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 #if SIMGRID_HAVE_STATEFUL_MC
15 #include "src/mc/sosp/RemoteProcessMemory.hpp"
18 #include "src/smpi/include/private.hpp"
20 #include "src/sthread/sthread.h"
21 #include "src/xbt/coverage.h"
23 #include <simgrid/modelchecker.h>
26 #include <cstdio> // setvbuf
30 #include <sys/ptrace.h>
31 #include <sys/socket.h>
32 #include <sys/types.h>
36 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(mc_client, mc, "MC client logic");
37 XBT_LOG_EXTERNAL_CATEGORY(mc_global);
39 namespace simgrid::mc {
41 std::unique_ptr<AppSide> AppSide::instance_;
43 AppSide* AppSide::get()
45 // Only initialize the MC world once
46 if (instance_ != nullptr)
47 return instance_.get();
49 if (std::getenv(MC_ENV_SOCKET_FD) == nullptr) // We are not in MC mode: don't initialize the MC world
52 XBT_DEBUG("Initialize the MC world. MC_NEED_PTRACE=%s", std::getenv("MC_NEED_PTRACE"));
54 simgrid::mc::set_model_checking_mode(ModelCheckingMode::APP_SIDE);
56 setvbuf(stdout, nullptr, _IOLBF, 0);
58 // Fetch socket from MC_ENV_SOCKET_FD:
59 const char* fd_env = std::getenv(MC_ENV_SOCKET_FD);
60 int fd = xbt_str_parse_int(fd_env, "Not a number in variable '" MC_ENV_SOCKET_FD "'");
61 XBT_DEBUG("Model-checked application found socket FD %i", fd);
63 instance_ = std::make_unique<simgrid::mc::AppSide>(fd);
65 // Wait for the model-checker:
66 if (getenv("MC_NEED_PTRACE") != nullptr) {
69 ptrace(PTRACE_TRACEME, 0, nullptr, nullptr);
71 ptrace(PT_TRACE_ME, 0, nullptr, 0);
73 xbt_die("no ptrace equivalent coded for this platform, please don't use the liveness checker here.");
76 xbt_assert(errno == 0 && raise(SIGSTOP) == 0, "Could not wait for the model-checker (errno = %d: %s)", errno,
80 instance_->handle_messages();
81 return instance_.get();
84 void AppSide::handle_deadlock_check(const s_mc_message_t*) const
86 const auto* engine = kernel::EngineImpl::get_instance();
87 const auto& actor_list = engine->get_actor_list();
88 bool deadlock = not actor_list.empty() && std::none_of(begin(actor_list), end(actor_list), [](const auto& kv) {
89 return mc::actor_is_enabled(kv.second);
93 XBT_CINFO(mc_global, "**************************");
94 XBT_CINFO(mc_global, "*** DEADLOCK DETECTED ***");
95 XBT_CINFO(mc_global, "**************************");
96 engine->display_all_actor_status();
99 s_mc_message_int_t answer = {};
100 answer.type = MessageType::DEADLOCK_CHECK_REPLY;
101 answer.value = deadlock;
102 xbt_assert(channel_.send(answer) == 0, "Could not send response");
104 void AppSide::handle_simcall_execute(const s_mc_message_simcall_execute_t* message) const
106 kernel::actor::ActorImpl* actor = kernel::EngineImpl::get_instance()->get_actor_by_pid(message->aid_);
107 xbt_assert(actor != nullptr, "Invalid pid %ld", message->aid_);
109 // The client may send some messages to the server while processing the transition
110 actor->simcall_handle(message->times_considered_);
111 // Say the server that the transition is over and that it should proceed
112 xbt_assert(channel_.send(MessageType::WAITING) == 0, "Could not send MESSAGE_WAITING to model-checker");
114 // Finish the RPC from the server: return a serialized observer, to build a Transition on Checker side
115 s_mc_message_simcall_execute_answer_t answer = {};
116 answer.type = MessageType::SIMCALL_EXECUTE_REPLY;
117 std::stringstream stream;
118 if (actor->simcall_.observer_ != nullptr) {
119 actor->simcall_.observer_->serialize(stream);
121 stream << (short)mc::Transition::Type::UNKNOWN;
123 std::string str = stream.str();
124 xbt_assert(str.size() + 1 <= answer.buffer.size(),
125 "The serialized simcall is too large for the buffer. Please fix the code.");
126 strncpy(answer.buffer.data(), str.c_str(), answer.buffer.size() - 1);
127 answer.buffer.back() = '\0';
129 XBT_DEBUG("send SIMCALL_EXECUTE_ANSWER(%s) ~> '%s'", actor->get_cname(), str.c_str());
130 xbt_assert(channel_.send(answer) == 0, "Could not send response");
133 void AppSide::handle_finalize(const s_mc_message_int_t* msg) const
135 bool terminate_asap = msg->value;
136 XBT_DEBUG("Finalize (terminate = %d)", (int)terminate_asap);
137 if (not terminate_asap) {
138 if (XBT_LOG_ISENABLED(mc_client, xbt_log_priority_debug))
139 kernel::EngineImpl::get_instance()->display_all_actor_status();
141 XBT_DEBUG("Smpi_enabled: %d", SMPI_is_inited());
142 if (SMPI_is_inited())
146 coverage_checkpoint();
147 xbt_assert(channel_.send(MessageType::FINALIZE_REPLY) == 0, "Could not answer to FINALIZE");
152 void AppSide::handle_fork(const s_mc_message_int_t* msg)
155 xbt_assert(pid >= 0, "Could not fork application sub-process: %s.", strerror(errno));
157 if (pid == 0) { // Child
158 int sock = socket(AF_UNIX,
160 SOCK_STREAM, /* Mac OSX does not have AF_UNIX + SOCK_SEQPACKET, even if that's faster*/
166 struct sockaddr_un addr = {};
167 addr.sun_family = AF_UNIX;
168 snprintf(addr.sun_path, 64, "/tmp/simgrid-mc-%lu", static_cast<unsigned long>(msg->value));
169 auto addr_size = offsetof(struct sockaddr_un, sun_path) + strlen(addr.sun_path);
171 xbt_assert(connect(sock, (struct sockaddr*)&addr, addr_size) >= 0,
172 "Cannot connect to Checker on %s: %s.", addr.sun_path, strerror(errno));
174 channel_.reset_socket(sock);
176 s_mc_message_int_t answer = {};
177 answer.type = MessageType::FORK_REPLY;
178 answer.value = getpid();
179 xbt_assert(channel_.send(answer) == 0, "Could not send response to WAIT_CHILD_REPLY: %s", strerror(errno));
181 XBT_DEBUG("App %d forks subprocess %d.", getpid(), pid);
184 void AppSide::handle_wait_child(const s_mc_message_int_t* msg)
188 waitpid(msg->value, &status, 0);
189 xbt_assert(errno == 0, "Cannot wait on behalf of the checker: %s.", strerror(errno));
191 s_mc_message_int_t answer = {};
192 answer.type = MessageType::WAIT_CHILD_REPLY;
193 answer.value = status;
194 xbt_assert(channel_.send(answer) == 0, "Could not send response to WAIT_CHILD: %s", strerror(errno));
196 void AppSide::handle_need_meminfo()
198 #if SIMGRID_HAVE_STATEFUL_MC
199 this->need_memory_info_ = true;
200 s_mc_message_need_meminfo_reply_t answer = {};
201 answer.type = MessageType::NEED_MEMINFO_REPLY;
202 answer.mmalloc_default_mdp = mmalloc_get_current_heap();
203 xbt_assert(channel_.send(answer) == 0, "Could not send response to the request for meminfo.");
205 xbt_die("SimGrid was compiled without MC suppport, so liveness and similar features are not available.");
208 void AppSide::handle_actors_status() const
210 auto const& actor_list = kernel::EngineImpl::get_instance()->get_actor_list();
211 XBT_DEBUG("Serialize the actors to answer ACTORS_STATUS from the checker. %zu actors to go.", actor_list.size());
213 std::vector<s_mc_message_actors_status_one_t> status;
214 for (auto const& [aid, actor] : actor_list) {
215 s_mc_message_actors_status_one_t one = {};
216 one.type = MessageType::ACTORS_STATUS_REPLY_TRANSITION;
218 one.enabled = mc::actor_is_enabled(actor);
219 one.max_considered = actor->simcall_.observer_->get_max_consider();
220 status.push_back(one);
223 struct s_mc_message_actors_status_answer_t answer = {};
224 answer.type = MessageType::ACTORS_STATUS_REPLY_COUNT;
225 answer.count = static_cast<int>(status.size());
227 xbt_assert(channel_.send(answer) == 0, "Could not send ACTORS_STATUS_REPLY msg");
228 if (answer.count > 0) {
229 size_t size = status.size() * sizeof(s_mc_message_actors_status_one_t);
230 xbt_assert(channel_.send(status.data(), size) == 0, "Could not send ACTORS_STATUS_REPLY data");
233 // Serialize each transition to describe what each actor is doing
234 XBT_DEBUG("Deliver ACTOR_TRANSITION_PROBE payload");
235 for (const auto& actor_status : status) {
236 if (not actor_status.enabled)
239 const auto& actor = actor_list.at(actor_status.aid);
240 const int max_considered = actor_status.max_considered;
242 for (int times_considered = 0; times_considered < max_considered; times_considered++) {
243 std::stringstream stream;
244 s_mc_message_simcall_probe_one_t probe;
245 probe.type = MessageType::ACTORS_STATUS_REPLY_SIMCALL;
247 if (actor->simcall_.observer_ != nullptr) {
248 actor->simcall_.observer_->prepare(times_considered);
249 actor->simcall_.observer_->serialize(stream);
251 stream << (short)mc::Transition::Type::UNKNOWN;
254 std::string str = stream.str();
255 xbt_assert(str.size() + 1 <= probe.buffer.size(),
256 "The serialized transition is too large for the buffer. Please fix the code.");
257 strncpy(probe.buffer.data(), str.c_str(), probe.buffer.size() - 1);
258 probe.buffer.back() = '\0';
260 xbt_assert(channel_.send(probe) == 0, "Could not send ACTOR_TRANSITION_PROBE payload");
262 // NOTE: We do NOT need to reset `times_considered` for each actor's
263 // simcall observer here to the "original" value (i.e. the value BEFORE
264 // multiple prepare() calls were made for serialization purposes) since
265 // each SIMCALL_EXECUTE provides a `times_considered` to be used to prepare
266 // the transition before execution.
269 void AppSide::handle_actors_maxpid() const
271 s_mc_message_int_t answer = {};
272 answer.type = MessageType::ACTORS_MAXPID_REPLY;
273 answer.value = kernel::actor::ActorImpl::get_maxpid();
274 xbt_assert(channel_.send(answer) == 0, "Could not send response");
277 #define assert_msg_size(_name_, _type_) \
278 xbt_assert(received_size == sizeof(_type_), "Unexpected size for " _name_ " (%zd != %zu)", received_size, \
281 void AppSide::handle_messages()
283 while (true) { // Until we get a CONTINUE message
284 XBT_DEBUG("Waiting messages from the model-checker");
286 std::array<char, MC_MESSAGE_LENGTH> message_buffer;
287 ssize_t received_size = channel_.receive(message_buffer.data(), message_buffer.size());
289 if (received_size == 0) {
290 XBT_DEBUG("Socket closed on the Checker side, bailing out.");
291 ::_Exit(0); // Nobody's listening to that process anymore => exit as quickly as possible.
293 xbt_assert(received_size >= 0, "Could not receive commands from the model-checker");
294 xbt_assert(static_cast<size_t>(received_size) >= sizeof(s_mc_message_t), "Cannot handle short message (size=%zd)",
297 const s_mc_message_t* message = (s_mc_message_t*)message_buffer.data();
298 switch (message->type) {
299 case MessageType::CONTINUE:
300 assert_msg_size("MESSAGE_CONTINUE", s_mc_message_t);
303 case MessageType::DEADLOCK_CHECK:
304 assert_msg_size("DEADLOCK_CHECK", s_mc_message_t);
305 handle_deadlock_check(message);
308 case MessageType::SIMCALL_EXECUTE:
309 assert_msg_size("SIMCALL_EXECUTE", s_mc_message_simcall_execute_t);
310 handle_simcall_execute((s_mc_message_simcall_execute_t*)message_buffer.data());
313 case MessageType::FINALIZE:
314 assert_msg_size("FINALIZE", s_mc_message_int_t);
315 handle_finalize((s_mc_message_int_t*)message_buffer.data());
318 case MessageType::FORK:
319 assert_msg_size("FORK", s_mc_message_int_t);
320 handle_fork((s_mc_message_int_t*)message_buffer.data());
323 case MessageType::WAIT_CHILD:
324 assert_msg_size("WAIT_CHILD", s_mc_message_int_t);
325 handle_wait_child((s_mc_message_int_t*)message_buffer.data());
328 case MessageType::NEED_MEMINFO:
329 assert_msg_size("NEED_MEMINFO", s_mc_message_t);
330 handle_need_meminfo();
333 case MessageType::ACTORS_STATUS:
334 assert_msg_size("ACTORS_STATUS", s_mc_message_t);
335 handle_actors_status();
338 case MessageType::ACTORS_MAXPID:
339 assert_msg_size("ACTORS_MAXPID", s_mc_message_t);
340 handle_actors_maxpid();
344 xbt_die("Received unexpected message %s (%i)", to_c_str(message->type), static_cast<int>(message->type));
350 void AppSide::main_loop()
352 simgrid::mc::processes_time.resize(simgrid::kernel::actor::ActorImpl::get_maxpid());
353 MC_ignore_heap(simgrid::mc::processes_time.data(),
354 simgrid::mc::processes_time.size() * sizeof(simgrid::mc::processes_time[0]));
357 coverage_checkpoint();
360 simgrid::mc::execute_actors();
361 xbt_assert(channel_.send(MessageType::WAITING) == 0, "Could not send WAITING message to model-checker");
362 this->handle_messages();
366 void AppSide::report_assertion_failure()
368 xbt_assert(channel_.send(MessageType::ASSERTION_FAILED) == 0, "Could not send assertion to model-checker");
369 this->handle_messages();
372 void AppSide::ignore_memory(void* addr, std::size_t size) const
374 if (not MC_is_active() || not need_memory_info_)
377 #if SIMGRID_HAVE_STATEFUL_MC
378 s_mc_message_ignore_memory_t message = {};
379 message.type = MessageType::IGNORE_MEMORY;
380 message.addr = (std::uintptr_t)addr;
382 xbt_assert(channel_.send(message) == 0, "Could not send IGNORE_MEMORY message to model-checker");
384 xbt_die("Cannot really call ignore_heap() in non-SIMGRID_MC mode.");
388 void AppSide::ignore_heap(void* address, std::size_t size) const
390 if (not MC_is_active() || not need_memory_info_)
393 #if SIMGRID_HAVE_STATEFUL_MC
394 const s_xbt_mheap_t* heap = mmalloc_get_current_heap();
396 s_mc_message_ignore_heap_t message = {};
397 message.type = MessageType::IGNORE_HEAP;
398 message.address = address;
400 message.block = ((char*)address - (char*)heap->heapbase) / BLOCKSIZE + 1;
401 if (heap->heapinfo[message.block].type == 0) {
402 message.fragment = -1;
403 heap->heapinfo[message.block].busy_block.ignore++;
405 message.fragment = (ADDR2UINT(address) % BLOCKSIZE) >> heap->heapinfo[message.block].type;
406 heap->heapinfo[message.block].busy_frag.ignore[message.fragment]++;
409 xbt_assert(channel_.send(message) == 0, "Could not send ignored region to MCer");
411 xbt_die("Cannot really call ignore_heap() in non-SIMGRID_MC mode.");
415 void AppSide::unignore_heap(void* address, std::size_t size) const
417 if (not MC_is_active() || not need_memory_info_)
420 #if SIMGRID_HAVE_STATEFUL_MC
421 s_mc_message_ignore_memory_t message = {};
422 message.type = MessageType::UNIGNORE_HEAP;
423 message.addr = (std::uintptr_t)address;
425 xbt_assert(channel_.send(message) == 0, "Could not send IGNORE_HEAP message to model-checker");
427 xbt_die("Cannot really call unignore_heap() in non-SIMGRID_MC mode.");
431 void AppSide::declare_symbol(const char* name, int* value) const
433 if (not MC_is_active() || not need_memory_info_) {
434 XBT_CRITICAL("Ignore AppSide::declare_symbol(%s)", name);
438 #if SIMGRID_HAVE_STATEFUL_MC
439 s_mc_message_register_symbol_t message = {};
440 message.type = MessageType::REGISTER_SYMBOL;
441 xbt_assert(strlen(name) + 1 <= message.name.size(), "Symbol is too long");
442 strncpy(message.name.data(), name, message.name.size() - 1);
443 message.callback = nullptr;
444 message.data = value;
445 xbt_assert(channel_.send(message) == 0, "Could send REGISTER_SYMBOL message to model-checker");
447 xbt_die("Cannot really call declare_symbol() in non-SIMGRID_MC mode.");
451 /** Register a stack in the model checker
453 * The stacks are allocated in the heap. The MC handle them specifically
454 * when we analyze/compare the content of the heap so it must be told where
455 * they are with this function.
457 #if HAVE_UCONTEXT_H /* Apple don't want us to use ucontexts */
458 void AppSide::declare_stack(void* stack, size_t size, ucontext_t* context) const
460 if (not MC_is_active() || not need_memory_info_)
463 #if SIMGRID_HAVE_STATEFUL_MC
464 const s_xbt_mheap_t* heap = mmalloc_get_current_heap();
466 s_stack_region_t region = {};
467 region.address = stack;
468 region.context = context;
470 region.block = ((char*)stack - (char*)heap->heapbase) / BLOCKSIZE + 1;
472 s_mc_message_stack_region_t message = {};
473 message.type = MessageType::STACK_REGION;
474 message.stack_region = region;
475 xbt_assert(channel_.send(message) == 0, "Could not send STACK_REGION to model-checker");
477 xbt_die("Cannot really call declare_stack() in non-SIMGRID_MC mode.");
482 } // namespace simgrid::mc