/* lock */
-void __bch2_btree_node_lock_write(struct btree_trans *trans,
- struct btree_bkey_cached_common *b)
+/*
+ * @trans wants to lock @b with type @type
+ */
+struct trans_waiting_for_lock {
+ struct btree_trans *trans;
+ struct btree_bkey_cached_common *node_want;
+ enum six_lock_type lock_want;
+
+ /* for iterating over held locks :*/
+ u8 path_idx;
+ u8 level;
+ u64 lock_start_time;
+};
+
+struct lock_graph {
+ struct trans_waiting_for_lock g[8];
+ unsigned nr;
+};
+
+static noinline void print_cycle(struct printbuf *out, struct lock_graph *g)
{
- int readers = bch2_btree_node_lock_counts(trans, NULL, b, b->level).n[SIX_LOCK_read];
+ struct trans_waiting_for_lock *i;
- /*
- * Must drop our read locks before calling six_lock_write() -
- * six_unlock() won't do wakeups until the reader count
- * goes to 0, and it's safe because we have the node intent
- * locked:
- */
- six_lock_readers_add(&b->lock, -readers);
- btree_node_lock_nopath_nofail(trans, b, SIX_LOCK_write);
- six_lock_readers_add(&b->lock, readers);
+ prt_printf(out, "Found lock cycle (%u entries):", g->nr);
+ prt_newline(out);
+
+ for (i = g->g; i < g->g + g->nr; i++)
+ bch2_btree_trans_to_text(out, i->trans);
}
-static inline bool path_has_read_locks(struct btree_path *path)
+static noinline void print_chain(struct printbuf *out, struct lock_graph *g)
{
- unsigned l;
+ struct trans_waiting_for_lock *i;
- for (l = 0; l < BTREE_MAX_DEPTH; l++)
- if (btree_node_read_locked(path, l))
+ for (i = g->g; i != g->g + g->nr; i++) {
+ if (i != g->g)
+ prt_str(out, "<- ");
+ prt_printf(out, "%u ", i->trans->locking_wait.task->pid);
+ }
+ prt_newline(out);
+}
+
+static void lock_graph_up(struct lock_graph *g)
+{
+ closure_put(&g->g[--g->nr].trans->ref);
+}
+
+static noinline void lock_graph_pop_all(struct lock_graph *g)
+{
+ while (g->nr)
+ lock_graph_up(g);
+}
+
+static void lock_graph_down(struct lock_graph *g, struct btree_trans *trans)
+{
+ closure_get(&trans->ref);
+
+ g->g[g->nr++] = (struct trans_waiting_for_lock) {
+ .trans = trans,
+ .node_want = trans->locking,
+ .lock_want = trans->locking_wait.lock_want,
+ };
+}
+
+static bool lock_graph_remove_non_waiters(struct lock_graph *g)
+{
+ struct trans_waiting_for_lock *i;
+
+ for (i = g->g + 1; i < g->g + g->nr; i++)
+ if (i->trans->locking != i->node_want ||
+ i->trans->locking_wait.start_time != i[-1].lock_start_time) {
+ while (g->g + g->nr > i)
+ lock_graph_up(g);
return true;
+ }
+
return false;
}
-/* Slowpath: */
-int __bch2_btree_node_lock(struct btree_trans *trans,
- struct btree_path *path,
- struct btree_bkey_cached_common *b,
- struct bpos pos, unsigned level,
- enum six_lock_type type,
- six_lock_should_sleep_fn should_sleep_fn, void *p,
- unsigned long ip)
+static int abort_lock(struct lock_graph *g, struct trans_waiting_for_lock *i)
{
- struct btree_path *linked;
- unsigned reason;
+ if (i == g->g) {
+ trace_and_count(i->trans->c, trans_restart_would_deadlock, i->trans, _RET_IP_);
+ return btree_trans_restart(i->trans, BCH_ERR_transaction_restart_would_deadlock);
+ } else {
+ i->trans->lock_must_abort = true;
+ wake_up_process(i->trans->locking_wait.task);
+ return 0;
+ }
+}
+
+static int btree_trans_abort_preference(struct btree_trans *trans)
+{
+ if (trans->lock_may_not_fail)
+ return 0;
+ if (trans->locking_wait.lock_want == SIX_LOCK_write)
+ return 1;
+ if (!trans->in_traverse_all)
+ return 2;
+ return 3;
+}
+
+static noinline int break_cycle(struct lock_graph *g, struct printbuf *cycle)
+{
+ struct trans_waiting_for_lock *i, *abort = NULL;
+ unsigned best = 0, pref;
+ int ret;
+
+ if (lock_graph_remove_non_waiters(g))
+ return 0;
+
+ /* Only checking, for debugfs: */
+ if (cycle) {
+ print_cycle(cycle, g);
+ ret = -1;
+ goto out;
+ }
+
+ for (i = g->g; i < g->g + g->nr; i++) {
+ pref = btree_trans_abort_preference(i->trans);
+ if (pref > best) {
+ abort = i;
+ best = pref;
+ }
+ }
+
+ if (unlikely(!best)) {
+ struct printbuf buf = PRINTBUF;
+
+ prt_printf(&buf, bch2_fmt(g->g->trans->c, "cycle of nofail locks"));
+
+ for (i = g->g; i < g->g + g->nr; i++) {
+ struct btree_trans *trans = i->trans;
+
+ bch2_btree_trans_to_text(&buf, trans);
+
+ prt_printf(&buf, "backtrace:");
+ prt_newline(&buf);
+ printbuf_indent_add(&buf, 2);
+ bch2_prt_task_backtrace(&buf, trans->locking_wait.task);
+ printbuf_indent_sub(&buf, 2);
+ prt_newline(&buf);
+ }
+
+ bch2_print_string_as_lines(KERN_ERR, buf.buf);
+ printbuf_exit(&buf);
+ BUG();
+ }
+
+ ret = abort_lock(g, abort);
+out:
+ if (ret)
+ while (g->nr)
+ lock_graph_up(g);
+ return ret;
+}
+
+static int lock_graph_descend(struct lock_graph *g, struct btree_trans *trans,
+ struct printbuf *cycle)
+{
+ struct btree_trans *orig_trans = g->g->trans;
+ struct trans_waiting_for_lock *i;
+
+ for (i = g->g; i < g->g + g->nr; i++)
+ if (i->trans == trans)
+ return break_cycle(g, cycle);
+
+ if (g->nr == ARRAY_SIZE(g->g)) {
+ if (orig_trans->lock_may_not_fail)
+ return 0;
+
+ while (g->nr)
+ lock_graph_up(g);
+
+ if (cycle)
+ return 0;
+
+ trace_and_count(trans->c, trans_restart_would_deadlock_recursion_limit, trans, _RET_IP_);
+ return btree_trans_restart(orig_trans, BCH_ERR_transaction_restart_deadlock_recursion_limit);
+ }
+
+ lock_graph_down(g, trans);
+ return 0;
+}
+
+static bool lock_type_conflicts(enum six_lock_type t1, enum six_lock_type t2)
+{
+ return t1 + t2 > 1;
+}
+
+int bch2_check_for_deadlock(struct btree_trans *trans, struct printbuf *cycle)
+{
+ struct lock_graph g;
+ struct trans_waiting_for_lock *top;
+ struct btree_bkey_cached_common *b;
+ struct btree_path *path;
+ int ret;
+
+ if (trans->lock_must_abort) {
+ if (cycle)
+ return -1;
+
+ trace_and_count(trans->c, trans_restart_would_deadlock, trans, _RET_IP_);
+ return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock);
+ }
+
+ g.nr = 0;
+ lock_graph_down(&g, trans);
+next:
+ if (!g.nr)
+ return 0;
- /* Check if it's safe to block: */
- trans_for_each_path(trans, linked) {
- if (!linked->nodes_locked)
+ top = &g.g[g.nr - 1];
+
+ trans_for_each_path_from(top->trans, path, top->path_idx) {
+ if (!path->nodes_locked)
continue;
- /*
- * Can't block taking an intent lock if we have _any_ nodes read
- * locked:
- *
- * - Our read lock blocks another thread with an intent lock on
- * the same node from getting a write lock, and thus from
- * dropping its intent lock
- *
- * - And the other thread may have multiple nodes intent locked:
- * both the node we want to intent lock, and the node we
- * already have read locked - deadlock:
- */
- if (type == SIX_LOCK_intent &&
- path_has_read_locks(linked)) {
- reason = 1;
- goto deadlock;
+ if (top->path_idx != path->idx) {
+ top->path_idx = path->idx;
+ top->level = 0;
+ top->lock_start_time = 0;
}
- if (linked->btree_id != path->btree_id) {
- if (linked->btree_id < path->btree_id)
+ for (;
+ top->level < BTREE_MAX_DEPTH;
+ top->level++, top->lock_start_time = 0) {
+ int lock_held = btree_node_locked_type(path, top->level);
+
+ if (lock_held == BTREE_NODE_UNLOCKED)
continue;
- reason = 3;
- goto deadlock;
- }
+ b = &READ_ONCE(path->l[top->level].b)->c;
+
+ if (IS_ERR_OR_NULL(b)) {
+ /*
+ * If we get here, it means we raced with the
+ * other thread updating its btree_path
+ * structures - which means it can't be blocked
+ * waiting on a lock:
+ */
+ if (!lock_graph_remove_non_waiters(&g)) {
+ /*
+ * If lock_graph_remove_non_waiters()
+ * didn't do anything, it must be
+ * because we're being called by debugfs
+ * checking for lock cycles, which
+ * invokes us on btree_transactions that
+ * aren't actually waiting on anything.
+ * Just bail out:
+ */
+ lock_graph_pop_all(&g);
+ }
+
+ goto next;
+ }
- /*
- * Within the same btree, non-cached paths come before cached
- * paths:
- */
- if (linked->cached != path->cached) {
- if (!linked->cached)
+ if (list_empty_careful(&b->lock.wait_list))
continue;
- reason = 4;
- goto deadlock;
- }
+ raw_spin_lock(&b->lock.wait_lock);
+ list_for_each_entry(trans, &b->lock.wait_list, locking_wait.list) {
+ BUG_ON(b != trans->locking);
- /*
- * Interior nodes must be locked before their descendants: if
- * another path has possible descendants locked of the node
- * we're about to lock, it must have the ancestors locked too:
- */
- if (level > btree_path_highest_level_locked(linked)) {
- reason = 5;
- goto deadlock;
- }
+ if (top->lock_start_time &&
+ time_after_eq64(top->lock_start_time, trans->locking_wait.start_time))
+ continue;
+
+ top->lock_start_time = trans->locking_wait.start_time;
+
+ /* Don't check for self deadlock: */
+ if (trans == top->trans ||
+ !lock_type_conflicts(lock_held, trans->locking_wait.lock_want))
+ continue;
+
+ ret = lock_graph_descend(&g, trans, cycle);
+ raw_spin_unlock(&b->lock.wait_lock);
- /* Must lock btree nodes in key order: */
- if (btree_node_locked(linked, level) &&
- bpos_cmp(pos, btree_node_pos(&linked->l[level].b->c)) <= 0) {
- reason = 7;
- goto deadlock;
+ if (ret)
+ return ret;
+ goto next;
+
+ }
+ raw_spin_unlock(&b->lock.wait_lock);
}
}
- return btree_node_lock_type(trans, path, b, pos, level,
- type, should_sleep_fn, p);
-deadlock:
- trace_and_count(trans->c, trans_restart_would_deadlock, trans, ip, reason, linked, path, &pos);
- return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock);
+ if (g.nr > 1 && cycle)
+ print_chain(cycle, &g);
+ lock_graph_up(&g);
+ goto next;
+}
+
+int bch2_six_check_for_deadlock(struct six_lock *lock, void *p)
+{
+ struct btree_trans *trans = p;
+
+ return bch2_check_for_deadlock(trans, NULL);
+}
+
+int __bch2_btree_node_lock_write(struct btree_trans *trans, struct btree_path *path,
+ struct btree_bkey_cached_common *b,
+ bool lock_may_not_fail)
+{
+ int readers = bch2_btree_node_lock_counts(trans, NULL, b, b->level).n[SIX_LOCK_read];
+ int ret;
+
+ /*
+ * Must drop our read locks before calling six_lock_write() -
+ * six_unlock() won't do wakeups until the reader count
+ * goes to 0, and it's safe because we have the node intent
+ * locked:
+ */
+ six_lock_readers_add(&b->lock, -readers);
+ ret = __btree_node_lock_nopath(trans, b, SIX_LOCK_write,
+ lock_may_not_fail, _RET_IP_);
+ six_lock_readers_add(&b->lock, readers);
+
+ if (ret)
+ mark_btree_node_locked_noreset(path, b->level, SIX_LOCK_intent);
+
+ return ret;
}
/* relock */
}
bool __bch2_btree_node_relock(struct btree_trans *trans,
- struct btree_path *path, unsigned level)
+ struct btree_path *path, unsigned level,
+ bool trace)
{
struct btree *b = btree_path_node(path, level);
int want = __btree_lock_want(path, level);
return true;
}
fail:
- trace_and_count(trans->c, btree_path_relock_fail, trans, _RET_IP_, path, level);
+ if (trace && !trans->notrace_relock_fail)
+ trace_and_count(trans->c, btree_path_relock_fail, trans, _RET_IP_, path, level);
return false;
}
struct btree_path *path, unsigned level)
{
struct btree *b = path->l[level].b;
+ struct six_lock_count count = bch2_btree_node_lock_counts(trans, path, &b->c, level);
if (!is_btree_node(path, level))
return false;
if (race_fault())
return false;
- if (btree_node_locked(path, level)
- ? six_lock_tryupgrade(&b->c.lock)
- : six_relock_type(&b->c.lock, SIX_LOCK_intent, path->l[level].lock_seq))
- goto success;
+ if (btree_node_locked(path, level)) {
+ bool ret;
+
+ six_lock_readers_add(&b->c.lock, -count.n[SIX_LOCK_read]);
+ ret = six_lock_tryupgrade(&b->c.lock);
+ six_lock_readers_add(&b->c.lock, count.n[SIX_LOCK_read]);
+
+ if (ret)
+ goto success;
+ } else {
+ if (six_relock_type(&b->c.lock, SIX_LOCK_intent, path->l[level].lock_seq))
+ goto success;
+ }
+ /*
+ * Do we already have an intent lock via another path? If so, just bump
+ * lock count:
+ */
if (btree_node_lock_seq_matches(path, b, level) &&
btree_node_lock_increment(trans, &b->c, level, BTREE_NODE_INTENT_LOCKED)) {
btree_node_unlock(trans, path, level);
return btree_path_get_locks(trans, path, false);
}
+int __bch2_btree_path_relock(struct btree_trans *trans,
+ struct btree_path *path, unsigned long trace_ip)
+{
+ if (!bch2_btree_path_relock_norestart(trans, path, trace_ip)) {
+ trace_and_count(trans->c, trans_restart_relock_path, trans, trace_ip, path);
+ return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_path);
+ }
+
+ return 0;
+}
+
__flatten
bool bch2_btree_path_upgrade_norestart(struct btree_trans *trans,
struct btree_path *path, unsigned long trace_ip)
struct btree_path *path;
if (unlikely(trans->restarted))
- return - ((int) trans->restarted);
+ return -((int) trans->restarted);
trans_for_each_path(trans, path)
if (path->should_be_locked &&
return 0;
}
+int bch2_trans_relock_notrace(struct btree_trans *trans)
+{
+ struct btree_path *path;
+
+ if (unlikely(trans->restarted))
+ return -((int) trans->restarted);
+
+ trans_for_each_path(trans, path)
+ if (path->should_be_locked &&
+ !bch2_btree_path_relock_norestart(trans, path, _RET_IP_)) {
+ return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
+ }
+ return 0;
+}
+
void bch2_trans_unlock(struct btree_trans *trans)
{
struct btree_path *path;
* bch2_gc_btree_init_recurse() doesn't use btree iterators for walking
* btree nodes, it implements its own walking:
*/
- BUG_ON(!trans->is_initial_gc &&
- lock_class_is_held(&bch2_btree_node_lock_key));
+ EBUG_ON(!trans->is_initial_gc &&
+ lock_class_is_held(&bch2_btree_node_lock_key));
+}
+
+bool bch2_trans_locked(struct btree_trans *trans)
+{
+ struct btree_path *path;
+
+ trans_for_each_path(trans, path)
+ if (path->nodes_locked)
+ return true;
+ return false;
}
/* Debug */