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());
139 if (next < 0) { // If there is no more transition in the current state, backtrack.
140 XBT_VERB("%lu actors remain, but none of them need to be interleaved (depth %zu).", state->get_actor_count(),
143 if (state->get_actor_count() == 0) {
144 get_remote_app().finalize_app();
145 XBT_VERB("Execution came to an end at %s (state: %ld, depth: %zu)", get_record_trace().to_string().c_str(),
146 state->get_num(), stack_.size());
153 if (XBT_LOG_ISENABLED(mc_dfs, xbt_log_priority_verbose)) {
154 XBT_VERB("Sleep set actually containing:");
155 for (const auto& [aid, transition] : state->get_sleep_set())
156 XBT_VERB(" <%ld,%s>", aid, transition->to_string().c_str());
159 auto todo = state->get_actors_list().at(next).get_transition();
160 XBT_DEBUG("wanna execute %ld: %.60s", next, todo->to_string().c_str());
162 /* Actually answer the request: let's execute the selected request (MCed does one step) */
163 auto executed_transition = state->execute_next(next, get_remote_app());
164 on_transition_execute_signal(state->get_transition_out().get(), get_remote_app());
166 // If there are processes to interleave and the maximum depth has not been
167 // reached then perform one step of the exploration algorithm.
168 XBT_VERB("Executed %ld: %.60s (stack depth: %zu, state: %ld, %zu interleaves)", state->get_transition_out()->aid_,
169 state->get_transition_out()->to_string().c_str(), stack_.size(), state->get_num(), state->count_todo());
171 /* Create the new expanded state (copy the state of MCed into our MCer data) */
172 auto next_state = std::make_shared<State>(get_remote_app(), state);
173 on_state_creation_signal(next_state.get(), get_remote_app());
175 if (reduction_mode_ == ReductionMode::odpor) {
176 // With ODPOR, after taking a step forward, we must assign a copy of that subtree to the next state.
178 // NOTE: We only add actions to the sleep set AFTER we've regenerated states. We must perform the search
179 // fully down a single path before we consider adding any elements to the sleep set according to the pseudocode
180 next_state->sprout_tree_from_parent_state();
182 /* Sleep set procedure:
183 * adding the taken transition to the sleep set of the original state.
184 * <!> Since the parent sleep set is used to compute the child sleep set, this need to be
185 * done after next_state creation */
186 XBT_DEBUG("Marking Transition >>%s<< of process %ld done and adding it to the sleep set",
187 state->get_transition_out()->to_string().c_str(), state->get_transition_out()->aid_);
188 state->add_sleep_set(
189 state->get_transition_out()); // Actors are marked done when they are considered in ActorState
192 /* DPOR persistent set procedure:
193 * for each new transition considered, check if it depends on any other previous transition executed before it
194 * on another process. If there exists one, find the more recent, and add its process to the interleave set.
195 * If the process is not enabled at this point, then add every enabled process to the interleave */
196 if (reduction_mode_ == ReductionMode::dpor) {
197 aid_t issuer_id = state->get_transition_out()->aid_;
198 stack_t tmp_stack = stack_;
199 while (not tmp_stack.empty()) {
200 if (const State* prev_state = tmp_stack.back().get();
201 state->get_transition_out()->aid_ == prev_state->get_transition_out()->aid_) {
202 XBT_DEBUG("Simcall >>%s<< and >>%s<< with same issuer %ld", state->get_transition_out()->to_string().c_str(),
203 prev_state->get_transition_out()->to_string().c_str(), issuer_id);
204 tmp_stack.pop_back();
206 } else if (prev_state->get_transition_out()->depends(state->get_transition_out().get())) {
207 XBT_VERB("Dependent Transitions:");
208 XBT_VERB(" #%ld %s (state=%ld)", prev_state->get_transition_out()->aid_,
209 prev_state->get_transition_out()->to_string().c_str(), prev_state->get_num());
210 XBT_VERB(" #%ld %s (state=%ld)", state->get_transition_out()->aid_,
211 state->get_transition_out()->to_string().c_str(), state->get_num());
213 if (prev_state->is_actor_enabled(issuer_id)) {
214 if (not prev_state->is_actor_done(issuer_id)) {
215 prev_state->consider_one(issuer_id);
216 opened_states_.emplace_back(tmp_stack.back());
218 XBT_DEBUG("Actor %ld is already in done set: no need to explore it again", issuer_id);
220 XBT_DEBUG("Actor %ld is not enabled: DPOR may be failing. To stay sound, we are marking every enabled "
221 "transition as todo",
223 // If we ended up marking at least a transition, explore it at some point
224 if (prev_state->consider_all() > 0)
225 opened_states_.emplace_back(tmp_stack.back());
229 XBT_VERB("INDEPENDENT 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 tmp_stack.pop_back();
237 } else if (reduction_mode_ == ReductionMode::sdpor) {
239 * SDPOR Source Set Procedure:
241 * Find "reversible races" in the current execution `E` with respect to the latest action `p`. For each such race,
242 * determine one thread not contained in the backtrack set at the "race point" `r` which "represents" the trace
243 * formed by first executing everything after `r` that doesn't depend on it (`v := notdep(r, E)`) and then `p` to
246 * The intuition is that some subsequence of `v` may enable `p`, so we want to be sure that search "in that
249 execution_seq_.push_transition(std::move(executed_transition));
250 xbt_assert(execution_seq_.get_latest_event_handle().has_value(), "No events are contained in the SDPOR execution "
251 "even though one was just added");
253 const auto next_E_p = execution_seq_.get_latest_event_handle().value();
254 for (const auto e_race : execution_seq_.get_reversible_races_of(next_E_p)) {
255 State* prev_state = stack_[e_race].get();
256 const auto choices = execution_seq_.get_missing_source_set_actors_from(e_race, prev_state->get_backtrack_set());
257 if (not choices.empty()) {
258 // NOTE: To incorporate the idea of attempting to select the "best" backtrack point into SDPOR, instead of
259 // selecting the `first` initial, we should instead compute all choices and decide which is best
261 // Here, we choose the actor with the lowest ID to ensure we get deterministic results
263 std::min_element(choices.begin(), choices.end(), [](const aid_t a1, const aid_t a2) { return a1 < a2; });
264 prev_state->consider_one(*q);
265 opened_states_.emplace_back(std::move(prev_state));
268 } else if (reduction_mode_ == ReductionMode::odpor) {
269 // In the case of ODPOR, we simply observe the transition that was executed until we've reached a maximal trace
270 execution_seq_.push_transition(std::move(executed_transition));
273 // Before leaving that state, if the transition we just took can be taken multiple times, we
274 // need to give it to the opened states
275 if (stack_.back()->count_todo_multiples() > 0)
276 opened_states_.emplace_back(stack_.back());
278 stack_.emplace_back(std::move(next_state));
280 /* If this is a new state (or if we don't care about state-equality reduction) */
281 /* Get an enabled process and insert it in the interleave set of the next state */
282 if (reduction_mode_ == ReductionMode::dpor)
283 stack_.back()->consider_best(); // Take only one transition if DPOR: others may be considered later if required
285 stack_.back()->consider_all();
288 dot_output("\"%ld\" -> \"%ld\" [%s];\n", state->get_num(), stack_.back()->get_num(),
289 state->get_transition_out()->dot_string().c_str());
294 std::shared_ptr<State> DFSExplorer::best_opened_state()
296 int best_prio = 0; // cache the value for the best priority found so far (initialized to silence gcc)
297 auto best = end(opened_states_); // iterator to the state to explore having the best priority
298 auto valid = begin(opened_states_); // iterator marking the limit between states still to explore, and already
301 // Keep only still non-explored states (aid != -1), and record the one with the best (greater) priority.
302 for (auto current = begin(opened_states_); current != end(opened_states_); ++current) {
303 auto [aid, prio] = (*current)->next_transition_guided();
306 if (valid != current)
307 *valid = std::move(*current);
308 if (best == end(opened_states_) || prio < best_prio) {
315 std::shared_ptr<State> best_state;
317 // There are non-explored states, and one of them has the best priority. Remove it from opened_states_ before
319 best_state = std::move(*best);
322 *best = std::move(*valid);
324 opened_states_.erase(valid, end(opened_states_));
329 std::shared_ptr<State> DFSExplorer::next_odpor_state()
331 for (auto iter = stack_.rbegin(); iter != stack_.rend(); ++iter) {
332 const auto& state = *iter;
333 state->do_odpor_unwind();
334 XBT_DEBUG("\tPerformed ODPOR 'clean-up'. Sleep set has:");
335 for (const auto& [aid, transition] : state->get_sleep_set())
336 XBT_DEBUG("\t <%ld,%s>", aid, transition->to_string().c_str());
337 if (not state->has_empty_tree()) {
344 void DFSExplorer::backtrack()
346 if (const auto last_event = execution_seq_.get_latest_event_handle();
347 reduction_mode_ == ReductionMode::odpor and last_event.has_value()) {
349 * ODPOR Race Detection Procedure:
351 * For each reversible race in the current execution, we note if there are any continuations `C` equivalent to that
352 * which would reverse the race that have already either a) been searched by ODPOR or b) been *noted* to be searched
353 * by the wakeup tree at the appropriate reversal point, either as `C` directly or an as equivalent to `C`
354 * ("eventually looks like C", viz. the `~_E` relation)
356 for (auto e_prime = static_cast<odpor::Execution::EventHandle>(0); e_prime <= last_event.value(); ++e_prime) {
357 for (const auto e : execution_seq_.get_reversible_races_of(e_prime)) {
358 XBT_DEBUG("ODPOR: Reversible race detected between events `%u` and `%u`", e, e_prime);
359 State& prev_state = *stack_[e];
360 if (const auto v = execution_seq_.get_odpor_extension_from(e, e_prime, prev_state); v.has_value()) {
361 switch (prev_state.insert_into_wakeup_tree(v.value(), execution_seq_.get_prefix_before(e))) {
362 case odpor::WakeupTree::InsertionResult::root: {
363 XBT_DEBUG("ODPOR: Reversible race with `%u` unaccounted for in the wakeup tree for "
364 "the execution prior to event `%u`:",
368 case odpor::WakeupTree::InsertionResult::interior_node: {
369 XBT_DEBUG("ODPOR: Reversible race with `%u` partially accounted for in the wakeup tree for "
370 "the execution prior to event `%u`:",
374 case odpor::WakeupTree::InsertionResult::leaf: {
375 XBT_DEBUG("ODPOR: Reversible race with `%u` accounted for in the wakeup tree for "
376 "the execution prior to event `%u`:",
381 for (const auto& seq : simgrid::mc::odpor::get_textual_trace(v.value())) {
382 XBT_DEBUG(" %s", seq.c_str());
385 XBT_DEBUG("ODPOR: Ignoring race: `sleep(E')` intersects `WI_[E'](v := notdep(%u, E))`", e);
386 XBT_DEBUG("Sleep set contains:");
387 for (const auto& [aid, transition] : prev_state.get_sleep_set())
388 XBT_DEBUG(" <%ld,%s>", aid, transition->to_string().c_str());
394 XBT_VERB("Backtracking from %s", get_record_trace().to_string().c_str());
395 XBT_DEBUG("%lu alternatives are yet to be explored:", opened_states_.size());
397 on_backtracking_signal(get_remote_app());
398 get_remote_app().check_deadlock();
400 // Take the point with smallest distance
401 auto backtracking_point = reduction_mode_ == ReductionMode::odpor ? next_odpor_state() : best_opened_state();
403 // if no backtracking point, then set the stack_ to empty so we can end the exploration
404 if (not backtracking_point) {
405 XBT_DEBUG("No more opened point of exploration, the search will end");
410 // We found a backtracking point, let's go to it
412 XBT_DEBUG("Backtracking to state#%ld", backtracking_point->get_num());
414 // Search how to restore the backtracking point
415 std::deque<Transition*> replay_recipe;
416 for (auto* s = backtracking_point.get(); s != nullptr; s = s->get_parent_state().get()) {
417 if (s->get_transition_in() != nullptr) // The root has no transition_in
418 replay_recipe.push_front(s->get_transition_in().get());
421 // Restore the initial state if no intermediate state was found
422 get_remote_app().restore_initial_state();
423 on_restore_initial_state_signal(get_remote_app());
425 /* if no snapshot, we need to restore the initial state and replay the transitions */
426 /* Traverse the stack from the state at position start and re-execute the transitions */
427 for (auto& transition : replay_recipe) {
428 transition->replay(get_remote_app());
429 on_transition_replay_signal(transition, get_remote_app());
430 visited_states_count_++;
432 this->restore_stack(backtracking_point);
435 DFSExplorer::DFSExplorer(const std::vector<char*>& args, ReductionMode mode) : Exploration(args), reduction_mode_(mode)
437 XBT_INFO("Start a DFS exploration. Reduction is: %s.", to_c_str(reduction_mode_));
439 auto initial_state = std::make_shared<State>(get_remote_app());
441 XBT_DEBUG("**************************************************");
443 stack_.emplace_back(std::move(initial_state));
445 /* Get an enabled actor and insert it in the interleave set of the initial state */
446 XBT_DEBUG("Initial state. %lu actors to consider", stack_.back()->get_actor_count());
447 if (reduction_mode_ == ReductionMode::dpor)
448 stack_.back()->consider_best();
450 stack_.back()->consider_all();
452 if (stack_.back()->count_todo_multiples() > 1)
453 opened_states_.emplace_back(stack_.back());
456 Exploration* create_dfs_exploration(const std::vector<char*>& args, ReductionMode mode)
458 return new DFSExplorer(args, mode);
461 } // namespace simgrid::mc