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/remote/mc_protocol.h"
12 #include "src/mc/transition/Transition.hpp"
14 #include "xbt/asserts.h"
16 #include "xbt/string.hpp"
17 #include "xbt/sysdep.h"
25 #include <unordered_set>
28 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(mc_dfs, mc, "DFS exploration algorithm of the model-checker");
30 namespace simgrid::mc {
32 xbt::signal<void(RemoteApp&)> DFSExplorer::on_exploration_start_signal;
33 xbt::signal<void(RemoteApp&)> DFSExplorer::on_backtracking_signal;
35 xbt::signal<void(State*, RemoteApp&)> DFSExplorer::on_state_creation_signal;
37 xbt::signal<void(State*, RemoteApp&)> DFSExplorer::on_restore_system_state_signal;
38 xbt::signal<void(RemoteApp&)> DFSExplorer::on_restore_initial_state_signal;
39 xbt::signal<void(Transition*, RemoteApp&)> DFSExplorer::on_transition_replay_signal;
40 xbt::signal<void(Transition*, RemoteApp&)> DFSExplorer::on_transition_execute_signal;
42 xbt::signal<void(RemoteApp&)> DFSExplorer::on_log_state_signal;
44 RecordTrace DFSExplorer::get_record_trace() // override
48 if (const auto trans = stack_.back()->get_transition_out(); trans != nullptr)
49 res.push_back(trans.get());
50 for (const auto* state = stack_.back().get(); state != nullptr; state = state->get_parent_state().get())
51 if (state->get_transition_in() != nullptr)
52 res.push_front(state->get_transition_in().get());
57 void DFSExplorer::restore_stack(std::shared_ptr<State> state)
60 execution_seq_ = odpor::Execution();
61 auto current_state = state;
62 stack_.emplace_front(current_state);
63 // condition corresponds to reaching initial state
64 while (current_state->get_parent_state() != nullptr) {
65 current_state = current_state->get_parent_state();
66 stack_.emplace_front(current_state);
68 XBT_DEBUG("Replaced stack by %s", get_record_trace().to_string().c_str());
69 if (reduction_mode_ == ReductionMode::sdpor || reduction_mode_ == ReductionMode::odpor) {
70 // NOTE: The outgoing transition for the top-most state of the stack refers to that which was taken
71 // as part of the last trace explored by the algorithm. Thus, only the sequence of transitions leading up to,
72 // but not including, the last state must be included when reconstructing the Exploration for SDPOR.
73 for (auto iter = std::next(stack_.begin()); iter != stack_.end(); ++iter) {
74 execution_seq_.push_transition((*iter)->get_transition_in());
76 XBT_DEBUG("Replaced SDPOR/ODPOR execution to reflect the new stack");
80 void DFSExplorer::log_state() // override
82 on_log_state_signal(get_remote_app());
83 XBT_INFO("DFS exploration ended. %ld unique states visited; %lu backtracks (%lu transition replays, %lu states "
85 State::get_expanded_states(), backtrack_count_, Transition::get_replayed_transitions(),
86 visited_states_count_);
87 Exploration::log_state();
90 void DFSExplorer::run()
92 on_exploration_start_signal(get_remote_app());
93 /* This function runs the DFS algorithm the state space.
94 * We do so iteratively instead of recursively, dealing with the call stack manually.
95 * This allows one to explore the call stack at will. */
97 while (not stack_.empty()) {
98 /* Get current state */
99 auto state = stack_.back();
101 XBT_DEBUG("**************************************************");
102 XBT_DEBUG("Exploration depth=%zu (state:#%ld; %zu interleaves todo)", stack_.size(), state->get_num(),
103 state->count_todo());
105 visited_states_count_++;
107 // Backtrack if we reached the maximum depth
108 if (stack_.size() > (std::size_t)_sg_mc_max_depth) {
109 if (reduction_mode_ == ReductionMode::dpor) {
110 XBT_ERROR("/!\\ Max depth of %d reached! THIS WILL PROBABLY BREAK the dpor reduction /!\\",
111 _sg_mc_max_depth.get());
112 XBT_ERROR("/!\\ If bad things happen, disable dpor with --cfg=model-check/reduction:none /!\\");
113 } else if (reduction_mode_ == ReductionMode::sdpor || reduction_mode_ == ReductionMode::odpor) {
114 XBT_ERROR("/!\\ Max depth of %d reached! THIS **WILL** BREAK the reduction, which is not sound "
115 "when stopping at a fixed depth /!\\",
116 _sg_mc_max_depth.get());
117 XBT_ERROR("/!\\ If bad things happen, disable the reduction with --cfg=model-check/reduction:none /!\\");
119 XBT_WARN("/!\\ Max depth reached ! /!\\ ");
125 if (reduction_mode_ == ReductionMode::odpor) {
126 // In the case of ODPOR, the wakeup tree for this state may be empty if we're exploring new territory
127 // (rather than following the partial execution of a wakeup tree). This corresponds to lines 9 to 13 of
128 // the ODPOR pseudocode
130 // INVARIANT: The execution sequence should be consistent with the state when seeding the tree. If the sequence
131 // gets out of sync with the state, selection will not work as we intend
132 state->seed_wakeup_tree_if_needed(execution_seq_);
135 // Search for the next transition
136 // next_transition returns a pair<aid_t, int>
137 // in case we want to consider multiple states (eg. during backtrack)
138 const aid_t next = reduction_mode_ == ReductionMode::odpor ? state->next_odpor_transition()
139 : std::get<0>(state->next_transition_guided());
141 if (next < 0 || not state->is_actor_enabled(next)) {
142 if (next >= 0) { // Actor is not enabled, then
144 "Reduction %s wants to execute a disabled transition %s. If it's ODPOR, ReversibleRace is suboptimal.",
145 to_c_str(reduction_mode_), state->get_actors_list().at(next).get_transition()->to_string(true).c_str());
146 if (reduction_mode_ == ReductionMode::odpor) {
147 // Remove the disabled transition from the wakeup tree so that ODPOR doesn't try it again
148 state->remove_subtree_at_aid(next);
149 state->add_sleep_set(state->get_actors_list().at(next).get_transition());
151 xbt_assert(false, "Only ODPOR should be confident enought in itself to try executing a disabled transition");
154 // If there is no more transition in the current state (or if ODPOR picked an actor that is not enabled --
155 // ReversibleRace is an overapproximation), backtrace
156 XBT_VERB("%lu actors remain, but none of them need to be interleaved (depth %zu).", state->get_actor_count(),
159 if (state->get_actor_count() == 0) {
160 get_remote_app().finalize_app();
161 XBT_VERB("Execution came to an end at %s (state: %ld, depth: %zu)", get_record_trace().to_string().c_str(),
162 state->get_num(), stack_.size());
169 if (XBT_LOG_ISENABLED(mc_dfs, xbt_log_priority_verbose)) {
170 XBT_VERB("Sleep set actually containing:");
171 for (const auto& [aid, transition] : state->get_sleep_set())
172 XBT_VERB(" <%ld,%s>", aid, transition->to_string().c_str());
175 auto todo = state->get_actors_list().at(next).get_transition();
176 XBT_DEBUG("wanna execute %ld: %.60s", next, todo->to_string().c_str());
178 /* Actually answer the request: let's execute the selected request (MCed does one step) */
179 auto executed_transition = state->execute_next(next, get_remote_app());
180 on_transition_execute_signal(state->get_transition_out().get(), get_remote_app());
182 // If there are processes to interleave and the maximum depth has not been
183 // reached then perform one step of the exploration algorithm.
184 XBT_VERB("Executed %ld: %.60s (stack depth: %zu, state: %ld, %zu interleaves)", state->get_transition_out()->aid_,
185 state->get_transition_out()->to_string().c_str(), stack_.size(), state->get_num(), state->count_todo());
187 /* Create the new expanded state (copy the state of MCed into our MCer data) */
188 auto next_state = std::make_shared<State>(get_remote_app(), state);
189 on_state_creation_signal(next_state.get(), get_remote_app());
191 if (reduction_mode_ == ReductionMode::odpor) {
192 // With ODPOR, after taking a step forward, we must assign a copy of that subtree to the next state.
194 // NOTE: We only add actions to the sleep set AFTER we've regenerated states. We must perform the search
195 // fully down a single path before we consider adding any elements to the sleep set according to the pseudocode
196 next_state->sprout_tree_from_parent_state();
198 /* Sleep set procedure:
199 * adding the taken transition to the sleep set of the original state.
200 * <!> Since the parent sleep set is used to compute the child sleep set, this need to be
201 * done after next_state creation */
202 XBT_DEBUG("Marking Transition >>%s<< of process %ld done and adding it to the sleep set",
203 state->get_transition_out()->to_string().c_str(), state->get_transition_out()->aid_);
204 state->add_sleep_set(
205 state->get_transition_out()); // Actors are marked done when they are considered in ActorState
208 /* DPOR persistent set procedure:
209 * for each new transition considered, check if it depends on any other previous transition executed before it
210 * on another process. If there exists one, find the more recent, and add its process to the interleave set.
211 * If the process is not enabled at this point, then add every enabled process to the interleave */
212 if (reduction_mode_ == ReductionMode::dpor) {
213 aid_t issuer_id = state->get_transition_out()->aid_;
214 stack_t tmp_stack = stack_;
215 while (not tmp_stack.empty()) {
216 if (const State* prev_state = tmp_stack.back().get();
217 state->get_transition_out()->aid_ == prev_state->get_transition_out()->aid_) {
218 XBT_DEBUG("Simcall >>%s<< and >>%s<< with same issuer %ld", state->get_transition_out()->to_string().c_str(),
219 prev_state->get_transition_out()->to_string().c_str(), issuer_id);
220 tmp_stack.pop_back();
222 } else if (prev_state->get_transition_out()->depends(state->get_transition_out().get())) {
223 XBT_VERB("Dependent Transitions:");
224 XBT_VERB(" #%ld %s (state=%ld)", prev_state->get_transition_out()->aid_,
225 prev_state->get_transition_out()->to_string().c_str(), prev_state->get_num());
226 XBT_VERB(" #%ld %s (state=%ld)", state->get_transition_out()->aid_,
227 state->get_transition_out()->to_string().c_str(), state->get_num());
229 if (prev_state->is_actor_enabled(issuer_id)) {
230 if (not prev_state->is_actor_done(issuer_id)) {
231 prev_state->consider_one(issuer_id);
232 opened_states_.emplace_back(tmp_stack.back());
234 XBT_DEBUG("Actor %ld is already in done set: no need to explore it again", issuer_id);
236 XBT_DEBUG("Actor %ld is not enabled: DPOR may be failing. To stay sound, we are marking every enabled "
237 "transition as todo",
239 // If we ended up marking at least a transition, explore it at some point
240 if (prev_state->consider_all() > 0)
241 opened_states_.emplace_back(tmp_stack.back());
245 XBT_VERB("INDEPENDENT Transitions:");
246 XBT_VERB(" #%ld %s (state=%ld)", prev_state->get_transition_out()->aid_,
247 prev_state->get_transition_out()->to_string().c_str(), prev_state->get_num());
248 XBT_VERB(" #%ld %s (state=%ld)", state->get_transition_out()->aid_,
249 state->get_transition_out()->to_string().c_str(), state->get_num());
251 tmp_stack.pop_back();
253 } else if (reduction_mode_ == ReductionMode::sdpor) {
255 * SDPOR Source Set Procedure:
257 * Find "reversible races" in the current execution `E` with respect to the latest action `p`. For each such race,
258 * determine one thread not contained in the backtrack set at the "race point" `r` which "represents" the trace
259 * formed by first executing everything after `r` that doesn't depend on it (`v := notdep(r, E)`) and then `p` to
262 * The intuition is that some subsequence of `v` may enable `p`, so we want to be sure that search "in that
265 execution_seq_.push_transition(std::move(executed_transition));
266 xbt_assert(execution_seq_.get_latest_event_handle().has_value(), "No events are contained in the SDPOR execution "
267 "even though one was just added");
269 const auto next_E_p = execution_seq_.get_latest_event_handle().value();
270 for (const auto e_race : execution_seq_.get_reversible_races_of(next_E_p)) {
271 State* prev_state = stack_[e_race].get();
272 const auto choices = execution_seq_.get_missing_source_set_actors_from(e_race, prev_state->get_backtrack_set());
273 if (not choices.empty()) {
274 // NOTE: To incorporate the idea of attempting to select the "best" backtrack point into SDPOR, instead of
275 // selecting the `first` initial, we should instead compute all choices and decide which is best
277 // Here, we choose the actor with the lowest ID to ensure we get deterministic results
279 std::min_element(choices.begin(), choices.end(), [](const aid_t a1, const aid_t a2) { return a1 < a2; });
280 prev_state->consider_one(*q);
281 opened_states_.emplace_back(std::move(prev_state));
284 } else if (reduction_mode_ == ReductionMode::odpor) {
285 // In the case of ODPOR, we simply observe the transition that was executed until we've reached a maximal trace
286 execution_seq_.push_transition(std::move(executed_transition));
289 // Before leaving that state, if the transition we just took can be taken multiple times, we
290 // need to give it to the opened states
291 if (stack_.back()->count_todo_multiples() > 0)
292 opened_states_.emplace_back(stack_.back());
294 stack_.emplace_back(std::move(next_state));
296 /* If this is a new state (or if we don't care about state-equality reduction) */
297 /* Get an enabled process and insert it in the interleave set of the next state */
298 if (reduction_mode_ == ReductionMode::dpor)
299 stack_.back()->consider_best(); // Take only one transition if DPOR: others may be considered later if required
301 stack_.back()->consider_all();
304 dot_output("\"%ld\" -> \"%ld\" [%s];\n", state->get_num(), stack_.back()->get_num(),
305 state->get_transition_out()->dot_string().c_str());
310 std::shared_ptr<State> DFSExplorer::best_opened_state()
312 int best_prio = 0; // cache the value for the best priority found so far (initialized to silence gcc)
313 auto best = end(opened_states_); // iterator to the state to explore having the best priority
314 auto valid = begin(opened_states_); // iterator marking the limit between states still to explore, and already
317 // Keep only still non-explored states (aid != -1), and record the one with the best (greater) priority.
318 for (auto current = begin(opened_states_); current != end(opened_states_); ++current) {
319 auto [aid, prio] = (*current)->next_transition_guided();
322 if (valid != current)
323 *valid = std::move(*current);
324 if (best == end(opened_states_) || prio < best_prio) {
331 std::shared_ptr<State> best_state;
333 // There are non-explored states, and one of them has the best priority. Remove it from opened_states_ before
335 best_state = std::move(*best);
338 *best = std::move(*valid);
340 opened_states_.erase(valid, end(opened_states_));
345 std::shared_ptr<State> DFSExplorer::next_odpor_state()
347 for (auto iter = stack_.rbegin(); iter != stack_.rend(); ++iter) {
348 const auto& state = *iter;
349 state->do_odpor_unwind();
350 XBT_DEBUG("\tPerformed ODPOR 'clean-up'. Sleep set has:");
351 for (const auto& [aid, transition] : state->get_sleep_set())
352 XBT_DEBUG("\t <%ld,%s>", aid, transition->to_string().c_str());
353 if (not state->has_empty_tree()) {
360 void DFSExplorer::backtrack()
362 if (const auto last_event = execution_seq_.get_latest_event_handle();
363 reduction_mode_ == ReductionMode::odpor and last_event.has_value()) {
365 * ODPOR Race Detection Procedure:
367 * For each reversible race in the current execution, we note if there are any continuations `C` equivalent to that
368 * which would reverse the race that have already either a) been searched by ODPOR or b) been *noted* to be searched
369 * by the wakeup tree at the appropriate reversal point, either as `C` directly or an as equivalent to `C`
370 * ("eventually looks like C", viz. the `~_E` relation)
372 for (auto e_prime = static_cast<odpor::Execution::EventHandle>(0); e_prime <= last_event.value(); ++e_prime) {
373 XBT_DEBUG("ODPOR: Now considering all possible race with `%u`", e_prime);
374 for (const auto e : execution_seq_.get_reversible_races_of(e_prime)) {
375 XBT_DEBUG("ODPOR: Reversible race detected between events `%u` and `%u`", e, e_prime);
376 State& prev_state = *stack_[e];
377 if (const auto v = execution_seq_.get_odpor_extension_from(e, e_prime, prev_state); v.has_value()) {
378 switch (prev_state.insert_into_wakeup_tree(v.value(), execution_seq_.get_prefix_before(e))) {
379 case odpor::WakeupTree::InsertionResult::root: {
380 XBT_DEBUG("ODPOR: Reversible race with `%u`(%ld: %.20s) unaccounted for in the wakeup tree for "
381 "the execution prior to event `%u`(%ld: %.20s):",
382 e_prime, stack_[e_prime]->get_transition_out()->aid_,
383 stack_[e_prime]->get_transition_out()->to_string(true).c_str(), e,
384 prev_state.get_transition_out()->aid_,
385 prev_state.get_transition_out()->to_string(true).c_str());
388 case odpor::WakeupTree::InsertionResult::interior_node: {
389 XBT_DEBUG("ODPOR: Reversible race with `%u` partially accounted for in the wakeup tree for "
390 "the execution prior to event `%u`:",
394 case odpor::WakeupTree::InsertionResult::leaf: {
395 XBT_DEBUG("ODPOR: Reversible race with `%u` accounted for in the wakeup tree for "
396 "the execution prior to event `%u`:",
401 for (const auto& seq : simgrid::mc::odpor::get_textual_trace(v.value())) {
402 XBT_DEBUG(" %s", seq.c_str());
405 XBT_DEBUG("ODPOR: Ignoring race: `sleep(E')` intersects `WI_[E'](v := notdep(%u, E))`", e);
406 XBT_DEBUG("Sleep set contains:");
407 for (const auto& [aid, transition] : prev_state.get_sleep_set())
408 XBT_DEBUG(" <%ld,%s>", aid, transition->to_string().c_str());
414 XBT_VERB("Backtracking from %s", get_record_trace().to_string().c_str());
415 XBT_DEBUG("%lu alternatives are yet to be explored:", opened_states_.size());
417 on_backtracking_signal(get_remote_app());
418 get_remote_app().check_deadlock();
420 // Take the point with smallest distance
421 auto backtracking_point = reduction_mode_ == ReductionMode::odpor ? next_odpor_state() : best_opened_state();
423 // if no backtracking point, then set the stack_ to empty so we can end the exploration
424 if (not backtracking_point) {
425 XBT_DEBUG("No more opened point of exploration, the search will end");
430 // We found a backtracking point, let's go to it
432 XBT_DEBUG("Backtracking to state#%ld", backtracking_point->get_num());
434 // Search how to restore the backtracking point
435 std::deque<Transition*> replay_recipe;
436 for (auto* s = backtracking_point.get(); s != nullptr; s = s->get_parent_state().get()) {
437 if (s->get_transition_in() != nullptr) // The root has no transition_in
438 replay_recipe.push_front(s->get_transition_in().get());
441 // Restore the initial state if no intermediate state was found
442 get_remote_app().restore_initial_state();
443 on_restore_initial_state_signal(get_remote_app());
445 /* if no snapshot, we need to restore the initial state and replay the transitions */
446 /* Traverse the stack from the state at position start and re-execute the transitions */
447 for (auto& transition : replay_recipe) {
448 transition->replay(get_remote_app());
449 on_transition_replay_signal(transition, get_remote_app());
450 visited_states_count_++;
452 this->restore_stack(backtracking_point);
455 DFSExplorer::DFSExplorer(const std::vector<char*>& args, ReductionMode mode) : Exploration(args), reduction_mode_(mode)
457 XBT_INFO("Start a DFS exploration. Reduction is: %s.", to_c_str(reduction_mode_));
459 auto initial_state = std::make_shared<State>(get_remote_app());
461 XBT_DEBUG("**************************************************");
463 stack_.emplace_back(std::move(initial_state));
465 /* Get an enabled actor and insert it in the interleave set of the initial state */
466 XBT_DEBUG("Initial state. %lu actors to consider", stack_.back()->get_actor_count());
467 if (reduction_mode_ == ReductionMode::dpor)
468 stack_.back()->consider_best();
470 stack_.back()->consider_all();
472 if (stack_.back()->count_todo_multiples() > 1)
473 opened_states_.emplace_back(stack_.back());
476 Exploration* create_dfs_exploration(const std::vector<char*>& args, ReductionMode mode)
478 return new DFSExplorer(args, mode);
481 } // namespace simgrid::mc