#include "btree_locking.h"
#include "btree_types.h"
-struct lock_class_key bch2_btree_node_lock_key;
+static struct lock_class_key bch2_btree_node_lock_key;
-/* Btree node locking: */
+void bch2_btree_lock_init(struct btree_bkey_cached_common *b,
+ enum six_lock_init_flags flags)
+{
+ __six_lock_init(&b->lock, "b->c.lock", &bch2_btree_node_lock_key, flags);
+ lockdep_set_novalidate_class(&b->lock);
+}
-static inline void six_lock_readers_add(struct six_lock *lock, int nr)
+#ifdef CONFIG_LOCKDEP
+void bch2_assert_btree_nodes_not_locked(void)
{
- if (lock->readers)
- this_cpu_add(*lock->readers, nr);
- else if (nr > 0)
- atomic64_add(__SIX_VAL(read_lock, nr), &lock->state.counter);
- else
- atomic64_sub(__SIX_VAL(read_lock, -nr), &lock->state.counter);
+#if 0
+ //Re-enable when lock_class_is_held() is merged:
+ BUG_ON(lock_class_is_held(&bch2_btree_node_lock_key));
+#endif
}
+#endif
+
+/* Btree node locking: */
struct six_lock_count bch2_btree_node_lock_counts(struct btree_trans *trans,
struct btree_path *skip,
{
struct btree_path *path;
struct six_lock_count ret;
+ unsigned i;
memset(&ret, 0, sizeof(ret));
if (IS_ERR_OR_NULL(b))
return ret;
- trans_for_each_path(trans, path)
+ trans_for_each_path(trans, path, i)
if (path != skip && &path->l[level].b->c == b) {
int t = btree_node_locked_type(path, level);
unsigned nr;
};
-static void lock_graph_pop(struct lock_graph *g)
-{
- closure_put(&g->g[--g->nr].trans->ref);
-}
-
static noinline void print_cycle(struct printbuf *out, struct lock_graph *g)
{
struct trans_waiting_for_lock *i;
bch2_btree_trans_to_text(out, i->trans);
}
-static int abort_lock(struct lock_graph *g, struct trans_waiting_for_lock *i)
+static noinline void print_chain(struct printbuf *out, struct lock_graph *g)
{
- int ret;
+ struct trans_waiting_for_lock *i;
+
+ for (i = g->g; i != g->g + g->nr; i++) {
+ struct task_struct *task = i->trans->locking_wait.task;
+ if (i != g->g)
+ prt_str(out, "<- ");
+ prt_printf(out, "%u ", task ?task->pid : 0);
+ }
+ 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)
+{
+ g->g[g->nr++] = (struct trans_waiting_for_lock) {
+ .trans = trans,
+ .node_want = trans->locking,
+ .lock_want = trans->locking_wait.lock_want,
+ };
+}
+
+static void lock_graph_down(struct lock_graph *g, struct btree_trans *trans)
+{
+ closure_get(&trans->ref);
+ __lock_graph_down(g, trans);
+}
+
+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;
+}
+
+static void trace_would_deadlock(struct lock_graph *g, struct btree_trans *trans,
+ unsigned long ip)
+{
+ struct bch_fs *c = trans->c;
+ count_event(c, trans_restart_would_deadlock);
+
+ if (trace_trans_restart_would_deadlock_enabled()) {
+ struct printbuf buf = PRINTBUF;
+
+ buf.atomic++;
+ print_cycle(&buf, g);
+
+ trace_trans_restart_would_deadlock(trans, ip, buf.buf);
+ printbuf_exit(&buf);
+ }
+}
+
+static int abort_lock(struct lock_graph *g, struct trans_waiting_for_lock *i)
+{
if (i == g->g) {
- trace_and_count(i->trans->c, trans_restart_would_deadlock, i->trans, _RET_IP_);
- ret = btree_trans_restart(i->trans, BCH_ERR_transaction_restart_would_deadlock);
+ trace_would_deadlock(g, i->trans, _RET_IP_);
+ return btree_trans_restart(i->trans, BCH_ERR_transaction_restart_would_deadlock);
} else {
i->trans->lock_must_abort = true;
- ret = 0;
+ wake_up_process(i->trans->locking_wait.task);
+ return 0;
}
+}
- for (i = g->g + 1; i < g->g + g->nr; i++)
- wake_up_process(i->trans->locking_wait.task);
- return ret;
+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)
+static noinline int break_cycle(struct lock_graph *g, struct printbuf *cycle)
{
- struct trans_waiting_for_lock *i;
+ struct trans_waiting_for_lock *i, *abort = NULL;
+ unsigned best = 0, pref;
+ int ret;
- for (i = g->g; i < g->g + g->nr; i++) {
- if (i->trans->lock_may_not_fail ||
- i->trans->locking_wait.lock_want == SIX_LOCK_write)
- continue;
+ if (lock_graph_remove_non_waiters(g))
+ return 0;
- return abort_lock(g, i);
+ /* Only checking, for debugfs: */
+ if (cycle) {
+ print_cycle(cycle, g);
+ ret = -1;
+ goto out;
}
for (i = g->g; i < g->g + g->nr; i++) {
- if (i->trans->lock_may_not_fail ||
- !i->trans->in_traverse_all)
- continue;
-
- return abort_lock(g, i);
+ pref = btree_trans_abort_preference(i->trans);
+ if (pref > best) {
+ abort = i;
+ best = pref;
+ }
}
- for (i = g->g; i < g->g + g->nr; i++) {
- if (i->trans->lock_may_not_fail)
- continue;
+ 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);
- return abort_lock(g, i);
+ 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();
}
- 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 btree_trans *orig_trans = g->g->trans;
struct trans_waiting_for_lock *i;
- int ret = 0;
-
- for (i = g->g; i < g->g + g->nr; i++) {
- if (i->trans->locking != i->node_want)
- while (g->g + g->nr >= i) {
- lock_graph_pop(g);
- return 0;
- }
+ for (i = g->g; i < g->g + g->nr; i++)
if (i->trans == trans) {
- if (cycle) {
- /* Only checking: */
- print_cycle(cycle, g);
- ret = -1;
- } else {
- ret = break_cycle(g);
- }
-
- if (ret)
- goto deadlock;
- /*
- * If we didn't abort (instead telling another
- * transaction to abort), keep checking:
- */
+ closure_put(&trans->ref);
+ return break_cycle(g, cycle);
}
- }
if (g->nr == ARRAY_SIZE(g->g)) {
+ closure_put(&trans->ref);
+
if (orig_trans->lock_may_not_fail)
return 0;
- trace_and_count(trans->c, trans_restart_would_deadlock_recursion_limit, trans, _RET_IP_);
- ret = btree_trans_restart(orig_trans, BCH_ERR_transaction_restart_deadlock_recursion_limit);
- goto deadlock;
- }
+ while (g->nr)
+ lock_graph_up(g);
- closure_get(&trans->ref);
+ if (cycle)
+ return 0;
- g->g[g->nr++] = (struct trans_waiting_for_lock) {
- .trans = trans,
- .node_want = trans->locking,
- .lock_want = trans->locking_wait.lock_want,
- };
+ 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;
-deadlock:
- while (g->nr)
- lock_graph_pop(g);
- return ret;
-}
-
-static noinline void 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_pop(g);
- return;
- }
- BUG();
}
static bool lock_type_conflicts(enum six_lock_type t1, enum six_lock_type t2)
struct lock_graph g;
struct trans_waiting_for_lock *top;
struct btree_bkey_cached_common *b;
- struct btree_path *path;
- int ret;
+ btree_path_idx_t path_idx;
+ int ret = 0;
+
+ g.nr = 0;
if (trans->lock_must_abort) {
- trace_and_count(trans->c, trans_restart_would_deadlock, trans, _RET_IP_);
+ if (cycle)
+ return -1;
+
+ trace_would_deadlock(&g, trans, _RET_IP_);
return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock);
}
- g.nr = 0;
- ret = lock_graph_descend(&g, trans, cycle);
- BUG_ON(ret);
+ lock_graph_down(&g, trans);
+
+ /* trans->paths is rcu protected vs. freeing */
+ rcu_read_lock();
+ if (cycle)
+ cycle->atomic++;
next:
if (!g.nr)
- return 0;
+ goto out;
top = &g.g[g.nr - 1];
- trans_for_each_path_from(top->trans, path, top->path_idx) {
+ struct btree_path *paths = rcu_dereference(top->trans->paths);
+ if (!paths)
+ goto up;
+
+ unsigned long *paths_allocated = trans_paths_allocated(paths);
+
+ trans_for_each_path_idx_from(paths_allocated, *trans_paths_nr(paths),
+ path_idx, top->path_idx) {
+ struct btree_path *path = paths + path_idx;
if (!path->nodes_locked)
continue;
- if (top->path_idx != path->idx) {
- top->path_idx = path->idx;
+ if (path_idx != top->path_idx) {
+ top->path_idx = path_idx;
top->level = 0;
top->lock_start_time = 0;
}
b = &READ_ONCE(path->l[top->level].b)->c;
- if (unlikely(IS_ERR_OR_NULL(b))) {
- lock_graph_remove_non_waiters(&g);
+ 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;
}
!lock_type_conflicts(lock_held, trans->locking_wait.lock_want))
continue;
- ret = lock_graph_descend(&g, trans, cycle);
+ closure_get(&trans->ref);
raw_spin_unlock(&b->lock.wait_lock);
+ ret = lock_graph_descend(&g, trans, cycle);
if (ret)
- return ret < 0 ? ret : 0;
+ goto out;
goto next;
}
raw_spin_unlock(&b->lock.wait_lock);
}
}
-
- lock_graph_pop(&g);
+up:
+ if (g.nr > 1 && cycle)
+ print_chain(cycle, &g);
+ lock_graph_up(&g);
goto next;
+out:
+ if (cycle)
+ --cycle->atomic;
+ rcu_read_unlock();
+ return ret;
}
int bch2_six_check_for_deadlock(struct six_lock *lock, void *p)
* locked:
*/
six_lock_readers_add(&b->lock, -readers);
- ret = __btree_node_lock_nopath(trans, b, SIX_LOCK_write, lock_may_not_fail);
+ 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);
+ mark_btree_node_locked_noreset(path, b->level, BTREE_NODE_INTENT_LOCKED);
return ret;
}
+void bch2_btree_node_lock_write_nofail(struct btree_trans *trans,
+ struct btree_path *path,
+ struct btree_bkey_cached_common *b)
+{
+ struct btree_path *linked;
+ unsigned i, iter;
+ int ret;
+
+ /*
+ * XXX BIG FAT NOTICE
+ *
+ * Drop all read locks before taking a write lock:
+ *
+ * This is a hack, because bch2_btree_node_lock_write_nofail() is a
+ * hack - but by dropping read locks first, this should never fail, and
+ * we only use this in code paths where whatever read locks we've
+ * already taken are no longer needed:
+ */
+
+ trans_for_each_path(trans, linked, iter) {
+ if (!linked->nodes_locked)
+ continue;
+
+ for (i = 0; i < BTREE_MAX_DEPTH; i++)
+ if (btree_node_read_locked(linked, i)) {
+ btree_node_unlock(trans, linked, i);
+ btree_path_set_dirty(linked, BTREE_ITER_NEED_RELOCK);
+ }
+ }
+
+ ret = __btree_node_lock_write(trans, path, b, true);
+ BUG_ON(ret);
+}
+
/* relock */
static inline bool btree_path_get_locks(struct btree_trans *trans,
struct btree_path *path,
- bool upgrade)
+ bool upgrade,
+ struct get_locks_fail *f)
{
unsigned l = path->level;
int fail_idx = -1;
if (!(upgrade
? bch2_btree_node_upgrade(trans, path, l)
- : bch2_btree_node_relock(trans, path, l)))
- fail_idx = l;
+ : bch2_btree_node_relock(trans, path, l))) {
+ fail_idx = l;
+
+ if (f) {
+ f->l = l;
+ f->b = path->l[l].b;
+ }
+ }
l++;
} while (l < path->locks_want);
return true;
}
fail:
- if (trace)
+ if (trace && !trans->notrace_relock_fail)
trace_and_count(trans->c, btree_path_relock_fail, trans, _RET_IP_, path, level);
return false;
}
trace_and_count(trans->c, btree_path_upgrade_fail, trans, _RET_IP_, path, level);
return false;
success:
- mark_btree_node_locked_noreset(path, level, SIX_LOCK_intent);
+ mark_btree_node_locked_noreset(path, level, BTREE_NODE_INTENT_LOCKED);
return true;
}
bool bch2_btree_path_relock_norestart(struct btree_trans *trans,
struct btree_path *path, unsigned long trace_ip)
{
- return btree_path_get_locks(trans, path, false);
+ struct get_locks_fail f;
+
+ return btree_path_get_locks(trans, path, false, &f);
}
-__flatten
-bool bch2_btree_path_upgrade_norestart(struct btree_trans *trans,
+int __bch2_btree_path_relock(struct btree_trans *trans,
struct btree_path *path, unsigned long trace_ip)
{
- return btree_path_get_locks(trans, path, true);
+ 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;
}
bool bch2_btree_path_upgrade_noupgrade_sibs(struct btree_trans *trans,
struct btree_path *path,
- unsigned new_locks_want)
+ unsigned new_locks_want,
+ struct get_locks_fail *f)
{
EBUG_ON(path->locks_want >= new_locks_want);
path->locks_want = new_locks_want;
- return btree_path_get_locks(trans, path, true);
+ return btree_path_get_locks(trans, path, true, f);
}
bool __bch2_btree_path_upgrade(struct btree_trans *trans,
struct btree_path *path,
- unsigned new_locks_want)
+ unsigned new_locks_want,
+ struct get_locks_fail *f)
{
- struct btree_path *linked;
-
- if (bch2_btree_path_upgrade_noupgrade_sibs(trans, path, new_locks_want))
+ if (bch2_btree_path_upgrade_noupgrade_sibs(trans, path, new_locks_want, f))
return true;
/*
* before interior nodes - now that's handled by
* bch2_btree_path_traverse_all().
*/
- if (!path->cached && !trans->in_traverse_all)
- trans_for_each_path(trans, linked)
+ if (!path->cached && !trans->in_traverse_all) {
+ struct btree_path *linked;
+ unsigned i;
+
+ trans_for_each_path(trans, linked, i)
if (linked != path &&
linked->cached == path->cached &&
linked->btree_id == path->btree_id &&
linked->locks_want < new_locks_want) {
linked->locks_want = new_locks_want;
- btree_path_get_locks(trans, linked, true);
+ btree_path_get_locks(trans, linked, true, NULL);
}
+ }
return false;
}
struct btree_path *path,
unsigned new_locks_want)
{
- unsigned l;
+ unsigned l, old_locks_want = path->locks_want;
+
+ if (trans->restarted)
+ return;
EBUG_ON(path->locks_want < new_locks_want);
} else {
if (btree_node_intent_locked(path, l)) {
six_lock_downgrade(&path->l[l].b->c.lock);
- mark_btree_node_locked_noreset(path, l, SIX_LOCK_read);
+ mark_btree_node_locked_noreset(path, l, BTREE_NODE_READ_LOCKED);
}
break;
}
}
bch2_btree_path_verify_locks(path);
+
+ trace_path_downgrade(trans, _RET_IP_, path, old_locks_want);
}
/* Btree transaction locking: */
void bch2_trans_downgrade(struct btree_trans *trans)
{
struct btree_path *path;
+ unsigned i;
+
+ if (trans->restarted)
+ return;
- trans_for_each_path(trans, path)
+ trans_for_each_path(trans, path, i)
bch2_btree_path_downgrade(trans, path);
}
int bch2_trans_relock(struct btree_trans *trans)
{
struct btree_path *path;
+ unsigned i;
if (unlikely(trans->restarted))
- return - ((int) trans->restarted);
+ return -((int) trans->restarted);
- trans_for_each_path(trans, path)
+ trans_for_each_path(trans, path, i)
if (path->should_be_locked &&
!bch2_btree_path_relock_norestart(trans, path, _RET_IP_)) {
trace_and_count(trans->c, trans_restart_relock, trans, _RET_IP_, path);
return 0;
}
+int bch2_trans_relock_notrace(struct btree_trans *trans)
+{
+ struct btree_path *path;
+ unsigned i;
+
+ if (unlikely(trans->restarted))
+ return -((int) trans->restarted);
+
+ trans_for_each_path(trans, path, i)
+ 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_noassert(struct btree_trans *trans)
+{
+ struct btree_path *path;
+ unsigned i;
+
+ trans_for_each_path(trans, path, i)
+ __bch2_btree_path_unlock(trans, path);
+}
+
void bch2_trans_unlock(struct btree_trans *trans)
{
struct btree_path *path;
+ unsigned i;
- trans_for_each_path(trans, path)
+ trans_for_each_path(trans, path, i)
__bch2_btree_path_unlock(trans, 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));
+void bch2_trans_unlock_long(struct btree_trans *trans)
+{
+ bch2_trans_unlock(trans);
+ bch2_trans_srcu_unlock(trans);
+}
+
+bool bch2_trans_locked(struct btree_trans *trans)
+{
+ struct btree_path *path;
+ unsigned i;
+
+ trans_for_each_path(trans, path, i)
+ if (path->nodes_locked)
+ return true;
+ return false;
+}
+
+int __bch2_trans_mutex_lock(struct btree_trans *trans,
+ struct mutex *lock)
+{
+ int ret = drop_locks_do(trans, (mutex_lock(lock), 0));
+
+ if (ret)
+ mutex_unlock(lock);
+ return ret;
}
/* Debug */
void bch2_trans_verify_locks(struct btree_trans *trans)
{
struct btree_path *path;
+ unsigned i;
- trans_for_each_path(trans, path)
+ trans_for_each_path(trans, path, i)
bch2_btree_path_verify_locks(path);
}