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());
96 if (reduction_mode_ == ReductionMode::sdpor) {
97 execution_seq_ = sdpor::Execution();
99 // NOTE: The outgoing transition for the top-most
100 // state of the stack refers to that which was taken
101 // as part of the last trace explored by the algorithm.
102 // Thus, only the sequence of transitions leading up to,
103 // but not including, the last state must be included
104 // when reconstructing the Exploration for SDPOR.
105 for (auto iter = stack_.begin(); iter != stack_.end() - 1 and iter != stack_.end(); ++iter) {
106 const auto& state = *(iter);
107 execution_seq_.push_transition(state->get_transition_out().get());
110 XBT_DEBUG("Additionally replaced corresponding SDPOR execution stack");
113 void DFSExplorer::log_state() // override
115 on_log_state_signal(get_remote_app());
116 XBT_INFO("DFS exploration ended. %ld unique states visited; %lu backtracks (%lu transition replays, %lu states "
118 State::get_expanded_states(), backtrack_count_, visited_states_count_,
119 Transition::get_replayed_transitions());
120 Exploration::log_state();
123 void DFSExplorer::run()
125 on_exploration_start_signal(get_remote_app());
126 /* This function runs the DFS algorithm the state space.
127 * We do so iteratively instead of recursively, dealing with the call stack manually.
128 * This allows one to explore the call stack at will. */
130 while (not stack_.empty()) {
131 /* Get current state */
132 auto state = stack_.back();
134 XBT_DEBUG("**************************************************");
135 XBT_DEBUG("Exploration depth=%zu (state:#%ld; %zu interleaves todo)", stack_.size(), state->get_num(),
136 state->count_todo());
138 visited_states_count_++;
140 // Backtrack if we reached the maximum depth
141 if (stack_.size() > (std::size_t)_sg_mc_max_depth) {
142 if (reduction_mode_ == ReductionMode::dpor) {
143 XBT_ERROR("/!\\ Max depth of %d reached! THIS WILL PROBABLY BREAK the dpor reduction /!\\",
144 _sg_mc_max_depth.get());
145 XBT_ERROR("/!\\ If bad things happen, disable dpor with --cfg=model-check/reduction:none /!\\");
147 XBT_WARN("/!\\ Max depth reached ! /!\\ ");
152 #if SIMGRID_HAVE_STATEFUL_MC
153 // Backtrack if we are revisiting a state we saw previously while applying state-equality reduction
154 if (visited_state_ != nullptr) {
155 XBT_DEBUG("State already visited (equal to state %ld), exploration stopped on this path.",
156 visited_state_->original_num_ == -1 ? visited_state_->num_ : visited_state_->original_num_);
158 visited_state_ = nullptr;
164 // Search for the next transition
165 // next_transition returns a pair<aid_t, int> in case we want to consider multiple state (eg. during backtrack)
166 auto [next, _] = state->next_transition_guided();
168 if (next < 0) { // If there is no more transition in the current state, backtrack.
169 XBT_VERB("%lu actors remain, but none of them need to be interleaved (depth %zu).", state->get_actor_count(),
172 if (state->get_actor_count() == 0) {
173 get_remote_app().finalize_app();
174 XBT_VERB("Execution came to an end at %s (state: %ld, depth: %zu)", get_record_trace().to_string().c_str(),
175 state->get_num(), stack_.size());
182 if (_sg_mc_sleep_set && XBT_LOG_ISENABLED(mc_dfs, xbt_log_priority_verbose)) {
183 XBT_VERB("Sleep set actually containing:");
184 for (auto& [aid, transition] : state->get_sleep_set())
185 XBT_VERB(" <%ld,%s>", aid, transition.to_string().c_str());
188 /* Actually answer the request: let's execute the selected request (MCed does one step) */
189 const auto executed_transition = state->execute_next(next, get_remote_app());
190 on_transition_execute_signal(state->get_transition_out().get(), get_remote_app());
192 // If there are processes to interleave and the maximum depth has not been
193 // reached then perform one step of the exploration algorithm.
194 XBT_VERB("Execute %ld: %.60s (stack depth: %zu, state: %ld, %zu interleaves)", state->get_transition_out()->aid_,
195 state->get_transition_out()->to_string().c_str(), stack_.size(), state->get_num(), state->count_todo());
197 /* Create the new expanded state (copy the state of MCed into our MCer data) */
198 auto next_state = std::make_shared<State>(get_remote_app(), state);
199 on_state_creation_signal(next_state.get(), get_remote_app());
201 /* Sleep set procedure:
202 * adding the taken transition to the sleep set of the original state.
203 * <!> Since the parent sleep set is used to compute the child sleep set, this need to be
204 * done after next_state creation */
205 XBT_DEBUG("Marking Transition >>%s<< of process %ld done and adding it to the sleep set",
206 state->get_transition_out()->to_string().c_str(), state->get_transition_out()->aid_);
207 state->add_sleep_set(state->get_transition_out()); // Actors are marked done when they are considerd in ActorState
209 /* DPOR persistent set procedure:
210 * for each new transition considered, check if it depends on any other previous transition executed before it
211 * on another process. If there exists one, find the more recent, and add its process to the interleave set.
212 * If the process is not enabled at this point, then add every enabled process to the interleave */
213 if (reduction_mode_ == ReductionMode::dpor) {
214 aid_t issuer_id = state->get_transition_out()->aid_;
215 stack_t tmp_stack = stack_;
216 while (not tmp_stack.empty()) {
217 if (const State* prev_state = tmp_stack.back().get();
218 state->get_transition_out()->aid_ == prev_state->get_transition_out()->aid_) {
219 XBT_DEBUG("Simcall >>%s<< and >>%s<< with same issuer %ld", state->get_transition_out()->to_string().c_str(),
220 prev_state->get_transition_out()->to_string().c_str(), issuer_id);
221 tmp_stack.pop_back();
223 } else if (prev_state->get_transition_out()->depends(state->get_transition_out().get())) {
224 XBT_VERB("Dependent Transitions:");
225 XBT_VERB(" %s (state=%ld)", prev_state->get_transition_out()->to_string().c_str(), prev_state->get_num());
226 XBT_VERB(" %s (state=%ld)", state->get_transition_out()->to_string().c_str(), state->get_num());
228 if (prev_state->is_actor_enabled(issuer_id)) {
229 if (not prev_state->is_actor_done(issuer_id)) {
230 prev_state->consider_one(issuer_id);
231 opened_states_.emplace_back(tmp_stack.back());
233 XBT_DEBUG("Actor %ld is already in done set: no need to explore it again", issuer_id);
235 XBT_DEBUG("Actor %ld is not enabled: DPOR may be failing. To stay sound, we are marking every enabled "
236 "transition as todo",
238 // If we ended up marking at least a transition, explore it at some point
239 if (prev_state->consider_all() > 0)
240 opened_states_.emplace_back(tmp_stack.back());
244 XBT_VERB("INDEPENDENT Transitions:");
245 XBT_VERB(" %s (state=%ld)", prev_state->get_transition_out()->to_string().c_str(), prev_state->get_num());
246 XBT_VERB(" %s (state=%ld)", state->get_transition_out()->to_string().c_str(), state->get_num());
248 tmp_stack.pop_back();
250 } else if (reduction_mode_ == ReductionMode::sdpor) {
252 * SDPOR Source Set Procedure:
254 * Find "reversible races" in the current execution `E` with respect
255 * to the latest action `p`. For each such race, determine one thread
256 * not contained in the backtrack set at the "race point" `r` which
257 * "represents" the trace formed by first executing everything after
258 * `r` that doesn't depend on it (`v := notdep(r, E)`) and then `p` to
261 * The intuition is that some subsequence of `v` may enable `p`, so
262 * we want to be sure that search "in that direction"
264 execution_seq_.push_transition(executed_transition.get());
266 xbt_assert(execution_seq_.get_latest_event_handle().has_value(),
267 "No events are contained in the SDPOR/OPDPOR execution "
268 "even though one was just added");
269 const aid_t p = executed_transition->aid_;
270 const auto next_E_p = execution_seq_.get_latest_event_handle().value();
272 for (const auto racing_event_handle : execution_seq_.get_racing_events_of(next_E_p)) {
273 // To determine if the race is reversible, we have to ensure
274 // that actor `p` running `next_E_p` (viz. the event such that
275 // `racing_event -> (E_p) next_E_p` and no other event
276 // "happens-between" the two) is enabled in any equivalent
277 // execution where `racing_event` happens before `next_E_p`.
279 // Importantly, it is equivalent to checking if in ANY
280 // such equivalent execution sequence where `racing_event`
281 // happens-before `next_E_p` that `p` is enabled in `pre(racing_event, E.p)`.
282 // Thus it suffices to check THIS execution
284 // If the actor `p` is not enabled at s_[E'], it is not a *reversible* race
285 const std::shared_ptr<State> prev_state = stack_[racing_event_handle];
286 if (prev_state->is_actor_enabled(p)) {
287 // NOTE: To incorporate the idea of attempting to select the "best"
288 // backtrack point into SDPOR, instead of selecting the `first` initial,
289 // we should instead compute all choices and decide which is bes
290 const std::optional<aid_t> q =
291 execution_seq_.get_first_sdpor_initial_from(racing_event_handle, prev_state->get_backtrack_set());
293 prev_state->consider_one(q.value());
294 opened_states_.emplace_back(std::move(prev_state));
300 // Before leaving that state, if the transition we just took can be taken multiple times, we
301 // need to give it to the opened states
302 if (stack_.back()->count_todo_multiples() > 0)
303 opened_states_.emplace_back(stack_.back());
305 if (_sg_mc_termination)
306 this->check_non_termination(next_state.get());
308 #if SIMGRID_HAVE_STATEFUL_MC
309 /* Check whether we already explored next_state in the past (but only if interested in state-equality reduction)
311 if (_sg_mc_max_visited_states > 0)
312 visited_state_ = visited_states_.addVisitedState(next_state->get_num(), next_state.get(), get_remote_app());
315 stack_.emplace_back(std::move(next_state));
317 /* If this is a new state (or if we don't care about state-equality reduction) */
318 if (visited_state_ == nullptr) {
319 /* Get an enabled process and insert it in the interleave set of the next state */
320 if (reduction_mode_ == ReductionMode::dpor)
321 stack_.back()->consider_best(); // Take only one transition if DPOR: others may be considered later if required
323 stack_.back()->consider_all();
326 dot_output("\"%ld\" -> \"%ld\" [%s];\n", state->get_num(), stack_.back()->get_num(),
327 state->get_transition_out()->dot_string().c_str());
328 #if SIMGRID_HAVE_STATEFUL_MC
330 dot_output("\"%ld\" -> \"%ld\" [%s];\n", state->get_num(),
331 visited_state_->original_num_ == -1 ? visited_state_->num_ : visited_state_->original_num_,
332 state->get_transition_out()->dot_string().c_str());
339 std::shared_ptr<State> DFSExplorer::best_opened_state()
341 int best_prio = 0; // cache the value for the best priority found so far (initialized to silence gcc)
342 auto best = end(opened_states_); // iterator to the state to explore having the best priority
343 auto valid = begin(opened_states_); // iterator marking the limit between states still to explore, and already
346 // Keep only still non-explored states (aid != -1), and record the one with the best (greater) priority.
347 for (auto current = begin(opened_states_); current != end(opened_states_); ++current) {
348 auto [aid, prio] = (*current)->next_transition_guided();
351 if (valid != current)
352 *valid = std::move(*current);
353 if (best == end(opened_states_) || prio > best_prio) {
360 std::shared_ptr<State> best_state;
362 // There are non-explored states, and one of them has the best priority. Remove it from opened_states_ before
364 best_state = std::move(*best);
367 *best = std::move(*valid);
369 opened_states_.erase(valid, end(opened_states_));
374 void DFSExplorer::backtrack()
376 XBT_VERB("Backtracking from %s", get_record_trace().to_string().c_str());
377 XBT_DEBUG("%lu alternatives are yet to be explored:", opened_states_.size());
379 on_backtracking_signal(get_remote_app());
380 get_remote_app().check_deadlock();
382 // Take the point with smallest distance
383 auto backtracking_point = best_opened_state();
385 // if no backtracking point, then set the stack_ to empty so we can end the exploration
386 if (not backtracking_point) {
387 XBT_DEBUG("No more opened point of exploration, the search will end");
392 // We found a backtracking point, let's go to it
394 XBT_DEBUG("Backtracking to state#%ld", backtracking_point->get_num());
396 #if SIMGRID_HAVE_STATEFUL_MC
397 /* If asked to rollback on a state that has a snapshot, restore it */
398 if (const auto* system_state = backtracking_point->get_system_state()) {
399 system_state->restore(*get_remote_app().get_remote_process_memory());
400 on_restore_system_state_signal(backtracking_point.get(), get_remote_app());
401 this->restore_stack(backtracking_point);
406 // Search how to restore the backtracking point
407 State* init_state = nullptr;
408 std::deque<Transition*> replay_recipe;
409 for (auto* s = backtracking_point.get(); s != nullptr; s = s->get_parent_state().get()) {
410 #if SIMGRID_HAVE_STATEFUL_MC
411 if (s->get_system_state() != nullptr) { // Found a state that I can restore
416 if (s->get_transition_in() != nullptr) // The root has no transition_in
417 replay_recipe.push_front(s->get_transition_in().get());
420 // Restore the init_state, if any
421 if (init_state != nullptr) {
422 #if SIMGRID_HAVE_STATEFUL_MC
423 const auto* system_state = init_state->get_system_state();
424 system_state->restore(*get_remote_app().get_remote_process_memory());
425 on_restore_system_state_signal(init_state, get_remote_app());
427 } else { // Restore the initial state if no intermediate state was found
428 get_remote_app().restore_initial_state();
429 on_restore_initial_state_signal(get_remote_app());
432 /* if no snapshot, we need to restore the initial state and replay the transitions */
433 /* Traverse the stack from the state at position start and re-execute the transitions */
434 for (auto& transition : replay_recipe) {
435 transition->replay(get_remote_app());
436 on_transition_replay_signal(transition, get_remote_app());
437 visited_states_count_++;
439 this->restore_stack(backtracking_point);
442 DFSExplorer::DFSExplorer(const std::vector<char*>& args, ReductionMode mode, bool need_memory_info)
443 : Exploration(args, need_memory_info || _sg_mc_termination
444 #if SIMGRID_HAVE_STATEFUL_MC
445 || _sg_mc_checkpoint > 0
448 , reduction_mode_(mode)
450 if (_sg_mc_termination) {
451 if (mode != ReductionMode::none) {
452 XBT_INFO("Check non progressive cycles (turning DPOR off)");
453 reduction_mode_ = ReductionMode::none;
455 XBT_INFO("Check non progressive cycles");
458 XBT_INFO("Start a DFS exploration. Reduction is: %s.", to_c_str(reduction_mode_));
460 auto initial_state = std::make_shared<State>(get_remote_app());
462 XBT_DEBUG("**************************************************");
464 stack_.emplace_back(std::move(initial_state));
466 /* Get an enabled actor and insert it in the interleave set of the initial state */
467 XBT_DEBUG("Initial state. %lu actors to consider", stack_.back()->get_actor_count());
468 if (reduction_mode_ == ReductionMode::dpor)
469 stack_.back()->consider_best();
471 stack_.back()->consider_all();
473 if (stack_.back()->count_todo_multiples() > 1)
474 opened_states_.emplace_back(stack_.back());
477 Exploration* create_dfs_exploration(const std::vector<char*>& args, ReductionMode mode)
479 return new DFSExplorer(args, mode);
482 } // namespace simgrid::mc