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 || reduction_mode_ == ReductionMode::odpor) {
97 execution_seq_ = odpor::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());
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 /!\\");
146 } else if (reduction_mode_ == ReductionMode::sdpor || reduction_mode_ == ReductionMode::odpor) {
147 XBT_ERROR("/!\\ Max depth of %d reached! THIS **WILL** BREAK the reduction, which is not sound "
148 "when stopping at a fixed depth /!\\",
149 _sg_mc_max_depth.get());
150 XBT_ERROR("/!\\ If bad things happen, disable dpor with --cfg=model-check/reduction:none /!\\");
152 XBT_WARN("/!\\ Max depth reached ! /!\\ ");
158 #if SIMGRID_HAVE_STATEFUL_MC
159 // Backtrack if we are revisiting a state we saw previously while applying state-equality reduction
160 if (visited_state_ != nullptr) {
161 XBT_DEBUG("State already visited (equal to state %ld), exploration stopped on this path.",
162 visited_state_->original_num_ == -1 ? visited_state_->num_ : visited_state_->original_num_);
164 visited_state_ = nullptr;
170 if (reduction_mode_ == ReductionMode::odpor) {
171 // In the case of ODPOR, the wakeup tree for this
172 // state may be empty if we're exploring new territory
173 // (rather than following the partial execution of a
174 // wakeup tree). This corresponds to lines 9 to 13 of
175 // the ODPOR pseudocode
177 // INVARIANT: The execution sequence should be consistent
178 // with the state when seeding the tree. If the sequence
179 // gets out of sync with the state, selection will not
181 state->seed_wakeup_tree_if_needed(execution_seq_);
184 // Search for the next transition
185 // next_transition returns a pair<aid_t, int>
186 // in case we want to consider multiple states (eg. during backtrack)
187 const aid_t next = reduction_mode_ == ReductionMode::odpor ? state->next_odpor_transition()
188 : std::get<0>(state->next_transition_guided());
190 if (next < 0) { // If there is no more transition in the current state, backtrack.
191 XBT_VERB("%lu actors remain, but none of them need to be interleaved (depth %zu).", state->get_actor_count(),
194 if (state->get_actor_count() == 0) {
195 get_remote_app().finalize_app();
196 XBT_VERB("Execution came to an end at %s (state: %ld, depth: %zu)", get_record_trace().to_string().c_str(),
197 state->get_num(), stack_.size());
204 if (_sg_mc_sleep_set && XBT_LOG_ISENABLED(mc_dfs, xbt_log_priority_verbose)) {
205 XBT_VERB("Sleep set actually containing:");
206 for (auto& [aid, transition] : state->get_sleep_set())
207 XBT_VERB(" <%ld,%s>", aid, transition.to_string().c_str());
210 /* Actually answer the request: let's execute the selected request (MCed does one step) */
211 const auto executed_transition = state->execute_next(next, get_remote_app());
212 on_transition_execute_signal(state->get_transition_out().get(), get_remote_app());
214 // If there are processes to interleave and the maximum depth has not been
215 // reached then perform one step of the exploration algorithm.
216 XBT_VERB("Execute %ld: %.60s (stack depth: %zu, state: %ld, %zu interleaves)", state->get_transition_out()->aid_,
217 state->get_transition_out()->to_string().c_str(), stack_.size(), state->get_num(), state->count_todo());
219 /* Create the new expanded state (copy the state of MCed into our MCer data) */
220 auto next_state = std::make_shared<State>(get_remote_app(), state);
221 on_state_creation_signal(next_state.get(), get_remote_app());
223 /* Sleep set procedure:
224 * adding the taken transition to the sleep set of the original state.
225 * <!> Since the parent sleep set is used to compute the child sleep set, this need to be
226 * done after next_state creation */
227 XBT_DEBUG("Marking Transition >>%s<< of process %ld done and adding it to the sleep set",
228 state->get_transition_out()->to_string().c_str(), state->get_transition_out()->aid_);
229 state->add_sleep_set(state->get_transition_out()); // Actors are marked done when they are considered in ActorState
231 if (reduction_mode_ == ReductionMode::odpor) {
232 // With ODPOR, after taking a step forward, we must:
233 // 1. remove the subtree whose root is a single-process
234 // node of actor `next` (viz. the action we took) from
235 // the wakeup tree of `state`
237 // 2. assign a copy of that subtree to the next state
239 // The latter evidently must be done BEFORE the former
240 next_state->sprout_tree_from_parent_state();
241 state->remove_subtree_starting_with(next);
243 // TODO: Consider what we have to do to handle transitions
244 // with multiple possible executions. We probably have to re-insert
245 // something into `state` and make note of that for later (opened_states_)
248 /* DPOR persistent set procedure:
249 * for each new transition considered, check if it depends on any other previous transition executed before it
250 * on another process. If there exists one, find the more recent, and add its process to the interleave set.
251 * If the process is not enabled at this point, then add every enabled process to the interleave */
252 if (reduction_mode_ == ReductionMode::dpor) {
253 aid_t issuer_id = state->get_transition_out()->aid_;
254 stack_t tmp_stack = stack_;
255 while (not tmp_stack.empty()) {
256 if (const State* prev_state = tmp_stack.back().get();
257 state->get_transition_out()->aid_ == prev_state->get_transition_out()->aid_) {
258 XBT_DEBUG("Simcall >>%s<< and >>%s<< with same issuer %ld", state->get_transition_out()->to_string().c_str(),
259 prev_state->get_transition_out()->to_string().c_str(), issuer_id);
260 tmp_stack.pop_back();
262 } else if (prev_state->get_transition_out()->depends(state->get_transition_out().get())) {
263 XBT_VERB("Dependent Transitions:");
264 XBT_VERB(" %s (state=%ld)", prev_state->get_transition_out()->to_string().c_str(), prev_state->get_num());
265 XBT_VERB(" %s (state=%ld)", state->get_transition_out()->to_string().c_str(), state->get_num());
267 if (prev_state->is_actor_enabled(issuer_id)) {
268 if (not prev_state->is_actor_done(issuer_id)) {
269 prev_state->consider_one(issuer_id);
270 opened_states_.emplace_back(tmp_stack.back());
272 XBT_DEBUG("Actor %ld is already in done set: no need to explore it again", issuer_id);
274 XBT_DEBUG("Actor %ld is not enabled: DPOR may be failing. To stay sound, we are marking every enabled "
275 "transition as todo",
277 // If we ended up marking at least a transition, explore it at some point
278 if (prev_state->consider_all() > 0)
279 opened_states_.emplace_back(tmp_stack.back());
283 XBT_VERB("INDEPENDENT Transitions:");
284 XBT_VERB(" %s (state=%ld)", prev_state->get_transition_out()->to_string().c_str(), prev_state->get_num());
285 XBT_VERB(" %s (state=%ld)", state->get_transition_out()->to_string().c_str(), state->get_num());
287 tmp_stack.pop_back();
289 } else if (reduction_mode_ == ReductionMode::sdpor) {
291 * SDPOR Source Set Procedure:
293 * Find "reversible races" in the current execution `E` with respect
294 * to the latest action `p`. For each such race, determine one thread
295 * not contained in the backtrack set at the "race point" `r` which
296 * "represents" the trace formed by first executing everything after
297 * `r` that doesn't depend on it (`v := notdep(r, E)`) and then `p` to
300 * The intuition is that some subsequence of `v` may enable `p`, so
301 * we want to be sure that search "in that direction"
303 const aid_t p = executed_transition->aid_;
304 execution_seq_.push_transition(std::move(executed_transition));
305 xbt_assert(execution_seq_.get_latest_event_handle().has_value(),
306 "No events are contained in the SDPOR/OPDPOR execution "
307 "even though one was just added");
309 const auto next_E_p = execution_seq_.get_latest_event_handle().value();
310 for (const auto racing_event_handle : execution_seq_.get_racing_events_of(next_E_p)) {
311 // To determine if the race is reversible, we have to ensure
312 // that actor `p` running `next_E_p` (viz. the event such that
313 // `racing_event -> (E_p) next_E_p` and no other event
314 // "happens-between" the two) is enabled in any equivalent
315 // execution where `racing_event` happens before `next_E_p`.
317 // Importantly, it is equivalent to checking if in ANY
318 // such equivalent execution sequence where `racing_event`
319 // happens-before `next_E_p` that `p` is enabled in `pre(racing_event, E.p)`.
320 // Thus it suffices to check THIS execution
322 // If the actor `p` is not enabled at s_[E'], it is not a *reversible* race
323 const std::shared_ptr<State> prev_state = stack_[racing_event_handle];
324 if (prev_state->is_actor_enabled(p)) {
325 // NOTE: To incorporate the idea of attempting to select the "best"
326 // backtrack point into SDPOR, instead of selecting the `first` initial,
327 // we should instead compute all choices and decide which is best
328 const std::optional<aid_t> q =
329 execution_seq_.get_first_sdpor_initial_from(racing_event_handle, prev_state->get_backtrack_set());
331 prev_state->consider_one(q.value());
332 opened_states_.emplace_back(std::move(prev_state));
336 } else if (reduction_mode_ == ReductionMode::odpor) {
337 // In the case of ODPOR, we simply observe the transition that was executed
338 // until we've reached a maximal trace
339 execution_seq_.push_transition(std::move(executed_transition));
342 // Before leaving that state, if the transition we just took can be taken multiple times, we
343 // need to give it to the opened states
344 if (stack_.back()->count_todo_multiples() > 0)
345 opened_states_.emplace_back(stack_.back());
347 if (_sg_mc_termination)
348 this->check_non_termination(next_state.get());
350 #if SIMGRID_HAVE_STATEFUL_MC
351 /* Check whether we already explored next_state in the past (but only if interested in state-equality reduction)
353 if (_sg_mc_max_visited_states > 0)
354 visited_state_ = visited_states_.addVisitedState(next_state->get_num(), next_state.get(), get_remote_app());
357 stack_.emplace_back(std::move(next_state));
359 /* If this is a new state (or if we don't care about state-equality reduction) */
360 if (visited_state_ == nullptr) {
361 /* Get an enabled process and insert it in the interleave set of the next state */
362 if (reduction_mode_ == ReductionMode::dpor)
363 stack_.back()->consider_best(); // Take only one transition if DPOR: others may be considered later if required
365 stack_.back()->consider_all();
368 dot_output("\"%ld\" -> \"%ld\" [%s];\n", state->get_num(), stack_.back()->get_num(),
369 state->get_transition_out()->dot_string().c_str());
370 #if SIMGRID_HAVE_STATEFUL_MC
372 dot_output("\"%ld\" -> \"%ld\" [%s];\n", state->get_num(),
373 visited_state_->original_num_ == -1 ? visited_state_->num_ : visited_state_->original_num_,
374 state->get_transition_out()->dot_string().c_str());
381 std::shared_ptr<State> DFSExplorer::best_opened_state()
383 if (reduction_mode_ == ReductionMode::odpor) {
385 std::find_if(stack_.rbegin(), stack_.rend(), [](const auto& state) { return !state->has_empty_tree(); });
389 int best_prio = 0; // cache the value for the best priority found so far (initialized to silence gcc)
390 auto best = end(opened_states_); // iterator to the state to explore having the best priority
391 auto valid = begin(opened_states_); // iterator marking the limit between states still to explore, and already
394 // Keep only still non-explored states (aid != -1), and record the one with the best (greater) priority.
395 for (auto current = begin(opened_states_); current != end(opened_states_); ++current) {
396 auto [aid, prio] = (*current)->next_transition_guided();
399 if (valid != current)
400 *valid = std::move(*current);
401 if (best == end(opened_states_) || prio > best_prio) {
408 std::shared_ptr<State> best_state;
410 // There are non-explored states, and one of them has the best priority. Remove it from opened_states_ before
412 best_state = std::move(*best);
415 *best = std::move(*valid);
417 opened_states_.erase(valid, end(opened_states_));
422 void DFSExplorer::backtrack()
424 if (const auto last_event = execution_seq_.get_latest_event_handle();
425 reduction_mode_ == ReductionMode::odpor and last_event.has_value()) {
427 * ODPOR Race Detection Procedure:
429 * For each reversible race in the current execution, we
430 * note if there are any continuations `C` equivalent to that which
431 * would reverse the race that have already either a) been searched by ODPOR or
432 * b) been *noted* to be searched by the wakeup tree at the
433 * appropriate reversal point, either as `C` directly or
434 * an as equivalent to `C` ("eventually looks like C", viz. the `~_E`
437 for (auto e_prime = static_cast<odpor::Execution::EventHandle>(0); e_prime <= last_event; e_prime++) {
438 for (const auto e : execution_seq_.get_racing_events_of(e_prime)) {
439 // To determine if the race is reversible, we have to ensure
440 // that actor `p` running `e_i` (viz. the event such that
441 // `racing_event -> (E_p) e_i` and no other event
442 // "happens-between" the two) is enabled in any equivalent
443 // execution where `racing_event` happens before `e_i`.
445 // Importantly, it is equivalent to checking if in ANY
446 // such equivalent execution sequence where `racing_event`
447 // happens-before `next_E_p` that `p` is enabled in `pre(racing_event, E.p)`.
448 // Thus it suffices to check THIS execution
450 // If the actor `p` is not enabled at s_[E'], it is not a *reversible* race
451 const aid_t p = execution_seq_.get_actor_with_handle(e_prime);
452 State& prev_state = *stack_[e];
453 if (prev_state.is_actor_enabled(p)) {
454 const std::optional<odpor::PartialExecution> v =
455 execution_seq_.get_odpor_extension_from(e, e_prime, prev_state);
457 prev_state.mark_path_interesting_for_odpor(v.value(), execution_seq_.get_prefix_before(e));
464 XBT_VERB("Backtracking from %s", get_record_trace().to_string().c_str());
465 XBT_DEBUG("%lu alternatives are yet to be explored:", opened_states_.size());
467 on_backtracking_signal(get_remote_app());
468 get_remote_app().check_deadlock();
470 // Take the point with smallest distance
471 auto backtracking_point = best_opened_state();
473 // if no backtracking point, then set the stack_ to empty so we can end the exploration
474 if (not backtracking_point) {
475 XBT_DEBUG("No more opened point of exploration, the search will end");
480 // We found a backtracking point, let's go to it
482 XBT_DEBUG("Backtracking to state#%ld", backtracking_point->get_num());
484 #if SIMGRID_HAVE_STATEFUL_MC
485 /* If asked to rollback on a state that has a snapshot, restore it */
486 if (const auto* system_state = backtracking_point->get_system_state()) {
487 system_state->restore(*get_remote_app().get_remote_process_memory());
488 on_restore_system_state_signal(backtracking_point.get(), get_remote_app());
489 this->restore_stack(backtracking_point);
494 // Search how to restore the backtracking point
495 State* init_state = nullptr;
496 std::deque<Transition*> replay_recipe;
497 for (auto* s = backtracking_point.get(); s != nullptr; s = s->get_parent_state().get()) {
498 #if SIMGRID_HAVE_STATEFUL_MC
499 if (s->get_system_state() != nullptr) { // Found a state that I can restore
504 if (s->get_transition_in() != nullptr) // The root has no transition_in
505 replay_recipe.push_front(s->get_transition_in().get());
508 // Restore the init_state, if any
509 if (init_state != nullptr) {
510 #if SIMGRID_HAVE_STATEFUL_MC
511 const auto* system_state = init_state->get_system_state();
512 system_state->restore(*get_remote_app().get_remote_process_memory());
513 on_restore_system_state_signal(init_state, get_remote_app());
515 } else { // Restore the initial state if no intermediate state was found
516 get_remote_app().restore_initial_state();
517 on_restore_initial_state_signal(get_remote_app());
520 /* if no snapshot, we need to restore the initial state and replay the transitions */
521 /* Traverse the stack from the state at position start and re-execute the transitions */
522 for (auto& transition : replay_recipe) {
523 transition->replay(get_remote_app());
524 on_transition_replay_signal(transition, get_remote_app());
525 visited_states_count_++;
527 this->restore_stack(backtracking_point);
530 DFSExplorer::DFSExplorer(const std::vector<char*>& args, ReductionMode mode, bool need_memory_info)
531 : Exploration(args, need_memory_info || _sg_mc_termination
532 #if SIMGRID_HAVE_STATEFUL_MC
533 || _sg_mc_checkpoint > 0
536 , reduction_mode_(mode)
538 if (_sg_mc_termination) {
539 if (mode != ReductionMode::none) {
540 XBT_INFO("Check non progressive cycles (turning DPOR off)");
541 reduction_mode_ = ReductionMode::none;
543 XBT_INFO("Check non progressive cycles");
546 XBT_INFO("Start a DFS exploration. Reduction is: %s.", to_c_str(reduction_mode_));
548 auto initial_state = std::make_shared<State>(get_remote_app());
550 XBT_DEBUG("**************************************************");
552 stack_.emplace_back(std::move(initial_state));
554 /* Get an enabled actor and insert it in the interleave set of the initial state */
555 XBT_DEBUG("Initial state. %lu actors to consider", stack_.back()->get_actor_count());
556 if (reduction_mode_ == ReductionMode::dpor)
557 stack_.back()->consider_best();
559 stack_.back()->consider_all();
561 if (stack_.back()->count_todo_multiples() > 1)
562 opened_states_.emplace_back(stack_.back());
565 Exploration* create_dfs_exploration(const std::vector<char*>& args, ReductionMode mode)
567 return new DFSExplorer(args, mode);
570 } // namespace simgrid::mc