1 /* Copyright (c) 2016-2023. 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/explo/DFSExplorer.hpp"
7 #include "src/mc/mc_config.hpp"
8 #include "src/mc/mc_exit.hpp"
9 #include "src/mc/mc_private.hpp"
10 #include "src/mc/mc_record.hpp"
11 #include "src/mc/transition/Transition.hpp"
13 #if SIMGRID_HAVE_STATEFUL_MC
14 #include "src/mc/VisitedState.hpp"
17 #include "src/xbt/mmalloc/mmprivate.h"
19 #include "xbt/string.hpp"
20 #include "xbt/sysdep.h"
28 #include <unordered_set>
31 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(mc_dfs, mc, "DFS exploration algorithm of the model-checker");
33 namespace simgrid::mc {
35 xbt::signal<void(RemoteApp&)> DFSExplorer::on_exploration_start_signal;
36 xbt::signal<void(RemoteApp&)> DFSExplorer::on_backtracking_signal;
38 xbt::signal<void(State*, RemoteApp&)> DFSExplorer::on_state_creation_signal;
40 xbt::signal<void(State*, RemoteApp&)> DFSExplorer::on_restore_system_state_signal;
41 xbt::signal<void(RemoteApp&)> DFSExplorer::on_restore_initial_state_signal;
42 xbt::signal<void(Transition*, RemoteApp&)> DFSExplorer::on_transition_replay_signal;
43 xbt::signal<void(Transition*, RemoteApp&)> DFSExplorer::on_transition_execute_signal;
45 xbt::signal<void(RemoteApp&)> DFSExplorer::on_log_state_signal;
47 void DFSExplorer::check_non_termination(const State* current_state)
49 #if SIMGRID_HAVE_STATEFUL_MC
50 for (auto const& state : stack_) {
51 if (state->get_system_state()->equals_to(*current_state->get_system_state(),
52 *get_remote_app().get_remote_process_memory())) {
53 XBT_INFO("Non-progressive cycle: state %ld -> state %ld", state->get_num(), current_state->get_num());
54 XBT_INFO("******************************************");
55 XBT_INFO("*** NON-PROGRESSIVE CYCLE DETECTED ***");
56 XBT_INFO("******************************************");
57 XBT_INFO("Counter-example execution trace:");
58 for (auto const& s : get_textual_trace())
59 XBT_INFO(" %s", s.c_str());
60 XBT_INFO("You can debug the problem (and see the whole details) by rerunning out of simgrid-mc with "
61 "--cfg=model-check/replay:'%s'",
62 get_record_trace().to_string().c_str());
65 throw McError(ExitStatus::NON_TERMINATION);
71 RecordTrace DFSExplorer::get_record_trace() // override
75 if (const auto trans = stack_.back()->get_transition_out(); trans != nullptr)
76 res.push_back(trans.get());
77 for (const auto* state = stack_.back().get(); state != nullptr; state = state->get_parent_state().get())
78 if (state->get_transition_in() != nullptr)
79 res.push_front(state->get_transition_in().get());
84 void DFSExplorer::restore_stack(std::shared_ptr<State> state)
87 auto current_state = state;
88 stack_.emplace_front(current_state);
89 // condition corresponds to reaching initial state
90 while (current_state->get_parent_state() != nullptr) {
91 current_state = current_state->get_parent_state();
92 stack_.emplace_front(current_state);
94 XBT_DEBUG("Replaced stack by %s", get_record_trace().to_string().c_str());
97 void DFSExplorer::log_state() // override
99 on_log_state_signal(get_remote_app());
100 XBT_INFO("DFS exploration ended. %ld unique states visited; %lu backtracks (%lu transition replays, %lu states "
102 State::get_expanded_states(), backtrack_count_, visited_states_count_,
103 Transition::get_replayed_transitions());
104 Exploration::log_state();
107 void DFSExplorer::run()
109 on_exploration_start_signal(get_remote_app());
110 /* This function runs the DFS algorithm the state space.
111 * We do so iteratively instead of recursively, dealing with the call stack manually.
112 * This allows one to explore the call stack at will. */
114 while (not stack_.empty()) {
115 /* Get current state */
116 auto state = stack_.back();
118 XBT_DEBUG("**************************************************");
119 XBT_DEBUG("Exploration depth=%zu (state:#%ld; %zu interleaves todo)", stack_.size(), state->get_num(),
120 state->count_todo());
122 visited_states_count_++;
124 // Backtrack if we reached the maximum depth
125 if (stack_.size() > (std::size_t)_sg_mc_max_depth) {
126 if (reduction_mode_ == ReductionMode::dpor) {
127 XBT_ERROR("/!\\ Max depth of %d reached! THIS WILL PROBABLY BREAK the dpor reduction /!\\",
128 _sg_mc_max_depth.get());
129 XBT_ERROR("/!\\ If bad things happen, disable dpor with --cfg=model-check/reduction:none /!\\");
131 XBT_WARN("/!\\ Max depth reached ! /!\\ ");
136 #if SIMGRID_HAVE_STATEFUL_MC
137 // Backtrack if we are revisiting a state we saw previously while applying state-equality reduction
138 if (visited_state_ != nullptr) {
139 XBT_DEBUG("State already visited (equal to state %ld), exploration stopped on this path.",
140 visited_state_->original_num_ == -1 ? visited_state_->num_ : visited_state_->original_num_);
142 visited_state_ = nullptr;
148 // Search for the next transition
149 // next_transition returns a pair<aid_t, int> in case we want to consider multiple state (eg. during backtrack)
150 auto [next, _] = state->next_transition_guided();
152 if (next < 0) { // If there is no more transition in the current state, backtrack.
153 XBT_VERB("%lu actors remain, but none of them need to be interleaved (depth %zu).", state->get_actor_count(),
156 if (state->get_actor_count() == 0) {
157 get_remote_app().finalize_app();
158 XBT_VERB("Execution came to an end at %s (state: %ld, depth: %zu)", get_record_trace().to_string().c_str(),
159 state->get_num(), stack_.size());
166 if (_sg_mc_sleep_set && XBT_LOG_ISENABLED(mc_dfs, xbt_log_priority_verbose)) {
167 XBT_VERB("Sleep set actually containing:");
168 for (auto& [aid, transition] : state->get_sleep_set())
169 XBT_VERB(" <%ld,%s>", aid, transition.to_string().c_str());
172 /* Actually answer the request: let's execute the selected request (MCed does one step) */
173 const auto executed_transition = state->execute_next(next, get_remote_app());
174 on_transition_execute_signal(state->get_transition_out().get(), get_remote_app());
176 // If there are processes to interleave and the maximum depth has not been
177 // reached then perform one step of the exploration algorithm.
178 XBT_VERB("Execute %ld: %.60s (stack depth: %zu, state: %ld, %zu interleaves)", state->get_transition_out()->aid_,
179 state->get_transition_out()->to_string().c_str(), stack_.size(), state->get_num(), state->count_todo());
181 /* Create the new expanded state (copy the state of MCed into our MCer data) */
182 auto next_state = std::make_shared<State>(get_remote_app(), state);
183 on_state_creation_signal(next_state.get(), get_remote_app());
185 /* Sleep set procedure:
186 * adding the taken transition to the sleep set of the original state.
187 * <!> Since the parent sleep set is used to compute the child sleep set, this need to be
188 * done after next_state creation */
189 XBT_DEBUG("Marking Transition >>%s<< of process %ld done and adding it to the sleep set",
190 state->get_transition_out()->to_string().c_str(), state->get_transition_out()->aid_);
191 state->add_sleep_set(state->get_transition_out()); // Actors are marked done when they are considerd in ActorState
193 /* DPOR persistent set procedure:
194 * for each new transition considered, check if it depends on any other previous transition executed before it
195 * on another process. If there exists one, find the more recent, and add its process to the interleave set.
196 * If the process is not enabled at this point, then add every enabled process to the interleave */
197 if (reduction_mode_ == ReductionMode::dpor) {
198 aid_t issuer_id = state->get_transition_out()->aid_;
199 stack_t tmp_stack = stack_;
200 while (not tmp_stack.empty()) {
201 if (const State* prev_state = tmp_stack.back().get();
202 state->get_transition_out()->aid_ == prev_state->get_transition_out()->aid_) {
203 XBT_DEBUG("Simcall >>%s<< and >>%s<< with same issuer %ld", state->get_transition_out()->to_string().c_str(),
204 prev_state->get_transition_out()->to_string().c_str(), issuer_id);
205 tmp_stack.pop_back();
207 } else if (prev_state->get_transition_out()->depends(state->get_transition_out().get())) {
208 XBT_VERB("Dependent Transitions:");
209 XBT_VERB(" %s (state=%ld)", prev_state->get_transition_out()->to_string().c_str(), prev_state->get_num());
210 XBT_VERB(" %s (state=%ld)", state->get_transition_out()->to_string().c_str(), state->get_num());
212 if (prev_state->is_actor_enabled(issuer_id)) {
213 if (not prev_state->is_actor_done(issuer_id)) {
214 prev_state->consider_one(issuer_id);
215 opened_states_.emplace_back(tmp_stack.back());
217 XBT_DEBUG("Actor %ld is already in done set: no need to explore it again", issuer_id);
219 XBT_DEBUG("Actor %ld is not enabled: DPOR may be failing. To stay sound, we are marking every enabled "
220 "transition as todo",
222 // If we ended up marking at least a transition, explore it at some point
223 if (prev_state->consider_all() > 0)
224 opened_states_.emplace_back(tmp_stack.back());
228 XBT_VERB("INDEPENDENT Transitions:");
229 XBT_VERB(" %s (state=%ld)", prev_state->get_transition_out()->to_string().c_str(), prev_state->get_num());
230 XBT_VERB(" %s (state=%ld)", state->get_transition_out()->to_string().c_str(), state->get_num());
232 tmp_stack.pop_back();
234 } else if (reduction_mode_ == ReductionMode::sdpor) {
236 * SDPOR Source Set Procedure:
238 execution_seq_.push_transition(executed_transition.get());
240 // To determine if the race is reversible, we have to ensure
241 // that actor `p` running `next_E_p` (viz. the event such that
242 // `racing_event -> (E_p) next_E_p` and no other event
243 // "happens-between" the two) is enabled in any equivalent
244 // execution where `racing_event` happens before `next_E_p`.
246 // Importantly, it is equivalent to checking if in ANY
247 // such equivalent execution sequence where `racing_event`
248 // happens-before `next_E_p` that `p` is enabled in `pre(racing_event, E.p)`.
249 // Thus it suffices to check THIS execution
250 xbt_assert(execution_seq_.get_latest_event_handle().has_value(),
251 "No events are contained in the SDPOR/OPDPOR execution "
252 "even though one was just added");
253 const aid_t p = executed_transition->aid_;
254 const auto next_E_p = execution_seq_.get_latest_event_handle().value();
256 for (const auto racing_event_handle : execution_seq_.get_racing_events_of(next_E_p)) {
257 // If the actor `p` is not enabled at s_[E'], it is not a *reversible* race
258 const std::shared_ptr<State> prev_state = stack_[racing_event_handle];
259 if (not prev_state->is_actor_enabled(p)) {
263 // This is a reversible race! First, grab `E' := pre(e, E)`
264 // TODO: Instead of copying around these big structs, it
265 // would behoove us to incorporate some way to reference
266 // portions of an execution. For simplicity and for a
267 // "proof of concept" version, we opt to simply copy
268 // the contents instead of making a view into the execution
269 const sdpor::Execution E_prime_v = execution_seq_.get_prefix_up_to(racing_event_handle);
271 // The vector `v` is constructed as `v := notdep(e, E)
272 std::vector<sdpor::Execution::EventHandle> v(execution_seq_.size());
273 std::unordered_set<aid_t> disqualified_actors = state->get_todo_actors();
275 for (auto e_prime = racing_event_handle; e_prime <= next_E_p; ++e_prime) {
276 // Any event `e` which occurs after `racing_event_handle` but which does not
277 // happen after `racing_event_handle` is a member of `v`
278 if (not E_prime_v.happens_before(racing_event_handle, e_prime) or e_prime == next_E_p) {
279 v.push_back(e_prime);
281 const aid_t q = E_prime_v.get_actor_with_handle(e_prime);
282 if (disqualified_actors.count(q) > 0) {
286 const bool is_initial = std::none_of(v.begin(), v.end(), [&E_prime_v, e_prime](const auto& e_star) {
287 return E_prime_v.happens_before(e_star, e_prime);
290 if (not prev_state->is_actor_done(q)) {
291 prev_state->consider_one(q);
292 opened_states_.emplace_back(std::move(prev_state));
296 disqualified_actors.insert(q);
302 // Before leaving that state, if the transition we just took can be taken multiple times, we
303 // need to give it to the opened states
304 if (stack_.back()->count_todo_multiples() > 0)
305 opened_states_.emplace_back(stack_.back());
307 if (_sg_mc_termination)
308 this->check_non_termination(next_state.get());
310 #if SIMGRID_HAVE_STATEFUL_MC
311 /* Check whether we already explored next_state in the past (but only if interested in state-equality reduction)
313 if (_sg_mc_max_visited_states > 0)
314 visited_state_ = visited_states_.addVisitedState(next_state->get_num(), next_state.get(), get_remote_app());
317 stack_.emplace_back(std::move(next_state));
319 /* If this is a new state (or if we don't care about state-equality reduction) */
320 if (visited_state_ == nullptr) {
321 /* Get an enabled process and insert it in the interleave set of the next state */
322 if (reduction_mode_ == ReductionMode::dpor)
323 stack_.back()->consider_best(); // Take only one transition if DPOR: others may be considered later if required
325 stack_.back()->consider_all();
328 dot_output("\"%ld\" -> \"%ld\" [%s];\n", state->get_num(), stack_.back()->get_num(),
329 state->get_transition_out()->dot_string().c_str());
330 #if SIMGRID_HAVE_STATEFUL_MC
332 dot_output("\"%ld\" -> \"%ld\" [%s];\n", state->get_num(),
333 visited_state_->original_num_ == -1 ? visited_state_->num_ : visited_state_->original_num_,
334 state->get_transition_out()->dot_string().c_str());
341 std::shared_ptr<State> DFSExplorer::best_opened_state()
343 int best_prio = 0; // cache the value for the best priority found so far (initialized to silence gcc)
344 auto best = end(opened_states_); // iterator to the state to explore having the best priority
345 auto valid = begin(opened_states_); // iterator marking the limit between states still to explore, and already
348 // Keep only still non-explored states (aid != -1), and record the one with the best (greater) priority.
349 for (auto current = begin(opened_states_); current != end(opened_states_); ++current) {
350 auto [aid, prio] = (*current)->next_transition_guided();
353 if (valid != current)
354 *valid = std::move(*current);
355 if (best == end(opened_states_) || prio > best_prio) {
362 std::shared_ptr<State> best_state;
364 // There are non-explored states, and one of them has the best priority. Remove it from opened_states_ before
366 best_state = std::move(*best);
369 *best = std::move(*valid);
371 opened_states_.erase(valid, end(opened_states_));
376 void DFSExplorer::backtrack()
378 XBT_VERB("Backtracking from %s", get_record_trace().to_string().c_str());
379 XBT_DEBUG("%lu alternatives are yet to be explored:", opened_states_.size());
381 on_backtracking_signal(get_remote_app());
382 get_remote_app().check_deadlock();
384 // Take the point with smallest distance
385 auto backtracking_point = best_opened_state();
387 // if no backtracking point, then set the stack_ to empty so we can end the exploration
388 if (not backtracking_point) {
389 XBT_DEBUG("No more opened point of exploration, the search will end");
394 // We found a backtracking point, let's go to it
396 XBT_DEBUG("Backtracking to state#%ld", backtracking_point->get_num());
398 #if SIMGRID_HAVE_STATEFUL_MC
399 /* If asked to rollback on a state that has a snapshot, restore it */
400 if (const auto* system_state = backtracking_point->get_system_state()) {
401 system_state->restore(*get_remote_app().get_remote_process_memory());
402 on_restore_system_state_signal(backtracking_point.get(), get_remote_app());
403 this->restore_stack(backtracking_point);
408 // Search how to restore the backtracking point
409 State* init_state = nullptr;
410 std::deque<Transition*> replay_recipe;
411 for (auto* s = backtracking_point.get(); s != nullptr; s = s->get_parent_state().get()) {
412 #if SIMGRID_HAVE_STATEFUL_MC
413 if (s->get_system_state() != nullptr) { // Found a state that I can restore
418 if (s->get_transition_in() != nullptr) // The root has no transition_in
419 replay_recipe.push_front(s->get_transition_in().get());
422 // Restore the init_state, if any
423 if (init_state != nullptr) {
424 #if SIMGRID_HAVE_STATEFUL_MC
425 const auto* system_state = init_state->get_system_state();
426 system_state->restore(*get_remote_app().get_remote_process_memory());
427 on_restore_system_state_signal(init_state, get_remote_app());
429 } else { // Restore the initial state if no intermediate state was found
430 get_remote_app().restore_initial_state();
431 on_restore_initial_state_signal(get_remote_app());
434 /* if no snapshot, we need to restore the initial state and replay the transitions */
435 /* Traverse the stack from the state at position start and re-execute the transitions */
436 for (auto& transition : replay_recipe) {
437 transition->replay(get_remote_app());
438 on_transition_replay_signal(transition, get_remote_app());
439 visited_states_count_++;
441 this->restore_stack(backtracking_point);
444 DFSExplorer::DFSExplorer(const std::vector<char*>& args, bool with_dpor, bool need_memory_info)
445 : Exploration(args, need_memory_info || _sg_mc_termination
446 #if SIMGRID_HAVE_STATEFUL_MC
447 || _sg_mc_checkpoint > 0
452 reduction_mode_ = ReductionMode::dpor;
454 reduction_mode_ = ReductionMode::none;
456 if (_sg_mc_termination) {
458 XBT_INFO("Check non progressive cycles (turning DPOR off)");
459 reduction_mode_ = ReductionMode::none;
461 XBT_INFO("Check non progressive cycles");
464 XBT_INFO("Start a DFS exploration. Reduction is: %s.", to_c_str(reduction_mode_));
466 auto initial_state = std::make_shared<State>(get_remote_app());
468 XBT_DEBUG("**************************************************");
470 stack_.emplace_back(std::move(initial_state));
472 /* Get an enabled actor and insert it in the interleave set of the initial state */
473 XBT_DEBUG("Initial state. %lu actors to consider", stack_.back()->get_actor_count());
474 if (reduction_mode_ == ReductionMode::dpor)
475 stack_.back()->consider_best();
477 stack_.back()->consider_all();
479 if (stack_.back()->count_todo_multiples() > 1)
480 opened_states_.emplace_back(stack_.back());
483 Exploration* create_dfs_exploration(const std::vector<char*>& args, bool with_dpor)
485 return new DFSExplorer(args, with_dpor);
488 } // namespace simgrid::mc