof(struct sockaddr_un, sun_path) + strlen(addr.sun_path);
180 addr.sun_path[0] = '\0'; // abstract socket
183 xbt_assert(connect(sock, (struct sockaddr*)&addr, addr_size) >= 0, "Cannot connect to Checker on %c%s: %s.",
184 (addr.sun_path[0] ? addr.sun_path[0] : '@'), addr.sun_path + 1, strerror(errno));
186 channel_.reset_socket(sock);
188 s_mc_message_int_t answer = {};
189 answer.type = MessageType::FORK_REPLY;
190 answer.value = getpid();
191 xbt_assert(channel_.send(answer) == 0, "Could not send response to WAIT_CHILD_REPLY: %s", strerror(errno));
193 XBT_VERB("App %d forks subprocess %d.", getpid(), pid);
196 void AppSide::handle_wait_child(const s_mc_message_int_t* msg)
200 if (auto search = child_statuses_.find(msg->value); search != child_statuses_.end()) {
201 status = search->second;
202 child_statuses_.erase(search); // We only need this info once
204 waitpid(msg->value, &status, 0);
206 xbt_assert(errno == 0, "Cannot wait on behalf of the checker: %s.", strerror(errno));
208 s_mc_message_int_t answer = {};
209 answer.type = MessageType::WAIT_CHILD_REPLY;
210 answer.value = status;
211 xbt_assert(channel_.send(answer) == 0, "Could not send response to WAIT_CHILD: %s", strerror(errno));
213 void AppSide::handle_need_meminfo()
215 #if SIMGRID_HAVE_STATEFUL_MC
216 this->need_memory_info_ = true;
217 s_mc_message_need_meminfo_reply_t answer = {};
218 answer.type = MessageType::NEED_MEMINFO_REPLY;
219 answer.mmalloc_default_mdp = mmalloc_get_current_heap();
220 xbt_assert(channel_.send(answer) == 0, "Could not send response to the request for meminfo: %s", strerror(errno));
222 xbt_die("SimGrid was compiled without MC suppport, so liveness and similar features are not available.");
225 void AppSide::handle_actors_status() const
227 auto const& actor_list = kernel::EngineImpl::get_instance()->get_actor_list();
228 XBT_DEBUG("Serialize the actors to answer ACTORS_STATUS from the checker. %zu actors to go.", actor_list.size());
230 std::vector<s_mc_message_actors_status_one_t> status;
231 for (auto const& [aid, actor] : actor_list) {
232 s_mc_message_actors_status_one_t one = {};
233 one.type = MessageType::ACTORS_STATUS_REPLY_TRANSITION;
235 one.enabled = mc::actor_is_enabled(actor);
236 one.max_considered = actor->simcall_.observer_->get_max_consider();
237 status.push_back(one);
240 struct s_mc_message_actors_status_answer_t answer = {};
241 answer.type = MessageType::ACTORS_STATUS_REPLY_COUNT;
242 answer.count = static_cast<int>(status.size());
244 xbt_assert(channel_.send(answer) == 0, "Could not send ACTORS_STATUS_REPLY msg: %s", strerror(errno));
245 if (answer.count > 0) {
246 size_t size = status.size() * sizeof(s_mc_message_actors_status_one_t);
247 xbt_assert(channel_.send(status.data(), size) == 0, "Could not send ACTORS_STATUS_REPLY data: %s", strerror(errno));
250 // Serialize each transition to describe what each actor is doing
251 XBT_DEBUG("Deliver ACTOR_TRANSITION_PROBE payload");
252 for (const auto& actor_status : status) {
253 const auto& actor = actor_list.at(actor_status.aid);
254 const int max_considered = actor_status.max_considered;
256 for (int times_considered = 0; times_considered < max_considered; times_considered++) {
257 std::stringstream stream;
258 s_mc_message_simcall_probe_one_t probe;
259 probe.type = MessageType::ACTORS_STATUS_REPLY_SIMCALL;
261 if (actor->simcall_.observer_ != nullptr) {
262 actor->simcall_.observer_->prepare(times_considered);
263 actor->simcall_.observer_->serialize(stream);
265 stream << (short)mc::Transition::Type::UNKNOWN;
268 std::string str = stream.str();
269 xbt_assert(str.size() + 1 <= probe.buffer.size(),
270 "The serialized transition is too large for the buffer. Please fix the code.");
271 strncpy(probe.buffer.data(), str.c_str(), probe.buffer.size() - 1);
272 probe.buffer.back() = '\0';
274 xbt_assert(channel_.send(probe) == 0, "Could not send ACTOR_TRANSITION_PROBE payload: %s", strerror(errno));
276 // NOTE: We do NOT need to reset `times_considered` for each actor's
277 // simcall observer here to the "original" value (i.e. the value BEFORE
278 // multiple prepare() calls were made for serialization purposes) since
279 // each SIMCALL_EXECUTE provides a `times_considered` to be used to prepare
280 // the transition before execution.
283 void AppSide::handle_actors_maxpid() const
285 s_mc_message_int_t answer = {};
286 answer.type = MessageType::ACTORS_MAXPID_REPLY;
287 answer.value = kernel::actor::ActorImpl::get_maxpid();
288 xbt_assert(channel_.send(answer) == 0, "Could not send response: %s", strerror(errno));
291 #define assert_msg_size(_name_, _type_) \
292 xbt_assert(received_size == sizeof(_type_), "Unexpected size for " _name_ " (%zd != %zu)", received_size, \
295 void AppSide::handle_messages()
297 while (true) { // Until we get a CONTINUE message
298 XBT_DEBUG("Waiting messages from the model-checker");
300 std::array<char, MC_MESSAGE_LENGTH> message_buffer;
301 ssize_t received_size = channel_.receive(message_buffer.data(), message_buffer.size());
303 if (received_size == 0) {
304 XBT_DEBUG("Socket closed on the Checker side, bailing out.");
305 ::_Exit(0); // Nobody's listening to that process anymore => exit as quickly as possible.
307 xbt_assert(received_size >= 0, "Could not receive commands from the model-checker: %s", strerror(errno));
308 xbt_assert(static_cast<size_t>(received_size) >= sizeof(s_mc_message_t), "Cannot handle short message (size=%zd)",
311 const s_mc_message_t* message = (s_mc_message_t*)message_buffer.data();
312 switch (message->type) {
313 case MessageType::CONTINUE:
314 assert_msg_size("MESSAGE_CONTINUE", s_mc_message_t);
317 case MessageType::DEADLOCK_CHECK:
318 assert_msg_size("DEADLOCK_CHECK", s_mc_message_t);
319 handle_deadlock_check(message);
322 case MessageType::SIMCALL_EXECUTE:
323 assert_msg_size("SIMCALL_EXECUTE", s_mc_message_simcall_execute_t);
324 handle_simcall_execute((s_mc_message_simcall_execute_t*)message_buffer.data());
327 case MessageType::FINALIZE:
328 assert_msg_size("FINALIZE", s_mc_message_int_t);
329 handle_finalize((s_mc_message_int_t*)message_buffer.data());
332 case MessageType::FORK:
333 assert_msg_size("FORK", s_mc_message_int_t);
334 handle_fork((s_mc_message_int_t*)message_buffer.data());
337 case MessageType::WAIT_CHILD:
338 assert_msg_size("WAIT_CHILD", s_mc_message_int_t);
339 handle_wait_child((s_mc_message_int_t*)message_buffer.data());
342 case MessageType::NEED_MEMINFO:
343 assert_msg_size("NEED_MEMINFO", s_mc_message_t);
344 handle_need_meminfo();
347 case MessageType::ACTORS_STATUS:
348 assert_msg_size("ACTORS_STATUS", s_mc_message_t);
349 handle_actors_status();
352 case MessageType::ACTORS_MAXPID:
353 assert_msg_size("ACTORS_MAXPID", s_mc_message_t);
354 handle_actors_maxpid();
358 xbt_die("Received unexpected message %s (%i)", to_c_str(message->type), static_cast<int>(message->type));
364 void AppSide::main_loop()
366 simgrid::mc::processes_time.resize(simgrid::kernel::actor::ActorImpl::get_maxpid());
367 MC_ignore_heap(simgrid::mc::processes_time.data(),
368 simgrid::mc::processes_time.size() * sizeof(simgrid::mc::processes_time[0]));
371 coverage_checkpoint();
374 simgrid::mc::execute_actors();
375 xbt_assert(channel_.send(MessageType::WAITING) == 0, "Could not send WAITING message to model-checker: %s",
377 this->handle_messages();
381 void AppSide::report_assertion_failure()
383 xbt_assert(channel_.send(MessageType::ASSERTION_FAILED) == 0, "Could not send assertion to model-checker: %s",
385 this->handle_messages();
388 void AppSide::ignore_memory(void* addr, std::size_t size) const
390 if (not MC_is_active() || not need_memory_info_)
393 #if SIMGRID_HAVE_STATEFUL_MC
394 s_mc_message_ignore_memory_t message = {};
395 message.type = MessageType::IGNORE_MEMORY;
396 message.addr = (std::uintptr_t)addr;
398 xbt_assert(channel_.send(message) == 0, "Could not send IGNORE_MEMORY message to model-checker: %s", strerror(errno));
400 xbt_die("Cannot really call ignore_heap() in non-SIMGRID_MC mode.");
404 void AppSide::ignore_heap(void* address, std::size_t size) const
406 if (not MC_is_active() || not need_memory_info_)
409 #if SIMGRID_HAVE_STATEFUL_MC
410 const s_xbt_mheap_t* heap = mmalloc_get_current_heap();
412 s_mc_message_ignore_heap_t message = {};
413 message.type = MessageType::IGNORE_HEAP;
414 message.address = address;
416 message.block = ((char*)address - (char*)heap->heapbase) / BLOCKSIZE + 1;
417 if (heap->heapinfo[message.block].type == 0) {
418 message.fragment = -1;
419 heap->heapinfo[message.block].busy_block.ignore++;
421 message.fragment = (ADDR2UINT(address) % BLOCKSIZE) >> heap->heapinfo[message.block].type;
422 heap->heapinfo[message.block].busy_frag.ignore[message.fragment]++;
425 xbt_assert(channel_.send(message) == 0, "Could not send ignored region to MCer: %s", strerror(errno));
427 xbt_die("Cannot really call ignore_heap() in non-SIMGRID_MC mode.");
431 void AppSide::unignore_heap(void* address, std::size_t size) const
433 if (not MC_is_active() || not need_memory_info_)
436 #if SIMGRID_HAVE_STATEFUL_MC
437 s_mc_message_ignore_memory_t message = {};
438 message.type = MessageType::UNIGNORE_HEAP;
439 message.addr = (std::uintptr_t)address;
441 xbt_assert(channel_.send(message) == 0, "Could not send IGNORE_HEAP message to model-checker: %s", strerror(errno));
443 xbt_die("Cannot really call unignore_heap() in non-SIMGRID_MC mode.");
447 void AppSide::declare_symbol(const char* name, int* value) const
449 if (not MC_is_active() || not need_memory_info_) {
450 XBT_CRITICAL("Ignore AppSide::declare_symbol(%s)", name);
454 #if SIMGRID_HAVE_STATEFUL_MC
455 s_mc_message_register_symbol_t message = {};
456 message.type = MessageType::REGISTER_SYMBOL;
457 xbt_assert(strlen(name) + 1 <= message.name.size(), "Symbol is too long");
458 strncpy(message.name.data(), name, message.name.size() - 1);
459 message.callback = nullptr;
460 message.data = value;
461 xbt_assert(channel_.send(message) == 0, "Could send REGISTER_SYMBOL message to model-checker: %s", strerror(errno));
463 xbt_die("Cannot really call declare_symbol() in non-SIMGRID_MC mode.");
467 /** Register a stack in the model checker
469 * The stacks are allocated in the heap. The MC handle them specifically
470 * when we analyze/compare the content of the heap so it must be told where
471 * they are with this function.
473 #if HAVE_UCONTEXT_H /* Apple don't want us to use ucontexts */
474 void AppSide::declare_stack(void* stack, size_t size, ucontext_t* context) const
476 if (not MC_is_active() || not need_memory_info_)
479 #if SIMGRID_HAVE_STATEFUL_MC
480 const s_xbt_mheap_t* heap = mmalloc_get_current_heap();
482 s_stack_region_t region = {};
483 region.address = stack;
484 region.context = context;
486 region.block = ((char*)stack - (char*)heap->heapbase) / BLOCKSIZE + 1;
488 s_mc_message_stack_region_t message = {};
489 message.type = MessageType::STACK_REGION;
490 message.stack_region = region;
491 xbt_assert(channel_.send(message) == 0, "Could not send STACK_REGION to model-checker: %s", strerror(errno));
493 xbt_die("Cannot really call declare_stack() in non-SIMGRID_MC mode.");
498 } // namespace simgrid::mc