*/
struct model_snapshot_members {
model_snapshot_members() :
- current_action(NULL),
/* First thread created will have id INITIAL_THREAD_ID */
next_thread_id(INITIAL_THREAD_ID),
used_sequence_numbers(0),
bugs.clear();
}
- ModelAction *current_action;
unsigned int next_thread_id;
modelclock_t used_sequence_numbers;
ModelAction *next_backtrack;
/* Print all model-checker output before rollback */
fflush(model_out);
+ /**
+ * FIXME: if we utilize partial rollback, we will need to free only
+ * those pending actions which were NOT pending before the rollback
+ * point
+ */
+ for (unsigned int i = 0; i < get_num_threads(); i++)
+ delete get_thread(int_to_id(i))->get_pending();
+
snapshot_backtrack_before(0);
}
* adjacency of read/write portions of a RMW action, force THREAD_CREATE to be
* followed by a THREAD_START, or it can enforce execution replay/backtracking.
* The model-checker may have no preference regarding the next thread (i.e.,
- * when exploring a new execution ordering), in which case this will return
- * NULL.
- * @param curr The current ModelAction. This action might guide the choice of
- * next thread.
- * @return The next thread to run. If the model-checker has no preference, NULL.
+ * when exploring a new execution ordering), in which case we defer to the
+ * scheduler.
+ *
+ * @param curr Optional: The current ModelAction. Only used if non-NULL and it
+ * might guide the choice of next thread (i.e., THREAD_CREATE should be
+ * followed by THREAD_START, or ATOMIC_RMWR followed by ATOMIC_{RMW,RMWC})
+ * @return The next chosen thread to run, if any exist. Or else if no threads
+ * remain to be executed, return NULL.
*/
Thread * ModelChecker::get_next_thread(ModelAction *curr)
{
if (curr != NULL) {
/* Do not split atomic actions. */
if (curr->is_rmwr())
- return thread_current();
+ return get_thread(curr);
else if (curr->get_type() == THREAD_CREATE)
return curr->get_thread_operand();
}
- /* Have we completed exploring the preselected path? */
+ /*
+ * Have we completed exploring the preselected path? Then let the
+ * scheduler decide
+ */
if (diverge == NULL)
- return NULL;
+ return scheduler->select_next_thread();
/* Else, we are trying to replay an execution */
ModelAction *next = node_stack->get_next()->get_action();
earliest_diverge = prevnode->get_action();
}
}
+ /* Start the round robin scheduler from this thread id */
+ scheduler->set_scheduler_thread(tid);
/* The correct sleep set is in the parent node. */
execute_sleep_set();
for (unsigned int i = 0; i < get_num_threads(); i++) {
thread_id_t tid = int_to_id(i);
Thread *thr = get_thread(tid);
- if (scheduler->is_sleep_set(thr) && thr->get_pending() == NULL) {
- thr->set_state(THREAD_RUNNING);
- scheduler->next_thread(thr);
- Thread::swap(&system_context, thr);
- priv->current_action->set_sleep_flag();
- thr->set_pending(priv->current_action);
+ if (scheduler->is_sleep_set(thr) && thr->get_pending()) {
+ thr->get_pending()->set_sleep_flag();
}
}
}
+/**
+ * @brief Should the current action wake up a given thread?
+ *
+ * @param curr The current action
+ * @param thread The thread that we might wake up
+ * @return True, if we should wake up the sleeping thread; false otherwise
+ */
+bool ModelChecker::should_wake_up(const ModelAction *curr, const Thread *thread) const
+{
+ const ModelAction *asleep = thread->get_pending();
+ /* Don't allow partial RMW to wake anyone up */
+ if (curr->is_rmwr())
+ return false;
+ /* Synchronizing actions may have been backtracked */
+ if (asleep->could_synchronize_with(curr))
+ return true;
+ /* All acquire/release fences and fence-acquire/store-release */
+ if (asleep->is_fence() && asleep->is_acquire() && curr->is_release())
+ return true;
+ /* Fence-release + store can awake load-acquire on the same location */
+ if (asleep->is_read() && asleep->is_acquire() && curr->same_var(asleep) && curr->is_write()) {
+ ModelAction *fence_release = get_last_fence_release(curr->get_tid());
+ if (fence_release && *(get_last_action(thread->get_id())) < *fence_release)
+ return true;
+ }
+ return false;
+}
+
void ModelChecker::wake_up_sleeping_actions(ModelAction *curr)
{
for (unsigned int i = 0; i < get_num_threads(); i++) {
Thread *thr = get_thread(int_to_id(i));
if (scheduler->is_sleep_set(thr)) {
- ModelAction *pending_act = thr->get_pending();
- if ((!curr->is_rmwr()) && pending_act->could_synchronize_with(curr))
- //Remove this thread from sleep set
+ if (should_wake_up(curr, thr))
+ /* Remove this thread from sleep set */
scheduler->remove_sleep(thr);
}
}
priv->bad_synchronization = true;
}
+/**
+ * Check whether the current trace has triggered an assertion which should halt
+ * its execution.
+ *
+ * @return True, if the execution should be aborted; false otherwise
+ */
bool ModelChecker::has_asserted() const
{
return priv->asserted;
}
+/**
+ * Trigger a trace assertion which should cause this execution to be halted.
+ * This can be due to a detected bug or due to an infeasibility that should
+ * halt ASAP.
+ */
void ModelChecker::set_assert()
{
priv->asserted = true;
return true;
}
-ModelAction * ModelChecker::get_last_conflict(ModelAction *act)
+/**
+ * @brief Find the last fence-related backtracking conflict for a ModelAction
+ *
+ * This function performs the search for the most recent conflicting action
+ * against which we should perform backtracking, as affected by fence
+ * operations. This includes pairs of potentially-synchronizing actions which
+ * occur due to fence-acquire or fence-release, and hence should be explored in
+ * the opposite execution order.
+ *
+ * @param act The current action
+ * @return The most recent action which conflicts with act due to fences
+ */
+ModelAction * ModelChecker::get_last_fence_conflict(ModelAction *act) const
+{
+ /* Only perform release/acquire fence backtracking for stores */
+ if (!act->is_write())
+ return NULL;
+
+ /* Find a fence-release (or, act is a release) */
+ ModelAction *last_release;
+ if (act->is_release())
+ last_release = act;
+ else
+ last_release = get_last_fence_release(act->get_tid());
+ if (!last_release)
+ return NULL;
+
+ /* Skip past the release */
+ action_list_t *list = action_trace;
+ action_list_t::reverse_iterator rit;
+ for (rit = list->rbegin(); rit != list->rend(); rit++)
+ if (*rit == last_release)
+ break;
+ ASSERT(rit != list->rend());
+
+ /* Find a prior:
+ * load-acquire
+ * or
+ * load --sb-> fence-acquire */
+ std::vector< ModelAction *, ModelAlloc<ModelAction *> > acquire_fences(get_num_threads(), NULL);
+ std::vector< ModelAction *, ModelAlloc<ModelAction *> > prior_loads(get_num_threads(), NULL);
+ bool found_acquire_fences = false;
+ for ( ; rit != list->rend(); rit++) {
+ ModelAction *prev = *rit;
+ if (act->same_thread(prev))
+ continue;
+
+ int tid = id_to_int(prev->get_tid());
+
+ if (prev->is_read() && act->same_var(prev)) {
+ if (prev->is_acquire()) {
+ /* Found most recent load-acquire, don't need
+ * to search for more fences */
+ if (!found_acquire_fences)
+ return NULL;
+ } else {
+ prior_loads[tid] = prev;
+ }
+ }
+ if (prev->is_acquire() && prev->is_fence() && !acquire_fences[tid]) {
+ found_acquire_fences = true;
+ acquire_fences[tid] = prev;
+ }
+ }
+
+ ModelAction *latest_backtrack = NULL;
+ for (unsigned int i = 0; i < acquire_fences.size(); i++)
+ if (acquire_fences[i] && prior_loads[i])
+ if (!latest_backtrack || *latest_backtrack < *acquire_fences[i])
+ latest_backtrack = acquire_fences[i];
+ return latest_backtrack;
+}
+
+/**
+ * @brief Find the last backtracking conflict for a ModelAction
+ *
+ * This function performs the search for the most recent conflicting action
+ * against which we should perform backtracking. This primary includes pairs of
+ * synchronizing actions which should be explored in the opposite execution
+ * order.
+ *
+ * @param act The current action
+ * @return The most recent action which conflicts with act
+ */
+ModelAction * ModelChecker::get_last_conflict(ModelAction *act) const
{
switch (act->get_type()) {
- case ATOMIC_FENCE:
+ /* case ATOMIC_FENCE: fences don't directly cause backtracking */
case ATOMIC_READ:
case ATOMIC_WRITE:
case ATOMIC_RMW: {
- /* Optimization: relaxed operations don't need backtracking */
- if (act->is_relaxed())
- return NULL;
+ ModelAction *ret = NULL;
+
/* linear search: from most recent to oldest */
action_list_t *list = get_safe_ptr_action(obj_map, act->get_location());
action_list_t::reverse_iterator rit;
for (rit = list->rbegin(); rit != list->rend(); rit++) {
ModelAction *prev = *rit;
- if (prev->could_synchronize_with(act))
- return prev;
+ if (prev->could_synchronize_with(act)) {
+ ret = prev;
+ break;
+ }
}
- break;
+
+ ModelAction *ret2 = get_last_fence_conflict(act);
+ if (!ret2)
+ return ret;
+ if (!ret)
+ return ret2;
+ if (*ret < *ret2)
+ return ret2;
+ return ret;
}
case ATOMIC_LOCK:
case ATOMIC_TRYLOCK: {
switch (curr->get_type()) {
case THREAD_CREATE: {
- Thread *th = curr->get_thread_operand();
+ thrd_t *thrd = (thrd_t *)curr->get_location();
+ struct thread_params *params = (struct thread_params *)curr->get_value();
+ Thread *th = new Thread(thrd, params->func, params->arg, get_thread(curr));
+ add_thread(th);
th->set_creation(curr);
/* Promises can be satisfied by children */
for (unsigned int i = 0; i < promises->size(); i++) {
return true;
}
-/**
- * Stores the ModelAction for the current thread action. Call this
- * immediately before switching from user- to system-context to pass
- * data between them.
- * @param act The ModelAction created by the user-thread action
- */
-void ModelChecker::set_current_action(ModelAction *act) {
- priv->current_action = act;
-}
-
/**
* This is the heart of the model checker routine. It performs model-checking
* actions corresponding to a given "current action." Among other processes, it
/* Build may_read_from set for newly-created actions */
if (newly_explored && curr->is_read())
- build_reads_from_past(curr);
+ build_may_read_from(curr);
/* Initialize work_queue with the "current action" work */
work_queue_t work_queue(1, CheckCurrWorkEntry(curr));
if (mo_graph->checkForCycles())
return false;
- while (rf) {
+ for ( ; rf != NULL; rf = rf->get_reads_from()) {
ASSERT(rf->is_write());
if (rf->is_release())
/* acq_rel RMW is a sufficient stopping condition */
if (rf->is_acquire() && rf->is_release())
return true; /* complete */
-
- rf = rf->get_reads_from();
};
if (!rf) {
/* read from future: need to settle this later */
for (unsigned int i = 0; i < promises->size(); i++) {
Promise *promise = (*promises)[i];
const ModelAction *act = promise->get_action();
+ ASSERT(act->is_read());
if (!act->happens_before(curr) &&
- act->is_read() &&
!act->could_synchronize_with(curr) &&
- !act->same_thread(curr) &&
- act->get_location() == curr->get_location() &&
+ promise->is_compatible(curr) &&
promise->get_value() == curr->get_value()) {
curr->get_node()->set_promise(i, act->is_rmw());
}
/**
* Build up an initial set of all past writes that this 'read' action may read
- * from. This set is determined by the clock vector's "happens before"
- * relationship.
+ * from, as well as any previously-observed future values that must still be valid.
+ *
* @param curr is the current ModelAction that we are exploring; it must be a
* 'read' operation.
*/
-void ModelChecker::build_reads_from_past(ModelAction *curr)
+void ModelChecker::build_may_read_from(ModelAction *curr)
{
std::vector<action_list_t> *thrd_lists = get_safe_ptr_vect_action(obj_thrd_map, curr->get_location());
unsigned int i;
break;
}
}
+
+ /* Inherit existing, promised future values */
+ for (i = 0; i < promises->size(); i++) {
+ const Promise *promise = (*promises)[i];
+ const ModelAction *promise_read = promise->get_action();
+ if (promise_read->same_var(curr)) {
+ /* Only add feasible future-values */
+ mo_graph->startChanges();
+ r_modification_order(curr, promise);
+ if (!is_infeasible()) {
+ const struct future_value fv = promise->get_fv();
+ curr->get_node()->add_future_value(fv);
+ }
+ mo_graph->rollbackChanges();
+ }
+ }
+
/* We may find no valid may-read-from only if the execution is doomed */
- if (!curr->get_node()->get_read_from_size()) {
+ if (!curr->get_node()->get_read_from_size() && curr->get_node()->future_value_empty()) {
priv->no_valid_reads = true;
set_assert();
}
bool ModelChecker::sleep_can_read_from(ModelAction *curr, const ModelAction *write)
{
- while (true) {
+ for ( ; write != NULL; write = write->get_reads_from()) {
/* UNINIT actions don't have a Node, and they never sleep */
if (write->is_uninitialized())
return true;
bool thread_sleep = prevnode->enabled_status(curr->get_tid()) == THREAD_SLEEP_SET;
if (write->is_release() && thread_sleep)
return true;
- if (!write->is_rmw()) {
+ if (!write->is_rmw())
return false;
- }
- if (write->get_reads_from() == NULL)
- return true;
- write = write->get_reads_from();
}
+ return true;
}
/**
return scheduler->is_enabled(tid);
}
+/**
+ * Switch from a model-checker context to a user-thread context. This is the
+ * complement of ModelChecker::switch_to_master and must be called from the
+ * model-checker context
+ *
+ * @param thread The user-thread to switch to
+ */
+void ModelChecker::switch_from_master(Thread *thread)
+{
+ scheduler->set_current_thread(thread);
+ Thread::swap(&system_context, thread);
+}
+
/**
* Switch from a user-context to the "master thread" context (a.k.a. system
* context). This switch is made with the intention of exploring a particular
{
DBG();
Thread *old = thread_current();
- set_current_action(act);
- old->set_state(THREAD_READY);
+ ASSERT(!old->get_pending());
+ old->set_pending(act);
if (Thread::swap(old, &system_context) < 0) {
perror("swap threads");
exit(EXIT_FAILURE);
/**
* Takes the next step in the execution, if possible.
* @param curr The current step to take
- * @return Returns true (success) if a step was taken and false otherwise.
+ * @return Returns the next Thread to run, if any; NULL if this execution
+ * should terminate
*/
-bool ModelChecker::take_step(ModelAction *curr)
+Thread * ModelChecker::take_step(ModelAction *curr)
{
- if (has_asserted())
- return false;
-
Thread *curr_thrd = get_thread(curr);
ASSERT(curr_thrd->get_state() == THREAD_READY);
/* Infeasible -> don't take any more steps */
if (is_infeasible())
- return false;
+ return NULL;
else if (isfeasibleprefix() && have_bug_reports()) {
set_assert();
- return false;
+ return NULL;
}
- if (params.bound != 0)
- if (priv->used_sequence_numbers > params.bound)
- return false;
+ if (params.bound != 0 && priv->used_sequence_numbers > params.bound)
+ return NULL;
if (curr_thrd->is_blocked() || curr_thrd->is_complete())
scheduler->remove_thread(curr_thrd);
Thread *next_thrd = get_next_thread(curr);
- next_thrd = scheduler->next_thread(next_thrd);
DEBUG("(%d, %d)\n", curr_thrd ? id_to_int(curr_thrd->get_id()) : -1,
next_thrd ? id_to_int(next_thrd->get_id()) : -1);
- /*
- * Launch end-of-execution release sequence fixups only when there are:
- *
- * (1) no more user threads to run (or when execution replay chooses
- * the 'model_thread')
- * (2) pending release sequences
- * (3) pending assertions (i.e., data races)
- * (4) no pending promises
- */
- if (!pending_rel_seqs->empty() && (!next_thrd || next_thrd->is_model_thread()) &&
- is_feasible_prefix_ignore_relseq() && !unrealizedraces.empty()) {
- model_print("*** WARNING: release sequence fixup action (%zu pending release seuqences) ***\n",
- pending_rel_seqs->size());
- ModelAction *fixup = new ModelAction(MODEL_FIXUP_RELSEQ,
- std::memory_order_seq_cst, NULL, VALUE_NONE,
- model_thread);
- set_current_action(fixup);
- return true;
- }
-
- /* next_thrd == NULL -> don't take any more steps */
- if (!next_thrd)
- return false;
-
- next_thrd->set_state(THREAD_RUNNING);
-
- if (next_thrd->get_pending() != NULL) {
- /* restart a pending action */
- set_current_action(next_thrd->get_pending());
- next_thrd->set_pending(NULL);
- next_thrd->set_state(THREAD_READY);
- return true;
- }
-
- /* Return false only if swap fails with an error */
- return (Thread::swap(&system_context, next_thrd) == 0);
+ return next_thrd;
}
/** Wrapper to run the user's main function, with appropriate arguments */
{
do {
thrd_t user_thread;
- Thread *t = new Thread(&user_thread, &user_main_wrapper, NULL);
-
+ Thread *t = new Thread(&user_thread, &user_main_wrapper, NULL, NULL);
add_thread(t);
- /* Run user thread up to its first action */
- scheduler->next_thread(t);
- Thread::swap(&system_context, t);
+ do {
+ /*
+ * Stash next pending action(s) for thread(s). There
+ * should only need to stash one thread's action--the
+ * thread which just took a step--plus the first step
+ * for any newly-created thread
+ */
+ for (unsigned int i = 0; i < get_num_threads(); i++) {
+ thread_id_t tid = int_to_id(i);
+ Thread *thr = get_thread(tid);
+ if (!thr->is_model_thread() && !thr->is_complete() && !thr->get_pending()) {
+ switch_from_master(thr);
+ }
+ }
+
+ /* Catch assertions from prior take_step or from
+ * between-ModelAction bugs (e.g., data races) */
+ if (has_asserted())
+ break;
+
+ /* Consume the next action for a Thread */
+ ModelAction *curr = t->get_pending();
+ t->set_pending(NULL);
+ t = take_step(curr);
+ } while (t && !t->is_model_thread());
- /* Wait for all threads to complete */
- while (take_step(priv->current_action));
+ /*
+ * Launch end-of-execution release sequence fixups only when
+ * the execution is otherwise feasible AND there are:
+ *
+ * (1) pending release sequences
+ * (2) pending assertions that could be invalidated by a change
+ * in clock vectors (i.e., data races)
+ * (3) no pending promises
+ */
+ while (!pending_rel_seqs->empty() &&
+ is_feasible_prefix_ignore_relseq() &&
+ !unrealizedraces.empty()) {
+ model_print("*** WARNING: release sequence fixup action "
+ "(%zu pending release seuqence(s)) ***\n",
+ pending_rel_seqs->size());
+ ModelAction *fixup = new ModelAction(MODEL_FIXUP_RELSEQ,
+ std::memory_order_seq_cst, NULL, VALUE_NONE,
+ model_thread);
+ take_step(fixup);
+ };
} while (next_execution());
+ model_print("******* Model-checking complete: *******\n");
print_stats();
}