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"
14 #include "xbt/string.hpp"
15 #include "xbt/sysdep.h"
23 #include <unordered_set>
26 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(mc_dfs, mc, "DFS exploration algorithm of the model-checker");
28 namespace simgrid::mc {
30 xbt::signal<void(RemoteApp&)> DFSExplorer::on_exploration_start_signal;
31 xbt::signal<void(RemoteApp&)> DFSExplorer::on_backtracking_signal;
33 xbt::signal<void(State*, RemoteApp&)> DFSExplorer::on_state_creation_signal;
35 xbt::signal<void(State*, RemoteApp&)> DFSExplorer::on_restore_system_state_signal;
36 xbt::signal<void(RemoteApp&)> DFSExplorer::on_restore_initial_state_signal;
37 xbt::signal<void(Transition*, RemoteApp&)> DFSExplorer::on_transition_replay_signal;
38 xbt::signal<void(Transition*, RemoteApp&)> DFSExplorer::on_transition_execute_signal;
40 xbt::signal<void(RemoteApp&)> DFSExplorer::on_log_state_signal;
42 RecordTrace DFSExplorer::get_record_trace() // override
46 if (const auto trans = stack_.back()->get_transition_out(); trans != nullptr)
47 res.push_back(trans.get());
48 for (const auto* state = stack_.back().get(); state != nullptr; state = state->get_parent_state().get())
49 if (state->get_transition_in() != nullptr)
50 res.push_front(state->get_transition_in().get());
55 void DFSExplorer::restore_stack(std::shared_ptr<State> state)
58 execution_seq_ = odpor::Execution();
59 auto current_state = state;
60 stack_.emplace_front(current_state);
61 // condition corresponds to reaching initial state
62 while (current_state->get_parent_state() != nullptr) {
63 current_state = current_state->get_parent_state();
64 stack_.emplace_front(current_state);
66 XBT_DEBUG("Replaced stack by %s", get_record_trace().to_string().c_str());
67 if (reduction_mode_ == ReductionMode::sdpor || reduction_mode_ == ReductionMode::odpor) {
68 // NOTE: The outgoing transition for the top-most state of the stack refers to that which was taken
69 // as part of the last trace explored by the algorithm. Thus, only the sequence of transitions leading up to,
70 // but not including, the last state must be included when reconstructing the Exploration for SDPOR.
71 for (auto iter = std::next(stack_.begin()); iter != stack_.end(); ++iter) {
72 execution_seq_.push_transition((*iter)->get_transition_in());
74 XBT_DEBUG("Replaced SDPOR/ODPOR execution to reflect the new stack");
78 void DFSExplorer::log_state() // override
80 on_log_state_signal(get_remote_app());
81 XBT_INFO("DFS exploration ended. %ld unique states visited; %lu backtracks (%lu transition replays, %lu states "
83 State::get_expanded_states(), backtrack_count_, Transition::get_replayed_transitions(),
84 visited_states_count_);
85 Exploration::log_state();
88 void DFSExplorer::run()
90 on_exploration_start_signal(get_remote_app());
91 /* This function runs the DFS algorithm the state space.
92 * We do so iteratively instead of recursively, dealing with the call stack manually.
93 * This allows one to explore the call stack at will. */
95 while (not stack_.empty()) {
96 /* Get current state */
97 auto state = stack_.back();
99 XBT_DEBUG("**************************************************");
100 XBT_DEBUG("Exploration depth=%zu (state:#%ld; %zu interleaves todo)", stack_.size(), state->get_num(),
101 state->count_todo());
103 visited_states_count_++;
105 // Backtrack if we reached the maximum depth
106 if (stack_.size() > (std::size_t)_sg_mc_max_depth) {
107 if (reduction_mode_ == ReductionMode::dpor) {
108 XBT_ERROR("/!\\ Max depth of %d reached! THIS WILL PROBABLY BREAK the dpor reduction /!\\",
109 _sg_mc_max_depth.get());
110 XBT_ERROR("/!\\ If bad things happen, disable dpor with --cfg=model-check/reduction:none /!\\");
111 } else if (reduction_mode_ == ReductionMode::sdpor || reduction_mode_ == ReductionMode::odpor) {
112 XBT_ERROR("/!\\ Max depth of %d reached! THIS **WILL** BREAK the reduction, which is not sound "
113 "when stopping at a fixed depth /!\\",
114 _sg_mc_max_depth.get());
115 XBT_ERROR("/!\\ If bad things happen, disable the reduction with --cfg=model-check/reduction:none /!\\");
117 XBT_WARN("/!\\ Max depth reached ! /!\\ ");
123 if (reduction_mode_ == ReductionMode::odpor) {
124 // In the case of ODPOR, the wakeup tree for this state may be empty if we're exploring new territory
125 // (rather than following the partial execution of a wakeup tree). This corresponds to lines 9 to 13 of
126 // the ODPOR pseudocode
128 // INVARIANT: The execution sequence should be consistent with the state when seeding the tree. If the sequence
129 // gets out of sync with the state, selection will not work as we intend
130 state->seed_wakeup_tree_if_needed(execution_seq_);
133 // Search for the next transition
134 // next_transition returns a pair<aid_t, int>
135 // in case we want to consider multiple states (eg. during backtrack)
136 const aid_t next = reduction_mode_ == ReductionMode::odpor ? state->next_odpor_transition()
137 : std::get<0>(state->next_transition_guided());
140 // If there is no more transition in the current state ), backtrace
141 XBT_VERB("%lu actors remain, but none of them need to be interleaved (depth %zu).", state->get_actor_count(),
144 if (state->get_actor_count() == 0) {
145 get_remote_app().finalize_app();
146 XBT_VERB("Execution came to an end at %s (state: %ld, depth: %zu)", get_record_trace().to_string().c_str(),
147 state->get_num(), stack_.size());
153 if (not state->is_actor_enabled(next)) {
154 // if ODPOR picked an actor that is not enabled -- ReversibleRace is an overapproximation
155 xbt_assert(reduction_mode_ == ReductionMode::odpor,
156 "Only ODPOR should be fool enough to try to execute a disabled transition");
157 XBT_VERB("Preventing ODPOR exploration from executing a disabled transition. The reversibility of the race "
158 "must have been overapproximated");
160 state->remove_subtree_at_aid(next);
161 state->add_sleep_set(state->get_actors_list().at(next).get_transition());
162 XBT_VERB("%lu actors remain, but none of them need to be interleaved (depth %zu).", state->get_actor_count(),
165 if (state->get_actor_count() == 0) {
166 get_remote_app().finalize_app();
167 XBT_VERB("Execution came to an end at %s (state: %ld, depth: %zu)", get_record_trace().to_string().c_str(),
168 state->get_num(), stack_.size());
175 if (XBT_LOG_ISENABLED(mc_dfs, xbt_log_priority_verbose)) {
176 XBT_VERB("Sleep set actually containing:");
177 for (const auto& [aid, transition] : state->get_sleep_set())
178 XBT_VERB(" <%ld,%s>", aid, transition->to_string().c_str());
181 auto todo = state->get_actors_list().at(next).get_transition();
182 XBT_DEBUG("wanna execute %ld: %.60s", next, todo->to_string().c_str());
184 /* Actually answer the request: let's execute the selected request (MCed does one step) */
185 auto executed_transition = state->execute_next(next, get_remote_app());
186 on_transition_execute_signal(state->get_transition_out().get(), get_remote_app());
188 // If there are processes to interleave and the maximum depth has not been
189 // reached then perform one step of the exploration algorithm.
190 XBT_VERB("Executed %ld: %.60s (stack depth: %zu, state: %ld, %zu interleaves)", state->get_transition_out()->aid_,
191 state->get_transition_out()->to_string().c_str(), stack_.size(), state->get_num(), state->count_todo());
193 /* Create the new expanded state (copy the state of MCed into our MCer data) */
194 auto next_state = std::make_shared<State>(get_remote_app(), state);
195 on_state_creation_signal(next_state.get(), get_remote_app());
197 if (reduction_mode_ == ReductionMode::odpor) {
198 // With ODPOR, after taking a step forward, we must assign a copy of that subtree to the next state.
200 // NOTE: We only add actions to the sleep set AFTER we've regenerated states. We must perform the search
201 // fully down a single path before we consider adding any elements to the sleep set according to the pseudocode
202 next_state->sprout_tree_from_parent_state();
204 /* Sleep set procedure:
205 * adding the taken transition to the sleep set of the original state.
206 * <!> Since the parent sleep set is used to compute the child sleep set, this need to be
207 * done after next_state creation */
208 XBT_DEBUG("Marking Transition >>%s<< of process %ld done and adding it to the sleep set",
209 state->get_transition_out()->to_string().c_str(), state->get_transition_out()->aid_);
210 state->add_sleep_set(
211 state->get_transition_out()); // Actors are marked done when they are considered in ActorState
214 /* DPOR persistent set procedure:
215 * for each new transition considered, check if it depends on any other previous transition executed before it
216 * on another process. If there exists one, find the more recent, and add its process to the interleave set.
217 * If the process is not enabled at this point, then add every enabled process to the interleave */
218 if (reduction_mode_ == ReductionMode::dpor) {
219 aid_t issuer_id = state->get_transition_out()->aid_;
220 stack_t tmp_stack = stack_;
221 while (not tmp_stack.empty()) {
222 if (const State* prev_state = tmp_stack.back().get();
223 state->get_transition_out()->aid_ == prev_state->get_transition_out()->aid_) {
224 XBT_DEBUG("Simcall >>%s<< and >>%s<< with same issuer %ld", state->get_transition_out()->to_string().c_str(),
225 prev_state->get_transition_out()->to_string().c_str(), issuer_id);
226 tmp_stack.pop_back();
228 } else if (prev_state->get_transition_out()->depends(state->get_transition_out().get())) {
229 XBT_VERB("Dependent Transitions:");
230 XBT_VERB(" #%ld %s (state=%ld)", prev_state->get_transition_out()->aid_,
231 prev_state->get_transition_out()->to_string().c_str(), prev_state->get_num());
232 XBT_VERB(" #%ld %s (state=%ld)", state->get_transition_out()->aid_,
233 state->get_transition_out()->to_string().c_str(), state->get_num());
235 if (prev_state->is_actor_enabled(issuer_id)) {
236 if (not prev_state->is_actor_done(issuer_id)) {
237 prev_state->consider_one(issuer_id);
238 opened_states_.emplace_back(tmp_stack.back());
240 XBT_DEBUG("Actor %ld is already in done set: no need to explore it again", issuer_id);
242 XBT_DEBUG("Actor %ld is not enabled: DPOR may be failing. To stay sound, we are marking every enabled "
243 "transition as todo",
245 // If we ended up marking at least a transition, explore it at some point
246 if (prev_state->consider_all() > 0)
247 opened_states_.emplace_back(tmp_stack.back());
251 XBT_VERB("INDEPENDENT Transitions:");
252 XBT_VERB(" #%ld %s (state=%ld)", prev_state->get_transition_out()->aid_,
253 prev_state->get_transition_out()->to_string().c_str(), prev_state->get_num());
254 XBT_VERB(" #%ld %s (state=%ld)", state->get_transition_out()->aid_,
255 state->get_transition_out()->to_string().c_str(), state->get_num());
257 tmp_stack.pop_back();
259 } else if (reduction_mode_ == ReductionMode::sdpor) {
261 * SDPOR Source Set Procedure:
263 * Find "reversible races" in the current execution `E` with respect to the latest action `p`. For each such race,
264 * determine one thread not contained in the backtrack set at the "race point" `r` which "represents" the trace
265 * formed by first executing everything after `r` that doesn't depend on it (`v := notdep(r, E)`) and then `p` to
268 * The intuition is that some subsequence of `v` may enable `p`, so we want to be sure that search "in that
271 execution_seq_.push_transition(std::move(executed_transition));
272 xbt_assert(execution_seq_.get_latest_event_handle().has_value(), "No events are contained in the SDPOR execution "
273 "even though one was just added");
275 const auto next_E_p = execution_seq_.get_latest_event_handle().value();
276 for (const auto e_race : execution_seq_.get_reversible_races_of(next_E_p)) {
277 State* prev_state = stack_[e_race].get();
278 const auto choices = execution_seq_.get_missing_source_set_actors_from(e_race, prev_state->get_backtrack_set());
279 if (not choices.empty()) {
280 // NOTE: To incorporate the idea of attempting to select the "best" backtrack point into SDPOR, instead of
281 // selecting the `first` initial, we should instead compute all choices and decide which is best
283 // Here, we choose the actor with the lowest ID to ensure we get deterministic results
285 std::min_element(choices.begin(), choices.end(), [](const aid_t a1, const aid_t a2) { return a1 < a2; });
286 prev_state->consider_one(*q);
287 opened_states_.emplace_back(std::move(prev_state));
290 } else if (reduction_mode_ == ReductionMode::odpor) {
291 // In the case of ODPOR, we simply observe the transition that was executed until we've reached a maximal trace
292 execution_seq_.push_transition(std::move(executed_transition));
295 // Before leaving that state, if the transition we just took can be taken multiple times, we
296 // need to give it to the opened states
297 if (stack_.back()->count_todo_multiples() > 0)
298 opened_states_.emplace_back(stack_.back());
300 stack_.emplace_back(std::move(next_state));
302 /* If this is a new state (or if we don't care about state-equality reduction) */
303 /* Get an enabled process and insert it in the interleave set of the next state */
304 if (reduction_mode_ == ReductionMode::dpor)
305 stack_.back()->consider_best(); // Take only one transition if DPOR: others may be considered later if required
307 stack_.back()->consider_all();
310 dot_output("\"%ld\" -> \"%ld\" [%s];\n", state->get_num(), stack_.back()->get_num(),
311 state->get_transition_out()->dot_string().c_str());
316 std::shared_ptr<State> DFSExplorer::best_opened_state()
318 int best_prio = 0; // cache the value for the best priority found so far (initialized to silence gcc)
319 auto best = end(opened_states_); // iterator to the state to explore having the best priority
320 auto valid = begin(opened_states_); // iterator marking the limit between states still to explore, and already
323 // Keep only still non-explored states (aid != -1), and record the one with the best (greater) priority.
324 for (auto current = begin(opened_states_); current != end(opened_states_); ++current) {
325 auto [aid, prio] = (*current)->next_transition_guided();
328 if (valid != current)
329 *valid = std::move(*current);
330 if (best == end(opened_states_) || prio < best_prio) {
337 std::shared_ptr<State> best_state;
339 // There are non-explored states, and one of them has the best priority. Remove it from opened_states_ before
341 best_state = std::move(*best);
344 *best = std::move(*valid);
346 opened_states_.erase(valid, end(opened_states_));
351 std::shared_ptr<State> DFSExplorer::next_odpor_state()
353 for (auto iter = stack_.rbegin(); iter != stack_.rend(); ++iter) {
354 const auto& state = *iter;
355 state->do_odpor_unwind();
356 XBT_DEBUG("\tPerformed ODPOR 'clean-up'. Sleep set has:");
357 for (const auto& [aid, transition] : state->get_sleep_set())
358 XBT_DEBUG("\t <%ld,%s>", aid, transition->to_string().c_str());
363 void DFSExplorer::backtrack()
365 if (const auto last_event = execution_seq_.get_latest_event_handle();
366 reduction_mode_ == ReductionMode::odpor and last_event.has_value()) {
368 * ODPOR Race Detection Procedure:
370 * For each reversible race in the current execution, we note if there are any continuations `C` equivalent to that
371 * which would reverse the race that have already either a) been searched by ODPOR or b) been *noted* to be searched
372 * by the wakeup tree at the appropriate reversal point, either as `C` directly or an as equivalent to `C`
373 * ("eventually looks like C", viz. the `~_E` relation)
375 for (auto e_prime = static_cast<odpor::Execution::EventHandle>(0); e_prime <= last_event.value(); ++e_prime) {
376 XBT_DEBUG("ODPOR: Now considering all possible race with `%u`", e_prime);
377 for (const auto e : execution_seq_.get_reversible_races_of(e_prime)) {
378 XBT_DEBUG("ODPOR: Reversible race detected between events `%u` and `%u`", e, e_prime);
379 State& prev_state = *stack_[e];
380 if (const auto v = execution_seq_.get_odpor_extension_from(e, e_prime, prev_state); v.has_value()) {
381 switch (prev_state.insert_into_wakeup_tree(v.value(), execution_seq_.get_prefix_before(e))) {
382 case odpor::WakeupTree::InsertionResult::root: {
383 XBT_DEBUG("ODPOR: Reversible race with `%u`(%ld: %.20s) unaccounted for in the wakeup tree for "
384 "the execution prior to event `%u`(%ld: %.20s):",
385 e_prime, stack_[e_prime]->get_transition_out()->aid_,
386 stack_[e_prime]->get_transition_out()->to_string(true).c_str(), e,
387 prev_state.get_transition_out()->aid_,
388 prev_state.get_transition_out()->to_string(true).c_str());
391 case odpor::WakeupTree::InsertionResult::interior_node: {
392 XBT_DEBUG("ODPOR: Reversible race with `%u` partially accounted for in the wakeup tree for "
393 "the execution prior to event `%u`:",
397 case odpor::WakeupTree::InsertionResult::leaf: {
398 XBT_DEBUG("ODPOR: Reversible race with `%u` accounted for in the wakeup tree for "
399 "the execution prior to event `%u`:",
404 for (const auto& seq : simgrid::mc::odpor::get_textual_trace(v.value())) {
405 XBT_DEBUG(" %s", seq.c_str());
408 XBT_DEBUG("ODPOR: Ignoring race: `sleep(E')` intersects `WI_[E'](v := notdep(%u, E))`", e);
409 XBT_DEBUG("Sleep set contains:");
410 for (const auto& [aid, transition] : prev_state.get_sleep_set())
411 XBT_DEBUG(" <%ld,%s>", aid, transition->to_string().c_str());
417 XBT_VERB("Backtracking from %s", get_record_trace().to_string().c_str());
418 XBT_DEBUG("%lu alternatives are yet to be explored:", opened_states_.size());
420 on_backtracking_signal(get_remote_app());
421 get_remote_app().check_deadlock();
423 // Take the point with smallest distance
424 auto backtracking_point = reduction_mode_ == ReductionMode::odpor ? next_odpor_state() : best_opened_state();
426 // if no backtracking point, then set the stack_ to empty so we can end the exploration
427 if (not backtracking_point) {
428 XBT_DEBUG("No more opened point of exploration, the search will end");
433 // We found a backtracking point, let's go to it
435 XBT_DEBUG("Backtracking to state#%ld", backtracking_point->get_num());
437 // Search how to restore the backtracking point
438 std::deque<Transition*> replay_recipe;
439 for (auto* s = backtracking_point.get(); s != nullptr; s = s->get_parent_state().get()) {
440 if (s->get_transition_in() != nullptr) // The root has no transition_in
441 replay_recipe.push_front(s->get_transition_in().get());
444 // Restore the initial state if no intermediate state was found
445 get_remote_app().restore_initial_state();
446 on_restore_initial_state_signal(get_remote_app());
448 /* if no snapshot, we need to restore the initial state and replay the transitions */
449 /* Traverse the stack from the state at position start and re-execute the transitions */
450 for (auto& transition : replay_recipe) {
451 transition->replay(get_remote_app());
452 on_transition_replay_signal(transition, get_remote_app());
453 visited_states_count_++;
455 this->restore_stack(backtracking_point);
458 DFSExplorer::DFSExplorer(const std::vector<char*>& args, ReductionMode mode) : Exploration(args), reduction_mode_(mode)
460 XBT_INFO("Start a DFS exploration. Reduction is: %s.", to_c_str(reduction_mode_));
462 auto initial_state = std::make_shared<State>(get_remote_app());
464 XBT_DEBUG("**************************************************");
466 stack_.emplace_back(std::move(initial_state));
468 /* Get an enabled actor and insert it in the interleave set of the initial state */
469 XBT_DEBUG("Initial state. %lu actors to consider", stack_.back()->get_actor_count());
470 if (reduction_mode_ == ReductionMode::dpor)
471 stack_.back()->consider_best();
473 stack_.back()->consider_all();
475 if (stack_.back()->count_todo_multiples() > 1)
476 opened_states_.emplace_back(stack_.back());
479 Exploration* create_dfs_exploration(const std::vector<char*>& args, ReductionMode mode)
481 return new DFSExplorer(args, mode);
484 } // namespace simgrid::mc