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 execution_seq_ = odpor::Execution();
88 auto current_state = state;
89 stack_.emplace_front(current_state);
90 // condition corresponds to reaching initial state
91 while (current_state->get_parent_state() != nullptr) {
92 current_state = current_state->get_parent_state();
93 stack_.emplace_front(current_state);
95 XBT_DEBUG("Replaced stack by %s", get_record_trace().to_string().c_str());
96 if (reduction_mode_ == ReductionMode::sdpor || reduction_mode_ == ReductionMode::odpor) {
97 // NOTE: The outgoing transition for the top-most
98 // state of the stack refers to that which was taken
99 // as part of the last trace explored by the algorithm.
100 // Thus, only the sequence of transitions leading up to,
101 // but not including, the last state must be included
102 // when reconstructing the Exploration for SDPOR.
103 for (auto iter = std::next(stack_.begin()); iter != stack_.end(); ++iter) {
104 execution_seq_.push_transition((*iter)->get_transition_in());
106 XBT_DEBUG("Replaced SDPOR/ODPOR execution to reflect the new stack");
110 void DFSExplorer::log_state() // override
112 on_log_state_signal(get_remote_app());
113 XBT_INFO("DFS exploration ended. %ld unique states visited; %lu backtracks (%lu transition replays, %lu states "
115 State::get_expanded_states(), backtrack_count_, visited_states_count_,
116 Transition::get_replayed_transitions());
117 Exploration::log_state();
120 void DFSExplorer::run()
122 on_exploration_start_signal(get_remote_app());
123 /* This function runs the DFS algorithm the state space.
124 * We do so iteratively instead of recursively, dealing with the call stack manually.
125 * This allows one to explore the call stack at will. */
127 while (not stack_.empty()) {
128 /* Get current state */
129 auto state = stack_.back();
131 XBT_DEBUG("**************************************************");
132 XBT_DEBUG("Exploration depth=%zu (state:#%ld; %zu interleaves todo)", stack_.size(), state->get_num(),
133 state->count_todo());
135 visited_states_count_++;
137 // Backtrack if we reached the maximum depth
138 if (stack_.size() > (std::size_t)_sg_mc_max_depth) {
139 if (reduction_mode_ == ReductionMode::dpor) {
140 XBT_ERROR("/!\\ Max depth of %d reached! THIS WILL PROBABLY BREAK the dpor reduction /!\\",
141 _sg_mc_max_depth.get());
142 XBT_ERROR("/!\\ If bad things happen, disable dpor with --cfg=model-check/reduction:none /!\\");
143 } else if (reduction_mode_ == ReductionMode::sdpor || reduction_mode_ == ReductionMode::odpor) {
144 XBT_ERROR("/!\\ Max depth of %d reached! THIS **WILL** BREAK the reduction, which is not sound "
145 "when stopping at a fixed depth /!\\",
146 _sg_mc_max_depth.get());
147 XBT_ERROR("/!\\ If bad things happen, disable the reduction with --cfg=model-check/reduction:none /!\\");
149 XBT_WARN("/!\\ Max depth reached ! /!\\ ");
155 #if SIMGRID_HAVE_STATEFUL_MC
156 // Backtrack if we are revisiting a state we saw previously while applying state-equality reduction
157 if (visited_state_ != nullptr) {
158 XBT_DEBUG("State already visited (equal to state %ld), exploration stopped on this path.",
159 visited_state_->original_num_ == -1 ? visited_state_->num_ : visited_state_->original_num_);
161 visited_state_ = nullptr;
167 if (reduction_mode_ == ReductionMode::odpor) {
168 // In the case of ODPOR, the wakeup tree for this
169 // state may be empty if we're exploring new territory
170 // (rather than following the partial execution of a
171 // wakeup tree). This corresponds to lines 9 to 13 of
172 // the ODPOR pseudocode
174 // INVARIANT: The execution sequence should be consistent
175 // with the state when seeding the tree. If the sequence
176 // gets out of sync with the state, selection will not
178 state->seed_wakeup_tree_if_needed(execution_seq_);
181 // Search for the next transition
182 // next_transition returns a pair<aid_t, int>
183 // in case we want to consider multiple states (eg. during backtrack)
184 const aid_t next = reduction_mode_ == ReductionMode::odpor ? state->next_odpor_transition()
185 : std::get<0>(state->next_transition_guided());
187 if (next < 0) { // If there is no more transition in the current state, backtrack.
188 XBT_VERB("%lu actors remain, but none of them need to be interleaved (depth %zu).", state->get_actor_count(),
191 if (state->get_actor_count() == 0) {
192 get_remote_app().finalize_app();
193 XBT_VERB("Execution came to an end at %s (state: %ld, depth: %zu)", get_record_trace().to_string().c_str(),
194 state->get_num(), stack_.size());
201 if (_sg_mc_sleep_set && XBT_LOG_ISENABLED(mc_dfs, xbt_log_priority_verbose)) {
202 XBT_VERB("Sleep set actually containing:");
203 for (auto& [aid, transition] : state->get_sleep_set())
204 XBT_VERB(" <%ld,%s>", aid, transition->to_string().c_str());
207 /* Actually answer the request: let's execute the selected request (MCed does one step) */
208 const auto executed_transition = state->execute_next(next, get_remote_app());
209 on_transition_execute_signal(state->get_transition_out().get(), get_remote_app());
211 // If there are processes to interleave and the maximum depth has not been
212 // reached then perform one step of the exploration algorithm.
213 XBT_VERB("Execute %ld: %.60s (stack depth: %zu, state: %ld, %zu interleaves)", state->get_transition_out()->aid_,
214 state->get_transition_out()->to_string().c_str(), stack_.size(), state->get_num(), state->count_todo());
216 /* Create the new expanded state (copy the state of MCed into our MCer data) */
217 auto next_state = std::make_shared<State>(get_remote_app(), state);
218 on_state_creation_signal(next_state.get(), get_remote_app());
220 if (reduction_mode_ == ReductionMode::odpor) {
221 // With ODPOR, after taking a step forward, we must
222 // assign a copy of that subtree to the next state.
224 // NOTE: We only add actions to the sleep set AFTER
225 // we've regenerated states. We must perform the search
226 // fully down a single path before we consider adding
227 // any elements to the sleep set according to the pseudocode
228 next_state->sprout_tree_from_parent_state();
230 /* Sleep set procedure:
231 * adding the taken transition to the sleep set of the original state.
232 * <!> Since the parent sleep set is used to compute the child sleep set, this need to be
233 * done after next_state creation */
234 XBT_DEBUG("Marking Transition >>%s<< of process %ld done and adding it to the sleep set",
235 state->get_transition_out()->to_string().c_str(), state->get_transition_out()->aid_);
236 state->add_sleep_set(
237 state->get_transition_out()); // Actors are marked done when they are considered in ActorState
240 /* DPOR persistent set procedure:
241 * for each new transition considered, check if it depends on any other previous transition executed before it
242 * on another process. If there exists one, find the more recent, and add its process to the interleave set.
243 * If the process is not enabled at this point, then add every enabled process to the interleave */
244 if (reduction_mode_ == ReductionMode::dpor) {
245 aid_t issuer_id = state->get_transition_out()->aid_;
246 stack_t tmp_stack = stack_;
247 while (not tmp_stack.empty()) {
248 if (const State* prev_state = tmp_stack.back().get();
249 state->get_transition_out()->aid_ == prev_state->get_transition_out()->aid_) {
250 XBT_DEBUG("Simcall >>%s<< and >>%s<< with same issuer %ld", state->get_transition_out()->to_string().c_str(),
251 prev_state->get_transition_out()->to_string().c_str(), issuer_id);
252 tmp_stack.pop_back();
254 } else if (prev_state->get_transition_out()->depends(state->get_transition_out().get())) {
255 XBT_VERB("Dependent Transitions:");
256 XBT_VERB(" %s (state=%ld)", prev_state->get_transition_out()->to_string().c_str(), prev_state->get_num());
257 XBT_VERB(" %s (state=%ld)", state->get_transition_out()->to_string().c_str(), state->get_num());
259 if (prev_state->is_actor_enabled(issuer_id)) {
260 if (not prev_state->is_actor_done(issuer_id)) {
261 prev_state->consider_one(issuer_id);
262 opened_states_.emplace_back(tmp_stack.back());
264 XBT_DEBUG("Actor %ld is already in done set: no need to explore it again", issuer_id);
266 XBT_DEBUG("Actor %ld is not enabled: DPOR may be failing. To stay sound, we are marking every enabled "
267 "transition as todo",
269 // If we ended up marking at least a transition, explore it at some point
270 if (prev_state->consider_all() > 0)
271 opened_states_.emplace_back(tmp_stack.back());
275 XBT_VERB("INDEPENDENT Transitions:");
276 XBT_VERB(" %s (state=%ld)", prev_state->get_transition_out()->to_string().c_str(), prev_state->get_num());
277 XBT_VERB(" %s (state=%ld)", state->get_transition_out()->to_string().c_str(), state->get_num());
279 tmp_stack.pop_back();
281 } else if (reduction_mode_ == ReductionMode::sdpor) {
283 * SDPOR Source Set Procedure:
285 * Find "reversible races" in the current execution `E` with respect
286 * to the latest action `p`. For each such race, determine one thread
287 * not contained in the backtrack set at the "race point" `r` which
288 * "represents" the trace formed by first executing everything after
289 * `r` that doesn't depend on it (`v := notdep(r, E)`) and then `p` to
292 * The intuition is that some subsequence of `v` may enable `p`, so
293 * we want to be sure that search "in that direction"
295 execution_seq_.push_transition(std::move(executed_transition));
296 xbt_assert(execution_seq_.get_latest_event_handle().has_value(), "No events are contained in the SDPOR execution "
297 "even though one was just added");
299 const auto next_E_p = execution_seq_.get_latest_event_handle().value();
300 for (const auto e_race : execution_seq_.get_reversible_races_of(next_E_p)) {
301 State* prev_state = stack_[e_race].get();
302 const auto choices = execution_seq_.get_missing_source_set_actors_from(e_race, prev_state->get_backtrack_set());
303 if (!choices.empty()) {
304 // NOTE: To incorporate the idea of attempting to select the "best"
305 // backtrack point into SDPOR, instead of selecting the `first` initial,
306 // we should instead compute all choices and decide which is best
308 // Here, we choose the actor with the lowest ID to ensure
309 // we get deterministic results
311 std::min_element(choices.begin(), choices.end(), [](const aid_t a1, const aid_t a2) { return a1 < a2; });
312 prev_state->consider_one(*q);
313 opened_states_.emplace_back(std::move(prev_state));
316 } else if (reduction_mode_ == ReductionMode::odpor) {
317 // In the case of ODPOR, we simply observe the transition that was executed
318 // until we've reached a maximal trace
319 execution_seq_.push_transition(std::move(executed_transition));
322 // Before leaving that state, if the transition we just took can be taken multiple times, we
323 // need to give it to the opened states
324 if (stack_.back()->count_todo_multiples() > 0)
325 opened_states_.emplace_back(stack_.back());
327 if (_sg_mc_termination)
328 this->check_non_termination(next_state.get());
330 #if SIMGRID_HAVE_STATEFUL_MC
331 /* Check whether we already explored next_state in the past (but only if interested in state-equality reduction)
333 if (_sg_mc_max_visited_states > 0)
334 visited_state_ = visited_states_.addVisitedState(next_state->get_num(), next_state.get(), get_remote_app());
337 stack_.emplace_back(std::move(next_state));
339 /* If this is a new state (or if we don't care about state-equality reduction) */
340 if (visited_state_ == nullptr) {
341 /* Get an enabled process and insert it in the interleave set of the next state */
342 if (reduction_mode_ == ReductionMode::dpor)
343 stack_.back()->consider_best(); // Take only one transition if DPOR: others may be considered later if required
345 stack_.back()->consider_all();
348 dot_output("\"%ld\" -> \"%ld\" [%s];\n", state->get_num(), stack_.back()->get_num(),
349 state->get_transition_out()->dot_string().c_str());
350 #if SIMGRID_HAVE_STATEFUL_MC
352 dot_output("\"%ld\" -> \"%ld\" [%s];\n", state->get_num(),
353 visited_state_->original_num_ == -1 ? visited_state_->num_ : visited_state_->original_num_,
354 state->get_transition_out()->dot_string().c_str());
361 std::shared_ptr<State> DFSExplorer::best_opened_state()
363 int best_prio = 0; // cache the value for the best priority found so far (initialized to silence gcc)
364 auto best = end(opened_states_); // iterator to the state to explore having the best priority
365 auto valid = begin(opened_states_); // iterator marking the limit between states still to explore, and already
368 // Keep only still non-explored states (aid != -1), and record the one with the best (greater) priority.
369 for (auto current = begin(opened_states_); current != end(opened_states_); ++current) {
370 auto [aid, prio] = (*current)->next_transition_guided();
373 if (valid != current)
374 *valid = std::move(*current);
375 if (best == end(opened_states_) || prio > best_prio) {
382 std::shared_ptr<State> best_state;
384 // There are non-explored states, and one of them has the best priority. Remove it from opened_states_ before
386 best_state = std::move(*best);
389 *best = std::move(*valid);
391 opened_states_.erase(valid, end(opened_states_));
396 std::shared_ptr<State> DFSExplorer::next_odpor_state()
398 for (auto iter = stack_.rbegin(); iter != stack_.rend(); ++iter) {
399 const auto& state = *iter;
400 state->do_odpor_unwind();
401 XBT_DEBUG("\tPerformed ODPOR 'clean-up'. Sleep set has:");
402 for (auto& [aid, transition] : state->get_sleep_set())
403 XBT_DEBUG("\t <%ld,%s>", aid, transition->to_string().c_str());
404 if (!state->has_empty_tree()) {
411 void DFSExplorer::backtrack()
413 if (const auto last_event = execution_seq_.get_latest_event_handle();
414 reduction_mode_ == ReductionMode::odpor and last_event.has_value()) {
416 * ODPOR Race Detection Procedure:
418 * For each reversible race in the current execution, we
419 * note if there are any continuations `C` equivalent to that which
420 * would reverse the race that have already either a) been searched by ODPOR or
421 * b) been *noted* to be searched by the wakeup tree at the
422 * appropriate reversal point, either as `C` directly or
423 * an as equivalent to `C` ("eventually looks like C", viz. the `~_E`
426 for (auto e_prime = static_cast<odpor::Execution::EventHandle>(0); e_prime <= last_event.value(); ++e_prime) {
427 for (const auto e : execution_seq_.get_reversible_races_of(e_prime)) {
428 XBT_DEBUG("ODPOR: Reversible race detected between events `%u` and `%u`", e, e_prime);
429 State& prev_state = *stack_[e];
430 if (const auto v = execution_seq_.get_odpor_extension_from(e, e_prime, prev_state); v.has_value()) {
431 const auto result = prev_state.insert_into_wakeup_tree(v.value(), execution_seq_.get_prefix_before(e));
433 case odpor::WakeupTree::InsertionResult::root: {
434 XBT_DEBUG("ODPOR: Reversible race with `%u` unaccounted for in the wakeup tree for "
435 "the execution prior to event `%u`:",
439 case odpor::WakeupTree::InsertionResult::interior_node: {
440 XBT_DEBUG("ODPOR: Reversible race with `%u` partially accounted for in the wakeup tree for "
441 "the execution prior to event `%u`:",
445 case odpor::WakeupTree::InsertionResult::leaf: {
446 XBT_DEBUG("ODPOR: Reversible race with `%u` accounted for in the wakeup tree for "
447 "the execution prior to event `%u`:",
452 for (const auto& seq : simgrid::mc::odpor::get_textual_trace(v.value())) {
453 XBT_DEBUG(" %s", seq.c_str());
456 XBT_DEBUG("ODPOR: Ignoring race: `sleep(E')` intersects `WI_[E'](v := notdep(%u, E))`", e);
457 XBT_DEBUG("Sleep set contains:");
458 for (auto& [aid, transition] : prev_state.get_sleep_set())
459 XBT_DEBUG(" <%ld,%s>", aid, transition->to_string().c_str());
465 XBT_VERB("Backtracking from %s", get_record_trace().to_string().c_str());
466 XBT_DEBUG("%lu alternatives are yet to be explored:", opened_states_.size());
468 on_backtracking_signal(get_remote_app());
469 get_remote_app().check_deadlock();
471 // Take the point with smallest distance
472 auto backtracking_point = reduction_mode_ == ReductionMode::odpor ? next_odpor_state() : best_opened_state();
474 // if no backtracking point, then set the stack_ to empty so we can end the exploration
475 if (not backtracking_point) {
476 XBT_DEBUG("No more opened point of exploration, the search will end");
481 // We found a backtracking point, let's go to it
483 XBT_DEBUG("Backtracking to state#%ld", backtracking_point->get_num());
485 #if SIMGRID_HAVE_STATEFUL_MC
486 /* If asked to rollback on a state that has a snapshot, restore it */
487 if (const auto* system_state = backtracking_point->get_system_state()) {
488 system_state->restore(*get_remote_app().get_remote_process_memory());
489 on_restore_system_state_signal(backtracking_point.get(), get_remote_app());
490 this->restore_stack(backtracking_point);
495 // Search how to restore the backtracking point
496 State* init_state = nullptr;
497 std::deque<Transition*> replay_recipe;
498 for (auto* s = backtracking_point.get(); s != nullptr; s = s->get_parent_state().get()) {
499 #if SIMGRID_HAVE_STATEFUL_MC
500 if (s->get_system_state() != nullptr) { // Found a state that I can restore
505 if (s->get_transition_in() != nullptr) // The root has no transition_in
506 replay_recipe.push_front(s->get_transition_in().get());
509 // Restore the init_state, if any
510 if (init_state != nullptr) {
511 #if SIMGRID_HAVE_STATEFUL_MC
512 const auto* system_state = init_state->get_system_state();
513 system_state->restore(*get_remote_app().get_remote_process_memory());
514 on_restore_system_state_signal(init_state, get_remote_app());
516 } else { // Restore the initial state if no intermediate state was found
517 get_remote_app().restore_initial_state();
518 on_restore_initial_state_signal(get_remote_app());
521 /* if no snapshot, we need to restore the initial state and replay the transitions */
522 /* Traverse the stack from the state at position start and re-execute the transitions */
523 for (auto& transition : replay_recipe) {
524 transition->replay(get_remote_app());
525 on_transition_replay_signal(transition, get_remote_app());
526 visited_states_count_++;
528 this->restore_stack(backtracking_point);
531 DFSExplorer::DFSExplorer(const std::vector<char*>& args, ReductionMode mode, bool need_memory_info)
532 : Exploration(args, need_memory_info || _sg_mc_termination
533 #if SIMGRID_HAVE_STATEFUL_MC
534 || _sg_mc_checkpoint > 0
537 , reduction_mode_(mode)
539 if (_sg_mc_termination) {
540 if (mode != ReductionMode::none) {
541 XBT_INFO("Check non progressive cycles (turning DPOR off)");
542 reduction_mode_ = ReductionMode::none;
544 XBT_INFO("Check non progressive cycles");
547 XBT_INFO("Start a DFS exploration. Reduction is: %s.", to_c_str(reduction_mode_));
549 auto initial_state = std::make_shared<State>(get_remote_app());
551 XBT_DEBUG("**************************************************");
553 stack_.emplace_back(std::move(initial_state));
555 /* Get an enabled actor and insert it in the interleave set of the initial state */
556 XBT_DEBUG("Initial state. %lu actors to consider", stack_.back()->get_actor_count());
557 if (reduction_mode_ == ReductionMode::dpor)
558 stack_.back()->consider_best();
560 stack_.back()->consider_all();
562 if (stack_.back()->count_todo_multiples() > 1)
563 opened_states_.emplace_back(stack_.back());
565 if (mode == ReductionMode::odpor && !_sg_mc_sleep_set) {
566 // ODPOR requires the use of sleep sets; SDPOR
567 // "likes" using sleep sets but it is not strictly
569 XBT_INFO("Forcing the use of sleep sets for use with ODPOR");
570 _sg_mc_sleep_set = true;
574 Exploration* create_dfs_exploration(const std::vector<char*>& args, ReductionMode mode)
576 return new DFSExplorer(args, mode);
579 } // namespace simgrid::mc