Quelle  btree_locking.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0

#include "bcachefs.h"
#include "btree_cache.h"
#include "btree_locking.h"
#include "btree_types.h"

static struct lock_class_key bch2_btree_node_lock_key;

void bch2_btree_lock_init(struct btree_bkey_cached_common *b,
     enum six_lock_init_flags flags,
     gfp_t gfp)
{
 __six_lock_init(&b->lock, "b->c.lock", &bch2_btree_node_lock_key, flags, gfp);
 lockdep_set_notrack_class(&b->lock);
}

/* Btree node locking: */

struct six_lock_count bch2_btree_node_lock_counts(struct btree_trans *trans,
        struct btree_path *skip,
        struct btree_bkey_cached_common *b,
        unsigned level)
{
 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, i)
  if (path != skip && &path->l[level].b->c == b) {
   int t = btree_node_locked_type(path, level);

   if (t != BTREE_NODE_UNLOCKED)
    ret.n[t]++;
  }

 return ret;
}

/* unlock */

void bch2_btree_node_unlock_write(struct btree_trans *trans,
   struct btree_path *path, struct btree *b)
{
 bch2_btree_node_unlock_write_inlined(trans, path, b);
}

/* lock */

/*
 * @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)
{
 struct trans_waiting_for_lock *i;

 prt_printf(out, "Found lock cycle (%u entries):\n", g->nr);

 for (i = g->g; i < g->g + g->nr; i++) {
  struct task_struct *task = READ_ONCE(i->trans->locking_wait.task);
  if (!task)
   continue;

  bch2_btree_trans_to_text(out, i->trans);
  bch2_prt_task_backtrace(out, task, i == g->g ? 5 : 1, GFP_NOWAIT);
 }
}

static noinline void print_chain(struct printbuf *out, struct lock_graph *g)
{
 struct trans_waiting_for_lock *i;

 for (i = g->g; i != g->g + g->nr; i++) {
  struct task_struct *task = READ_ONCE(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 noinline void lock_graph_pop_from(struct lock_graph *g, struct trans_waiting_for_lock *i)
{
 while (g->g + g->nr > i)
  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 *from)
{
 struct trans_waiting_for_lock *i;

 if (from->trans->locking != from->node_want) {
  lock_graph_pop_from(g, from);
  return true;
 }

 for (i = from + 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) {
   lock_graph_pop_from(g, i);
   return true;
  }

 return false;
}

static void trace_would_deadlock(struct lock_graph *g, struct btree_trans *trans)
{
 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, buf.buf);
  printbuf_exit(&buf);
 }
}

static int abort_lock(struct lock_graph *g, struct trans_waiting_for_lock *i)
{
 if (i == g->g) {
  trace_would_deadlock(g, i->trans);
  return btree_trans_restart_foreign_task(i->trans,
     BCH_ERR_transaction_restart_would_deadlock,
     _THIS_IP_);
 } 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 __noreturn void break_cycle_fail(struct lock_graph *g)
{
 struct printbuf buf = PRINTBUF;
 buf.atomic++;

 prt_printf(&buf, bch2_fmt(g->g->trans->c, "cycle of nofail locks"));

 for (struct trans_waiting_for_lock *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:\n");
  printbuf_indent_add(&buf, 2);
  bch2_prt_task_backtrace(&buf, trans->locking_wait.task, 2, GFP_NOWAIT);
  printbuf_indent_sub(&buf, 2);
  prt_newline(&buf);
 }

 bch2_print_str(g->g->trans->c, KERN_ERR, buf.buf);
 printbuf_exit(&buf);
 BUG();
}

static noinline int break_cycle(struct lock_graph *g, struct printbuf *cycle,
    struct trans_waiting_for_lock *from)
{
 struct trans_waiting_for_lock *i, *abort = NULL;
 unsigned best = 0, pref;
 int ret;

 if (lock_graph_remove_non_waiters(g, from))
  return 0;

 /* Only checking, for debugfs: */
 if (cycle) {
  print_cycle(cycle, g);
  ret = -1;
  goto out;
 }

 for (i = from; i < g->g + g->nr; i++) {
  pref = btree_trans_abort_preference(i->trans);
  if (pref > best) {
   abort = i;
   best = pref;
  }
 }

 if (unlikely(!best))
  break_cycle_fail(g);

 ret = abort_lock(g, abort);
out:
 if (ret)
  lock_graph_pop_all(g);
 else
  lock_graph_pop_from(g, abort);
 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;

 for (struct trans_waiting_for_lock *i = g->g; i < g->g + g->nr; i++)
  if (i->trans == trans) {
   closure_put(&trans->ref);
   return break_cycle(g, cycle, i);
  }

 if (unlikely(g->nr == ARRAY_SIZE(g->g))) {
  closure_put(&trans->ref);

  if (orig_trans->lock_may_not_fail)
   return 0;

  lock_graph_pop_all(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;
 btree_path_idx_t path_idx;
 int ret = 0;

 g.nr = 0;

 if (trans->lock_must_abort && !trans->lock_may_not_fail) {
  if (cycle)
   return -1;

  trace_would_deadlock(&g, trans);
  return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock);
 }

 lock_graph_down(&g, trans);

 /* trans->paths is rcu protected vs. freeing */
 guard(rcu)();
 if (cycle)
  cycle->atomic++;
next:
 if (!g.nr)
  goto out;

 top = &g.g[g.nr - 1];

 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 (path_idx != top->path_idx) {
   top->path_idx  = path_idx;
   top->level  = 0;
   top->lock_start_time = 0;
  }

  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;

   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, g.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;
   }

   if (list_empty_careful(&b->lock.wait_list))
    continue;

   raw_spin_lock(&b->lock.wait_lock);
   list_for_each_entry(trans, &b->lock.wait_list, locking_wait.list) {
    BUG_ON(b != trans->locking);

    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;

    closure_get(&trans->ref);
    raw_spin_unlock(&b->lock.wait_lock);

    ret = lock_graph_descend(&g, trans, cycle);
    if (ret)
     goto out;
    goto next;

   }
   raw_spin_unlock(&b->lock.wait_lock);
  }
 }
up:
 if (g.nr > 1 && cycle)
  print_chain(cycle, &g);
 lock_graph_up(&g);
 goto next;
out:
 if (cycle)
  --cycle->atomic;
 return ret;
}

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, 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)
{
 int ret = __btree_node_lock_write(trans, path, b, true);
 BUG_ON(ret);
}

/* relock */

static int btree_path_get_locks(struct btree_trans *trans,
    struct btree_path *path,
    bool upgrade,
    struct get_locks_fail *f,
    int restart_err)
{
 unsigned l = path->level;

 do {
  if (!btree_path_node(path, l))
   break;

  if (!(upgrade
        ? bch2_btree_node_upgrade(trans, path, l)
        : bch2_btree_node_relock(trans, path, l)))
   goto err;

  l++;
 } while (l < path->locks_want);

 if (path->uptodate == BTREE_ITER_NEED_RELOCK)
  path->uptodate = BTREE_ITER_UPTODATE;

 return path->uptodate < BTREE_ITER_NEED_RELOCK ? 0 : -1;
err:
 if (f) {
  f->l = l;
  f->b = path->l[l].b;
 }

 /*
 * Do transaction restart before unlocking, so we don't pop
 * should_be_locked asserts
 */
 if (restart_err) {
  btree_trans_restart(trans, restart_err);
 } else if (path->should_be_locked && !trans->restarted) {
  if (upgrade)
   path->locks_want = l;
  return -1;
 }

 __bch2_btree_path_unlock(trans, path);
 btree_path_set_dirty(trans, path, BTREE_ITER_NEED_TRAVERSE);

 /*
 * When we fail to get a lock, we have to ensure that any child nodes
 * can't be relocked so bch2_btree_path_traverse has to walk back up to
 * the node that we failed to relock:
 */
 do {
  path->l[l].b = upgrade
   ? ERR_PTR(-BCH_ERR_no_btree_node_upgrade)
   : ERR_PTR(-BCH_ERR_no_btree_node_relock);
 } while (l--);

 return -restart_err ?: -1;
}

bool __bch2_btree_node_relock(struct btree_trans *trans,
         struct btree_path *path, unsigned level,
         bool trace)
{
 struct btree *b = btree_path_node(path, level);
 int want = __btree_lock_want(path, level);

 if (race_fault())
  goto fail;

 if (six_relock_type(&b->c.lock, want, path->l[level].lock_seq) ||
     (btree_node_lock_seq_matches(path, b, level) &&
      btree_node_lock_increment(trans, &b->c, level, want))) {
  mark_btree_node_locked(trans, path, level, want);
  return true;
 }
fail:
 if (trace && !trans->notrace_relock_fail)
  trace_and_count(trans->c, btree_path_relock_fail, trans, _RET_IP_, path, level);
 return false;
}

/* upgrade */

bool bch2_btree_node_upgrade(struct btree_trans *trans,
        struct btree_path *path, unsigned level)
{
 struct btree *b = path->l[level].b;

 if (!is_btree_node(path, level))
  return false;

 switch (btree_lock_want(path, level)) {
 case BTREE_NODE_UNLOCKED:
  BUG_ON(btree_node_locked(path, level));
  return true;
 case BTREE_NODE_READ_LOCKED:
  BUG_ON(btree_node_intent_locked(path, level));
  return bch2_btree_node_relock(trans, path, level);
 case BTREE_NODE_INTENT_LOCKED:
  break;
 case BTREE_NODE_WRITE_LOCKED:
  BUG();
 }

 if (btree_node_intent_locked(path, level))
  return true;

 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_lock_seq_matches(path, b, level) &&
     btree_node_lock_increment(trans, &b->c, level, BTREE_NODE_INTENT_LOCKED)) {
  btree_node_unlock(trans, path, level);
  goto success;
 }

 trace_and_count(trans->c, btree_path_upgrade_fail, trans, _RET_IP_, path, level);
 return false;
success:
 mark_btree_node_locked_noreset(path, level, BTREE_NODE_INTENT_LOCKED);
 return true;
}

/* Btree path locking: */

/*
 * Only for btree_cache.c - only relocks intent locks
 */
int bch2_btree_path_relock_intent(struct btree_trans *trans,
      struct btree_path *path)
{
 unsigned l;

 for (l = path->level;
      l < path->locks_want && btree_path_node(path, l);
      l++) {
  if (!bch2_btree_node_relock(trans, path, l)) {
   __bch2_btree_path_unlock(trans, path);
   btree_path_set_dirty(trans, path, BTREE_ITER_NEED_TRAVERSE);
   trace_and_count(trans->c, trans_restart_relock_path_intent, trans, _RET_IP_, path);
   return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_path_intent);
  }
 }

 return 0;
}

__flatten
bool bch2_btree_path_relock_norestart(struct btree_trans *trans, struct btree_path *path)
{
 bool ret = !btree_path_get_locks(trans, path, false, NULL, 0);
 bch2_trans_verify_locks(trans);
 return ret;
}

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_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_norestart(struct btree_trans *trans,
      struct btree_path *path,
      unsigned new_locks_want)
{
 path->locks_want = new_locks_want;

 /*
 * If we need it locked, we can't touch it. Otherwise, we can return
 * success - bch2_path_get() will use this path, and it'll just be
 * retraversed:
 */
 bool ret = !btree_path_get_locks(trans, path, true, NULL, 0) ||
  !path->should_be_locked;

 bch2_btree_path_verify_locks(trans, path);
 return ret;
}

int __bch2_btree_path_upgrade(struct btree_trans *trans,
         struct btree_path *path,
         unsigned new_locks_want)
{
 unsigned old_locks = path->nodes_locked;
 unsigned old_locks_want = path->locks_want;

 path->locks_want = max_t(unsigned, path->locks_want, new_locks_want);

 struct get_locks_fail f = {};
 int ret = btree_path_get_locks(trans, path, true, &f,
    BCH_ERR_transaction_restart_upgrade);
 if (!ret)
  goto out;

 /*
 * XXX: this is ugly - we'd prefer to not be mucking with other
 * iterators in the btree_trans here.
 *
 * On failure to upgrade the iterator, setting iter->locks_want and
 * calling get_locks() is sufficient to make bch2_btree_path_traverse()
 * get the locks we want on transaction restart.
 *
 * But if this iterator was a clone, on transaction restart what we did
 * to this iterator isn't going to be preserved.
 *
 * Possibly we could add an iterator field for the parent iterator when
 * an iterator is a copy - for now, we'll just upgrade any other
 * iterators with the same btree id.
 *
 * The code below used to be needed to ensure ancestor nodes get locked
 * before interior nodes - now that's handled by
 * bch2_btree_path_traverse_all().
 */
 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, NULL, 0);
   }
 }

 count_event(trans->c, trans_restart_upgrade);
 if (trace_trans_restart_upgrade_enabled()) {
  struct printbuf buf = PRINTBUF;

  prt_printf(&buf, "%s %pS\n", trans->fn, (void *) _RET_IP_);
  prt_printf(&buf, "btree %s pos\n", bch2_btree_id_str(path->btree_id));
  bch2_bpos_to_text(&buf, path->pos);
  prt_printf(&buf, "locks want %u -> %u level %u\n",
      old_locks_want, new_locks_want, f.l);
  prt_printf(&buf, "nodes_locked %x -> %x\n",
      old_locks, path->nodes_locked);
  prt_printf(&buf, "node %s ", IS_ERR(f.b) ? bch2_err_str(PTR_ERR(f.b)) :
      !f.b ? "(null)" : "(node)");
  prt_printf(&buf, "path seq %u node seq %u\n",
      IS_ERR_OR_NULL(f.b) ? 0 : f.b->c.lock.seq,
      path->l[f.l].lock_seq);

  trace_trans_restart_upgrade(trans->c, buf.buf);
  printbuf_exit(&buf);
 }
out:
 bch2_trans_verify_locks(trans);
 return ret;
}

void __bch2_btree_path_downgrade(struct btree_trans *trans,
     struct btree_path *path,
     unsigned new_locks_want)
{
 unsigned l, old_locks_want = path->locks_want;

 if (trans->restarted)
  return;

 EBUG_ON(path->locks_want < new_locks_want);

 path->locks_want = new_locks_want;

 while (path->nodes_locked &&
        (l = btree_path_highest_level_locked(path)) >= path->locks_want) {
  if (l > path->level) {
   btree_node_unlock(trans, path, l);
  } else {
   if (btree_node_intent_locked(path, l)) {
    six_lock_downgrade(&path->l[l].b->c.lock);
    mark_btree_node_locked_noreset(path, l, BTREE_NODE_READ_LOCKED);
   }
   break;
  }
 }

 bch2_btree_path_verify_locks(trans, 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, i)
  if (path->ref)
   bch2_btree_path_downgrade(trans, path);
}

static inline void __bch2_trans_unlock(struct btree_trans *trans)
{
 struct btree_path *path;
 unsigned i;

 trans_for_each_path(trans, path, i)
  __bch2_btree_path_unlock(trans, path);
}

static noinline __cold void bch2_trans_relock_fail(struct btree_trans *trans, struct btree_path *path,
         struct get_locks_fail *f, bool trace, ulong ip)
{
 if (!trace)
  goto out;

 if (trace_trans_restart_relock_enabled()) {
  struct printbuf buf = PRINTBUF;

  bch2_bpos_to_text(&buf, path->pos);
  prt_printf(&buf, " %s l=%u seq=%u node seq=",
      bch2_btree_id_str(path->btree_id),
      f->l, path->l[f->l].lock_seq);
  if (IS_ERR_OR_NULL(f->b)) {
   prt_str(&buf, bch2_err_str(PTR_ERR(f->b)));
  } else {
   prt_printf(&buf, "%u", f->b->c.lock.seq);

   struct six_lock_count c =
    bch2_btree_node_lock_counts(trans, NULL, &f->b->c, f->l);
   prt_printf(&buf, " self locked %u.%u.%u", c.n[0], c.n[1], c.n[2]);

   c = six_lock_counts(&f->b->c.lock);
   prt_printf(&buf, " total locked %u.%u.%u", c.n[0], c.n[1], c.n[2]);
  }

  trace_trans_restart_relock(trans, ip, buf.buf);
  printbuf_exit(&buf);
 }

 count_event(trans->c, trans_restart_relock);
out:
 __bch2_trans_unlock(trans);
 bch2_trans_verify_locks(trans);
}

static inline int __bch2_trans_relock(struct btree_trans *trans, bool trace, ulong ip)
{
 bch2_trans_verify_locks(trans);

 if (unlikely(trans->restarted))
  return -((int) trans->restarted);
 if (unlikely(trans->locked))
  goto out;

 struct btree_path *path;
 unsigned i;

 trans_for_each_path(trans, path, i) {
  struct get_locks_fail f;
  int ret;

  if (path->should_be_locked &&
      (ret = btree_path_get_locks(trans, path, false, &f,
     BCH_ERR_transaction_restart_relock))) {
   bch2_trans_relock_fail(trans, path, &f, trace, ip);
   return ret;
  }
 }

 trans_set_locked(trans, true);
out:
 bch2_trans_verify_locks(trans);
 return 0;
}

int bch2_trans_relock(struct btree_trans *trans)
{
 return __bch2_trans_relock(trans, true, _RET_IP_);
}

int bch2_trans_relock_notrace(struct btree_trans *trans)
{
 return __bch2_trans_relock(trans, false, _RET_IP_);
}

void bch2_trans_unlock(struct btree_trans *trans)
{
 trans_set_unlocked(trans);

 __bch2_trans_unlock(trans);
}

void bch2_trans_unlock_long(struct btree_trans *trans)
{
 bch2_trans_unlock(trans);
 bch2_trans_srcu_unlock(trans);
}

void bch2_trans_unlock_write(struct btree_trans *trans)
{
 struct btree_path *path;
 unsigned i;

 trans_for_each_path(trans, path, i)
  for (unsigned l = 0; l < BTREE_MAX_DEPTH; l++)
   if (btree_node_write_locked(path, l))
    bch2_btree_node_unlock_write(trans, path, path->l[l].b);
}

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_btree_path_verify_locks(struct btree_trans *trans, struct btree_path *path)
{
 if (!path->nodes_locked && btree_path_node(path, path->level)) {
  /*
 * A path may be uptodate and yet have nothing locked if and only if
 * there is no node at path->level, which generally means we were
 * iterating over all nodes and got to the end of the btree
 */
  BUG_ON(path->uptodate == BTREE_ITER_UPTODATE);
  BUG_ON(path->should_be_locked && trans->locked && !trans->restarted);
 }

 if (!path->nodes_locked)
  return;

 for (unsigned l = 0; l < BTREE_MAX_DEPTH; l++) {
  int want = btree_lock_want(path, l);
  int have = btree_node_locked_type_nowrite(path, l);

  BUG_ON(!is_btree_node(path, l) && have != BTREE_NODE_UNLOCKED);

  BUG_ON(is_btree_node(path, l) && want != have);

  BUG_ON(btree_node_locked(path, l) &&
         path->l[l].lock_seq != six_lock_seq(&path->l[l].b->c.lock));
 }
}

static 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;
}

void __bch2_trans_verify_locks(struct btree_trans *trans)
{
 if (!trans->locked) {
  BUG_ON(bch2_trans_locked(trans));
  return;
 }

 struct btree_path *path;
 unsigned i;

 trans_for_each_path(trans, path, i)
  __bch2_btree_path_verify_locks(trans, path);
}