Quelle  btree_iter.c

// SPDX-License-Identifier: GPL-2.0

#include "bcachefs.h"
#include "bkey_methods.h"
#include "bkey_buf.h"
#include "btree_cache.h"
#include "btree_iter.h"
#include "btree_journal_iter.h"
#include "btree_key_cache.h"
#include "btree_locking.h"
#include "btree_update.h"
#include "debug.h"
#include "error.h"
#include "extents.h"
#include "journal.h"
#include "journal_io.h"
#include "replicas.h"
#include "snapshot.h"
#include "super.h"
#include "trace.h"

#include <linux/random.h>
#include <linux/prefetch.h>

static inline void btree_path_list_remove(struct btree_trans *, struct btree_path *);
static inline void btree_path_list_add(struct btree_trans *,
   btree_path_idx_t, btree_path_idx_t);

static inline unsigned long btree_iter_ip_allocated(struct btree_iter *iter)
{
#ifdef TRACK_PATH_ALLOCATED
 return iter->ip_allocated;
#else
 return 0;
#endif
}

static btree_path_idx_t btree_path_alloc(struct btree_trans *, btree_path_idx_t);
static void bch2_trans_srcu_lock(struct btree_trans *);

static inline int __btree_path_cmp(const struct btree_path *l,
       enum btree_id r_btree_id,
       bool   r_cached,
       struct bpos  r_pos,
       unsigned  r_level)
{
 /*
 * Must match lock ordering as defined by __bch2_btree_node_lock:
 */
 return   cmp_int(l->btree_id, r_btree_id) ?:
   cmp_int((int) l->cached, (int) r_cached) ?:
   bpos_cmp(l->pos, r_pos) ?:
  -cmp_int(l->level, r_level);
}

static inline int btree_path_cmp(const struct btree_path *l,
     const struct btree_path *r)
{
 return __btree_path_cmp(l, r->btree_id, r->cached, r->pos, r->level);
}

static inline struct bpos bkey_successor(struct btree_iter *iter, struct bpos p)
{
 /* Are we iterating over keys in all snapshots? */
 if (iter->flags & BTREE_ITER_all_snapshots) {
  p = bpos_successor(p);
 } else {
  p = bpos_nosnap_successor(p);
  p.snapshot = iter->snapshot;
 }

 return p;
}

static inline struct bpos bkey_predecessor(struct btree_iter *iter, struct bpos p)
{
 /* Are we iterating over keys in all snapshots? */
 if (iter->flags & BTREE_ITER_all_snapshots) {
  p = bpos_predecessor(p);
 } else {
  p = bpos_nosnap_predecessor(p);
  p.snapshot = iter->snapshot;
 }

 return p;
}

static inline struct bpos btree_iter_search_key(struct btree_iter *iter)
{
 struct bpos pos = iter->pos;

 if ((iter->flags & BTREE_ITER_is_extents) &&
     !bkey_eq(pos, POS_MAX))
  pos = bkey_successor(iter, pos);
 return pos;
}

static inline bool btree_path_pos_before_node(struct btree_path *path,
           struct btree *b)
{
 return bpos_lt(path->pos, b->data->min_key);
}

static inline bool btree_path_pos_after_node(struct btree_path *path,
          struct btree *b)
{
 return bpos_gt(path->pos, b->key.k.p);
}

static inline bool btree_path_pos_in_node(struct btree_path *path,
       struct btree *b)
{
 return path->btree_id == b->c.btree_id &&
  !btree_path_pos_before_node(path, b) &&
  !btree_path_pos_after_node(path, b);
}

/* Debug: */

static void __bch2_btree_path_verify_cached(struct btree_trans *trans,
       struct btree_path *path)
{
 struct bkey_cached *ck;
 bool locked = btree_node_locked(path, 0);

 if (!bch2_btree_node_relock(trans, path, 0))
  return;

 ck = (void *) path->l[0].b;
 BUG_ON(ck->key.btree_id != path->btree_id ||
        !bkey_eq(ck->key.pos, path->pos));

 if (!locked)
  btree_node_unlock(trans, path, 0);
}

static void __bch2_btree_path_verify_level(struct btree_trans *trans,
    struct btree_path *path, unsigned level)
{
 struct btree_path_level *l;
 struct btree_node_iter tmp;
 bool locked;
 struct bkey_packed *p, *k;
 struct printbuf buf1 = PRINTBUF;
 struct printbuf buf2 = PRINTBUF;
 struct printbuf buf3 = PRINTBUF;
 const char *msg;

 l = &path->l[level];
 tmp = l->iter;
 locked = btree_node_locked(path, level);

 if (path->cached) {
  if (!level)
   __bch2_btree_path_verify_cached(trans, path);
  return;
 }

 if (!btree_path_node(path, level))
  return;

 if (!bch2_btree_node_relock_notrace(trans, path, level))
  return;

 BUG_ON(!btree_path_pos_in_node(path, l->b));

 bch2_btree_node_iter_verify(&l->iter, l->b);

 /*
 * For interior nodes, the iterator will have skipped past deleted keys:
 */
 p = level
  ? bch2_btree_node_iter_prev(&tmp, l->b)
  : bch2_btree_node_iter_prev_all(&tmp, l->b);
 k = bch2_btree_node_iter_peek_all(&l->iter, l->b);

 if (p && bkey_iter_pos_cmp(l->b, p, &path->pos) >= 0) {
  msg = "before";
  goto err;
 }

 if (k && bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
  msg = "after";
  goto err;
 }

 if (!locked)
  btree_node_unlock(trans, path, level);
 return;
err:
 bch2_bpos_to_text(&buf1, path->pos);

 if (p) {
  struct bkey uk = bkey_unpack_key(l->b, p);

  bch2_bkey_to_text(&buf2, &uk);
 } else {
  prt_printf(&buf2, "(none)");
 }

 if (k) {
  struct bkey uk = bkey_unpack_key(l->b, k);

  bch2_bkey_to_text(&buf3, &uk);
 } else {
  prt_printf(&buf3, "(none)");
 }

 panic("path should be %s key at level %u:\n"
       "path pos %s\n"
       "prev key %s\n"
       "cur  key %s\n",
       msg, level, buf1.buf, buf2.buf, buf3.buf);
}

static void __bch2_btree_path_verify(struct btree_trans *trans,
       struct btree_path *path)
{
 struct bch_fs *c = trans->c;

 for (unsigned i = 0; i < (!path->cached ? BTREE_MAX_DEPTH : 1); i++) {
  if (!path->l[i].b) {
   BUG_ON(!path->cached &&
          bch2_btree_id_root(c, path->btree_id)->b->c.level > i);
   break;
  }

  __bch2_btree_path_verify_level(trans, path, i);
 }

 bch2_btree_path_verify_locks(trans, path);
}

void __bch2_trans_verify_paths(struct btree_trans *trans)
{
 struct btree_path *path;
 unsigned iter;

 trans_for_each_path(trans, path, iter)
  __bch2_btree_path_verify(trans, path);
}

static void __bch2_btree_iter_verify(struct btree_trans *trans, struct btree_iter *iter)
{
 BUG_ON(!!(iter->flags & BTREE_ITER_cached) != btree_iter_path(trans, iter)->cached);

 BUG_ON((iter->flags & BTREE_ITER_is_extents) &&
        (iter->flags & BTREE_ITER_all_snapshots));

 BUG_ON(!(iter->flags & BTREE_ITER_snapshot_field) &&
        (iter->flags & BTREE_ITER_all_snapshots) &&
        !btree_type_has_snapshot_field(iter->btree_id));

 if (iter->update_path)
  __bch2_btree_path_verify(trans, &trans->paths[iter->update_path]);
 __bch2_btree_path_verify(trans, btree_iter_path(trans, iter));
}

static void __bch2_btree_iter_verify_entry_exit(struct btree_iter *iter)
{
 BUG_ON((iter->flags & BTREE_ITER_filter_snapshots) &&
        !iter->pos.snapshot);

 BUG_ON(!(iter->flags & BTREE_ITER_all_snapshots) &&
        iter->pos.snapshot != iter->snapshot);

 BUG_ON(iter->flags & BTREE_ITER_all_snapshots ? !bpos_eq(iter->pos, iter->k.p) :
        !(iter->flags & BTREE_ITER_is_extents) ? !bkey_eq(iter->pos, iter->k.p) :
        (bkey_lt(iter->pos, bkey_start_pos(&iter->k)) ||
  bkey_gt(iter->pos, iter->k.p)));
}

static int __bch2_btree_iter_verify_ret(struct btree_trans *trans,
     struct btree_iter *iter, struct bkey_s_c k)
{
 struct btree_iter copy;
 struct bkey_s_c prev;
 int ret = 0;

 if (!(iter->flags & BTREE_ITER_filter_snapshots))
  return 0;

 if (bkey_err(k) || !k.k)
  return 0;

 BUG_ON(!bch2_snapshot_is_ancestor(trans->c,
       iter->snapshot,
       k.k->p.snapshot));

 bch2_trans_iter_init(trans, ©, iter->btree_id, iter->pos,
        BTREE_ITER_nopreserve|
        BTREE_ITER_all_snapshots);
 prev = bch2_btree_iter_prev(trans, ©);
 if (!prev.k)
  goto out;

 ret = bkey_err(prev);
 if (ret)
  goto out;

 if (bkey_eq(prev.k->p, k.k->p) &&
     bch2_snapshot_is_ancestor(trans->c, iter->snapshot,
          prev.k->p.snapshot) > 0) {
  struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;

  bch2_bkey_to_text(&buf1, k.k);
  bch2_bkey_to_text(&buf2, prev.k);

  panic("iter snap %u\n"
        "k    %s\n"
        "prev %s\n",
        iter->snapshot,
        buf1.buf, buf2.buf);
 }
out:
 bch2_trans_iter_exit(trans, ©);
 return ret;
}

void __bch2_assert_pos_locked(struct btree_trans *trans, enum btree_id id,
       struct bpos pos)
{
 bch2_trans_verify_not_unlocked_or_in_restart(trans);

 struct btree_path *path;
 struct trans_for_each_path_inorder_iter iter;
 struct printbuf buf = PRINTBUF;

 btree_trans_sort_paths(trans);

 trans_for_each_path_inorder(trans, path, iter) {
  if (path->btree_id != id ||
      !btree_node_locked(path, 0) ||
      !path->should_be_locked)
   continue;

  if (!path->cached) {
   if (bkey_ge(pos, path->l[0].b->data->min_key) &&
       bkey_le(pos, path->l[0].b->key.k.p))
    return;
  } else {
   if (bkey_eq(pos, path->pos))
    return;
  }
 }

 bch2_dump_trans_paths_updates(trans);
 bch2_bpos_to_text(&buf, pos);

 panic("not locked: %s %s\n", bch2_btree_id_str(id), buf.buf);
}

static inline void bch2_btree_path_verify_level(struct btree_trans *trans,
      struct btree_path *path, unsigned l)
{
 if (static_branch_unlikely(&bch2_debug_check_iterators))
  __bch2_btree_path_verify_level(trans, path, l);
}

static inline void bch2_btree_path_verify(struct btree_trans *trans,
       struct btree_path *path)
{
 if (static_branch_unlikely(&bch2_debug_check_iterators))
  __bch2_btree_path_verify(trans, path);
}

static inline void bch2_btree_iter_verify(struct btree_trans *trans,
       struct btree_iter *iter)
{
 if (static_branch_unlikely(&bch2_debug_check_iterators))
  __bch2_btree_iter_verify(trans, iter);
}

static inline void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter)
{
 if (static_branch_unlikely(&bch2_debug_check_iterators))
  __bch2_btree_iter_verify_entry_exit(iter);
}

static inline int bch2_btree_iter_verify_ret(struct btree_trans *trans, struct btree_iter *iter,
          struct bkey_s_c k)
{
 return static_branch_unlikely(&bch2_debug_check_iterators)
  ? __bch2_btree_iter_verify_ret(trans, iter, k)
  : 0;
}

/* Btree path: fixups after btree updates */

static void btree_node_iter_set_set_pos(struct btree_node_iter *iter,
     struct btree *b,
     struct bset_tree *t,
     struct bkey_packed *k)
{
 struct btree_node_iter_set *set;

 btree_node_iter_for_each(iter, set)
  if (set->end == t->end_offset) {
   set->k = __btree_node_key_to_offset(b, k);
   bch2_btree_node_iter_sort(iter, b);
   return;
  }

 bch2_btree_node_iter_push(iter, b, k, btree_bkey_last(b, t));
}

static void __bch2_btree_path_fix_key_modified(struct btree_path *path,
            struct btree *b,
            struct bkey_packed *where)
{
 struct btree_path_level *l = &path->l[b->c.level];

 if (where != bch2_btree_node_iter_peek_all(&l->iter, l->b))
  return;

 if (bkey_iter_pos_cmp(l->b, where, &path->pos) < 0)
  bch2_btree_node_iter_advance(&l->iter, l->b);
}

void bch2_btree_path_fix_key_modified(struct btree_trans *trans,
          struct btree *b,
          struct bkey_packed *where)
{
 struct btree_path *path;
 unsigned i;

 trans_for_each_path_with_node(trans, b, path, i) {
  __bch2_btree_path_fix_key_modified(path, b, where);
  bch2_btree_path_verify_level(trans, path, b->c.level);
 }
}

static void __bch2_btree_node_iter_fix(struct btree_path *path,
           struct btree *b,
           struct btree_node_iter *node_iter,
           struct bset_tree *t,
           struct bkey_packed *where,
           unsigned clobber_u64s,
           unsigned new_u64s)
{
 const struct bkey_packed *end = btree_bkey_last(b, t);
 struct btree_node_iter_set *set;
 unsigned offset = __btree_node_key_to_offset(b, where);
 int shift = new_u64s - clobber_u64s;
 unsigned old_end = t->end_offset - shift;
 unsigned orig_iter_pos = node_iter->data[0].k;
 bool iter_current_key_modified =
  orig_iter_pos >= offset &&
  orig_iter_pos <= offset + clobber_u64s;

 btree_node_iter_for_each(node_iter, set)
  if (set->end == old_end)
   goto found;

 /* didn't find the bset in the iterator - might have to readd it: */
 if (new_u64s &&
     bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
  bch2_btree_node_iter_push(node_iter, b, where, end);
  goto fixup_done;
 } else {
  /* Iterator is after key that changed */
  return;
 }
found:
 set->end = t->end_offset;

 /* Iterator hasn't gotten to the key that changed yet: */
 if (set->k < offset)
  return;

 if (new_u64s &&
     bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
  set->k = offset;
 } else if (set->k < offset + clobber_u64s) {
  set->k = offset + new_u64s;
  if (set->k == set->end)
   bch2_btree_node_iter_set_drop(node_iter, set);
 } else {
  /* Iterator is after key that changed */
  set->k = (int) set->k + shift;
  return;
 }

 bch2_btree_node_iter_sort(node_iter, b);
fixup_done:
 if (node_iter->data[0].k != orig_iter_pos)
  iter_current_key_modified = true;

 /*
 * When a new key is added, and the node iterator now points to that
 * key, the iterator might have skipped past deleted keys that should
 * come after the key the iterator now points to. We have to rewind to
 * before those deleted keys - otherwise
 * bch2_btree_node_iter_prev_all() breaks:
 */
 if (!bch2_btree_node_iter_end(node_iter) &&
     iter_current_key_modified &&
     b->c.level) {
  struct bkey_packed *k, *k2, *p;

  k = bch2_btree_node_iter_peek_all(node_iter, b);

  for_each_bset(b, t) {
   bool set_pos = false;

   if (node_iter->data[0].end == t->end_offset)
    continue;

   k2 = bch2_btree_node_iter_bset_pos(node_iter, b, t);

   while ((p = bch2_bkey_prev_all(b, t, k2)) &&
          bkey_iter_cmp(b, k, p) < 0) {
    k2 = p;
    set_pos = true;
   }

   if (set_pos)
    btree_node_iter_set_set_pos(node_iter,
           b, t, k2);
  }
 }
}

void bch2_btree_node_iter_fix(struct btree_trans *trans,
         struct btree_path *path,
         struct btree *b,
         struct btree_node_iter *node_iter,
         struct bkey_packed *where,
         unsigned clobber_u64s,
         unsigned new_u64s)
{
 struct bset_tree *t = bch2_bkey_to_bset_inlined(b, where);
 struct btree_path *linked;
 unsigned i;

 if (node_iter != &path->l[b->c.level].iter) {
  __bch2_btree_node_iter_fix(path, b, node_iter, t,
        where, clobber_u64s, new_u64s);

  if (static_branch_unlikely(&bch2_debug_check_iterators))
   bch2_btree_node_iter_verify(node_iter, b);
 }

 trans_for_each_path_with_node(trans, b, linked, i) {
  __bch2_btree_node_iter_fix(linked, b,
        &linked->l[b->c.level].iter, t,
        where, clobber_u64s, new_u64s);
  bch2_btree_path_verify_level(trans, linked, b->c.level);
 }
}

/* Btree path level: pointer to a particular btree node and node iter */

static inline struct bkey_s_c __btree_iter_unpack(struct bch_fs *c,
        struct btree_path_level *l,
        struct bkey *u,
        struct bkey_packed *k)
{
 if (unlikely(!k)) {
  /*
 * signal to bch2_btree_iter_peek_slot() that we're currently at
 * a hole
 */
  u->type = KEY_TYPE_deleted;
  return bkey_s_c_null;
 }

 return bkey_disassemble(l->b, k, u);
}

static inline struct bkey_s_c btree_path_level_peek_all(struct bch_fs *c,
       struct btree_path_level *l,
       struct bkey *u)
{
 return __btree_iter_unpack(c, l, u,
   bch2_btree_node_iter_peek_all(&l->iter, l->b));
}

static inline struct bkey_s_c btree_path_level_prev(struct btree_trans *trans,
          struct btree_path *path,
          struct btree_path_level *l,
          struct bkey *u)
{
 struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u,
   bch2_btree_node_iter_prev(&l->iter, l->b));

 path->pos = k.k ? k.k->p : l->b->data->min_key;
 trans->paths_sorted = false;
 bch2_btree_path_verify_level(trans, path, l - path->l);
 return k;
}

static inline bool btree_path_advance_to_pos(struct btree_path *path,
          struct btree_path_level *l,
          int max_advance)
{
 struct bkey_packed *k;
 int nr_advanced = 0;

 while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
        bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
  if (max_advance > 0 && nr_advanced >= max_advance)
   return false;

  bch2_btree_node_iter_advance(&l->iter, l->b);
  nr_advanced++;
 }

 return true;
}

static inline void __btree_path_level_init(struct btree_path *path,
        unsigned level)
{
 struct btree_path_level *l = &path->l[level];

 bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);

 /*
 * Iterators to interior nodes should always be pointed at the first non
 * whiteout:
 */
 if (level)
  bch2_btree_node_iter_peek(&l->iter, l->b);
}

void bch2_btree_path_level_init(struct btree_trans *trans,
    struct btree_path *path,
    struct btree *b)
{
 BUG_ON(path->cached);

 EBUG_ON(!btree_path_pos_in_node(path, b));

 path->l[b->c.level].lock_seq = six_lock_seq(&b->c.lock);
 path->l[b->c.level].b = b;
 __btree_path_level_init(path, b->c.level);
}

/* Btree path: fixups after btree node updates: */

static void bch2_trans_revalidate_updates_in_node(struct btree_trans *trans, struct btree *b)
{
 struct bch_fs *c = trans->c;

 trans_for_each_update(trans, i)
  if (!i->cached &&
      i->level == b->c.level &&
      i->btree_id == b->c.btree_id &&
      bpos_cmp(i->k->k.p, b->data->min_key) >= 0 &&
      bpos_cmp(i->k->k.p, b->data->max_key) <= 0) {
   i->old_v = bch2_btree_path_peek_slot(trans->paths + i->path, &i->old_k).v;

   if (unlikely(trans->journal_replay_not_finished)) {
    struct bkey_i *j_k =
     bch2_journal_keys_peek_slot(c, i->btree_id, i->level,
            i->k->k.p);

    if (j_k) {
     i->old_k = j_k->k;
     i->old_v = &j_k->v;
    }
   }
  }
}

/*
 * A btree node is being replaced - update the iterator to point to the new
 * node:
 */
void bch2_trans_node_add(struct btree_trans *trans,
    struct btree_path *path,
    struct btree *b)
{
 struct btree_path *prev;

 BUG_ON(!btree_path_pos_in_node(path, b));

 while ((prev = prev_btree_path(trans, path)) &&
        btree_path_pos_in_node(prev, b))
  path = prev;

 for (;
      path && btree_path_pos_in_node(path, b);
      path = next_btree_path(trans, path))
  if (path->uptodate == BTREE_ITER_UPTODATE && !path->cached) {
   enum btree_node_locked_type t =
    btree_lock_want(path, b->c.level);

   if (t != BTREE_NODE_UNLOCKED) {
    btree_node_unlock(trans, path, b->c.level);
    six_lock_increment(&b->c.lock, (enum six_lock_type) t);
    mark_btree_node_locked(trans, path, b->c.level, t);
   }

   bch2_btree_path_level_init(trans, path, b);
  }

 bch2_trans_revalidate_updates_in_node(trans, b);
}

void bch2_trans_node_drop(struct btree_trans *trans,
     struct btree *b)
{
 struct btree_path *path;
 unsigned i, level = b->c.level;

 trans_for_each_path(trans, path, i)
  if (path->l[level].b == b) {
   btree_node_unlock(trans, path, level);
   path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
  }
}

/*
 * A btree node has been modified in such a way as to invalidate iterators - fix
 * them:
 */
void bch2_trans_node_reinit_iter(struct btree_trans *trans, struct btree *b)
{
 struct btree_path *path;
 unsigned i;

 trans_for_each_path_with_node(trans, b, path, i)
  __btree_path_level_init(path, b->c.level);

 bch2_trans_revalidate_updates_in_node(trans, b);
}

/* Btree path: traverse, set_pos: */

static inline int btree_path_lock_root(struct btree_trans *trans,
           struct btree_path *path,
           unsigned depth_want,
           unsigned long trace_ip)
{
 struct bch_fs *c = trans->c;
 struct btree_root *r = bch2_btree_id_root(c, path->btree_id);
 enum six_lock_type lock_type;
 unsigned i;
 int ret;

 EBUG_ON(path->nodes_locked);

 while (1) {
  struct btree *b = READ_ONCE(r->b);
  if (unlikely(!b)) {
   BUG_ON(!r->error);
   return r->error;
  }

  path->level = READ_ONCE(b->c.level);

  if (unlikely(path->level < depth_want)) {
   /*
 * the root is at a lower depth than the depth we want:
 * got to the end of the btree, or we're walking nodes
 * greater than some depth and there are no nodes >=
 * that depth
 */
   path->level = depth_want;
   for (i = path->level; i < BTREE_MAX_DEPTH; i++)
    path->l[i].b = NULL;
   return 1;
  }

  lock_type = __btree_lock_want(path, path->level);
  ret = btree_node_lock(trans, path, &b->c,
          path->level, lock_type, trace_ip);
  if (unlikely(ret)) {
   if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
    return ret;
   BUG();
  }

  if (likely(b == READ_ONCE(r->b) &&
      b->c.level == path->level &&
      !race_fault())) {
   for (i = 0; i < path->level; i++)
    path->l[i].b = ERR_PTR(-BCH_ERR_no_btree_node_lock_root);
   path->l[path->level].b = b;
   for (i = path->level + 1; i < BTREE_MAX_DEPTH; i++)
    path->l[i].b = NULL;

   mark_btree_node_locked(trans, path, path->level,
            (enum btree_node_locked_type) lock_type);
   bch2_btree_path_level_init(trans, path, b);
   return 0;
  }

  six_unlock_type(&b->c.lock, lock_type);
 }
}

noinline
static int btree_path_prefetch(struct btree_trans *trans, struct btree_path *path)
{
 struct bch_fs *c = trans->c;
 struct btree_path_level *l = path_l(path);
 struct btree_node_iter node_iter = l->iter;
 struct bkey_packed *k;
 struct bkey_buf tmp;
 unsigned nr = test_bit(BCH_FS_started, &c->flags)
  ? (path->level > 1 ? 0 :  2)
  : (path->level > 1 ? 1 : 16);
 bool was_locked = btree_node_locked(path, path->level);
 int ret = 0;

 bch2_bkey_buf_init(&tmp);

 while (nr-- && !ret) {
  if (!bch2_btree_node_relock(trans, path, path->level))
   break;

  bch2_btree_node_iter_advance(&node_iter, l->b);
  k = bch2_btree_node_iter_peek(&node_iter, l->b);
  if (!k)
   break;

  bch2_bkey_buf_unpack(&tmp, c, l->b, k);
  ret = bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id,
            path->level - 1);
 }

 if (!was_locked)
  btree_node_unlock(trans, path, path->level);

 bch2_bkey_buf_exit(&tmp, c);
 return ret;
}

static int btree_path_prefetch_j(struct btree_trans *trans, struct btree_path *path,
     struct btree_and_journal_iter *jiter)
{
 struct bch_fs *c = trans->c;
 struct bkey_s_c k;
 struct bkey_buf tmp;
 unsigned nr = test_bit(BCH_FS_started, &c->flags)
  ? (path->level > 1 ? 0 :  2)
  : (path->level > 1 ? 1 : 16);
 bool was_locked = btree_node_locked(path, path->level);
 int ret = 0;

 bch2_bkey_buf_init(&tmp);

 jiter->fail_if_too_many_whiteouts = true;

 while (nr-- && !ret) {
  if (!bch2_btree_node_relock(trans, path, path->level))
   break;

  bch2_btree_and_journal_iter_advance(jiter);
  k = bch2_btree_and_journal_iter_peek(jiter);
  if (!k.k)
   break;

  bch2_bkey_buf_reassemble(&tmp, c, k);
  ret = bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id,
            path->level - 1);
 }

 if (!was_locked)
  btree_node_unlock(trans, path, path->level);

 bch2_bkey_buf_exit(&tmp, c);
 return ret;
}

static noinline void btree_node_mem_ptr_set(struct btree_trans *trans,
         struct btree_path *path,
         unsigned plevel, struct btree *b)
{
 struct btree_path_level *l = &path->l[plevel];
 bool locked = btree_node_locked(path, plevel);
 struct bkey_packed *k;
 struct bch_btree_ptr_v2 *bp;

 if (!bch2_btree_node_relock(trans, path, plevel))
  return;

 k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 BUG_ON(k->type != KEY_TYPE_btree_ptr_v2);

 bp = (void *) bkeyp_val(&l->b->format, k);
 bp->mem_ptr = (unsigned long)b;

 if (!locked)
  btree_node_unlock(trans, path, plevel);
}

static noinline int btree_node_iter_and_journal_peek(struct btree_trans *trans,
           struct btree_path *path,
           unsigned flags)
{
 struct bch_fs *c = trans->c;
 struct btree_path_level *l = path_l(path);
 struct btree_and_journal_iter jiter;
 struct bkey_s_c k;
 int ret = 0;

 __bch2_btree_and_journal_iter_init_node_iter(trans, &jiter, l->b, l->iter, path->pos);

 k = bch2_btree_and_journal_iter_peek(&jiter);
 if (!k.k) {
  struct printbuf buf = PRINTBUF;

  prt_str(&buf, "node not found at pos ");
  bch2_bpos_to_text(&buf, path->pos);
  prt_str(&buf, " at btree ");
  bch2_btree_pos_to_text(&buf, c, l->b);

  ret = bch2_fs_topology_error(c, "%s", buf.buf);
  printbuf_exit(&buf);
  goto err;
 }

 bkey_reassemble(&trans->btree_path_down, k);

 if ((flags & BTREE_ITER_prefetch) &&
     c->opts.btree_node_prefetch)
  ret = btree_path_prefetch_j(trans, path, &jiter);

err:
 bch2_btree_and_journal_iter_exit(&jiter);
 return ret;
}

static noinline_for_stack int btree_node_missing_err(struct btree_trans *trans,
           struct btree_path *path)
{
 struct bch_fs *c = trans->c;
 struct printbuf buf = PRINTBUF;

 prt_str(&buf, "node not found at pos ");
 bch2_bpos_to_text(&buf, path->pos);
 prt_str(&buf, " within parent node ");
 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&path_l(path)->b->key));

 bch2_fs_fatal_error(c, "%s", buf.buf);
 printbuf_exit(&buf);
 return bch_err_throw(c, btree_need_topology_repair);
}

static __always_inline int btree_path_down(struct btree_trans *trans,
        struct btree_path *path,
        unsigned flags,
        unsigned long trace_ip)
{
 struct bch_fs *c = trans->c;
 struct btree_path_level *l = path_l(path);
 struct btree *b;
 unsigned level = path->level - 1;
 enum six_lock_type lock_type = __btree_lock_want(path, level);
 int ret;

 EBUG_ON(!btree_node_locked(path, path->level));

 if (unlikely(trans->journal_replay_not_finished)) {
  ret = btree_node_iter_and_journal_peek(trans, path, flags);
  if (ret)
   return ret;
 } else {
  struct bkey_packed *k = bch2_btree_node_iter_peek(&l->iter, l->b);
  if (unlikely(!k))
   return btree_node_missing_err(trans, path);

  bch2_bkey_unpack(l->b, &trans->btree_path_down, k);

  if (unlikely((flags & BTREE_ITER_prefetch)) &&
      c->opts.btree_node_prefetch) {
   ret = btree_path_prefetch(trans, path);
   if (ret)
    return ret;
  }
 }

 b = bch2_btree_node_get(trans, path, &trans->btree_path_down,
    level, lock_type, trace_ip);
 ret = PTR_ERR_OR_ZERO(b);
 if (unlikely(ret))
  return ret;

 if (unlikely(b != btree_node_mem_ptr(&trans->btree_path_down)) &&
     likely(!trans->journal_replay_not_finished &&
     trans->btree_path_down.k.type == KEY_TYPE_btree_ptr_v2))
  btree_node_mem_ptr_set(trans, path, level + 1, b);

 if (btree_node_read_locked(path, level + 1))
  btree_node_unlock(trans, path, level + 1);

 mark_btree_node_locked(trans, path, level,
          (enum btree_node_locked_type) lock_type);
 path->level = level;
 bch2_btree_path_level_init(trans, path, b);

 bch2_btree_path_verify_locks(trans, path);
 return 0;
}

static int bch2_btree_path_traverse_all(struct btree_trans *trans)
{
 struct bch_fs *c = trans->c;
 struct btree_path *path;
 unsigned long trace_ip = _RET_IP_;
 unsigned i;
 int ret = 0;

 if (trans->in_traverse_all)
  return bch_err_throw(c, transaction_restart_in_traverse_all);

 trans->in_traverse_all = true;
retry_all:
 trans->restarted = 0;
 trans->last_restarted_ip = 0;

 trans_for_each_path(trans, path, i)
  path->should_be_locked = false;

 btree_trans_sort_paths(trans);

 bch2_trans_unlock(trans);
 cond_resched();
 trans_set_locked(trans, false);

 if (unlikely(trans->memory_allocation_failure)) {
  struct closure cl;

  closure_init_stack(&cl);

  do {
   ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
   closure_sync(&cl);
  } while (ret);
 }

 /* Now, redo traversals in correct order: */
 i = 0;
 while (i < trans->nr_sorted) {
  btree_path_idx_t idx = trans->sorted[i];

  /*
 * Traversing a path can cause another path to be added at about
 * the same position:
 */
  if (trans->paths[idx].uptodate) {
   __btree_path_get(trans, &trans->paths[idx], false);
   ret = bch2_btree_path_traverse_one(trans, idx, 0, _THIS_IP_);
   __btree_path_put(trans, &trans->paths[idx], false);

   if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
       bch2_err_matches(ret, ENOMEM))
    goto retry_all;
   if (ret)
    goto err;
  } else {
   i++;
  }
 }

 /*
 * We used to assert that all paths had been traversed here
 * (path->uptodate < BTREE_ITER_NEED_TRAVERSE); however, since
 * path->should_be_locked is not set yet, we might have unlocked and
 * then failed to relock a path - that's fine.
 */
err:
 bch2_btree_cache_cannibalize_unlock(trans);

 trans->in_traverse_all = false;

 trace_and_count(c, trans_traverse_all, trans, trace_ip);
 return ret;
}

static inline bool btree_path_check_pos_in_node(struct btree_path *path,
      unsigned l, int check_pos)
{
 if (check_pos < 0 && btree_path_pos_before_node(path, path->l[l].b))
  return false;
 if (check_pos > 0 && btree_path_pos_after_node(path, path->l[l].b))
  return false;
 return true;
}

static inline bool btree_path_good_node(struct btree_trans *trans,
     struct btree_path *path,
     unsigned l, int check_pos)
{
 return is_btree_node(path, l) &&
  bch2_btree_node_relock(trans, path, l) &&
  btree_path_check_pos_in_node(path, l, check_pos);
}

static void btree_path_set_level_down(struct btree_trans *trans,
          struct btree_path *path,
          unsigned new_level)
{
 unsigned l;

 path->level = new_level;

 for (l = path->level + 1; l < BTREE_MAX_DEPTH; l++)
  if (btree_lock_want(path, l) == BTREE_NODE_UNLOCKED)
   btree_node_unlock(trans, path, l);

 btree_path_set_dirty(trans, path, BTREE_ITER_NEED_TRAVERSE);
 bch2_btree_path_verify(trans, path);
}

static noinline unsigned __btree_path_up_until_good_node(struct btree_trans *trans,
        struct btree_path *path,
        int check_pos)
{
 unsigned i, l = path->level;
again:
 while (btree_path_node(path, l) &&
        !btree_path_good_node(trans, path, l, check_pos))
  __btree_path_set_level_up(trans, path, l++);

 /* If we need intent locks, take them too: */
 for (i = l + 1;
      i < path->locks_want && btree_path_node(path, i);
      i++)
  if (!bch2_btree_node_relock(trans, path, i)) {
   while (l <= i)
    __btree_path_set_level_up(trans, path, l++);
   goto again;
  }

 return l;
}

static inline unsigned btree_path_up_until_good_node(struct btree_trans *trans,
           struct btree_path *path,
           int check_pos)
{
 return likely(btree_node_locked(path, path->level) &&
        btree_path_check_pos_in_node(path, path->level, check_pos))
  ? path->level
  : __btree_path_up_until_good_node(trans, path, check_pos);
}

/*
 * This is the main state machine for walking down the btree - walks down to a
 * specified depth
 *
 * Returns 0 on success, -EIO on error (error reading in a btree node).
 *
 * On error, caller (peek_node()/peek_key()) must return NULL; the error is
 * stashed in the iterator and returned from bch2_trans_exit().
 */
int bch2_btree_path_traverse_one(struct btree_trans *trans,
     btree_path_idx_t path_idx,
     unsigned flags,
     unsigned long trace_ip)
{
 struct btree_path *path = &trans->paths[path_idx];
 unsigned depth_want = path->level;
 int ret = -((int) trans->restarted);

 if (unlikely(ret))
  goto out;

 if (unlikely(!trans->srcu_held))
  bch2_trans_srcu_lock(trans);

 trace_btree_path_traverse_start(trans, path);

 /*
 * Ensure we obey path->should_be_locked: if it's set, we can't unlock
 * and re-traverse the path without a transaction restart:
 */
 if (path->should_be_locked) {
  ret = bch2_btree_path_relock(trans, path, trace_ip);
  goto out;
 }

 if (path->cached) {
  ret = bch2_btree_path_traverse_cached(trans, path_idx, flags);
  goto out;
 }

 path = &trans->paths[path_idx];

 if (unlikely(path->level >= BTREE_MAX_DEPTH))
  goto out_uptodate;

 path->level = btree_path_up_until_good_node(trans, path, 0);
 unsigned max_level = path->level;

 EBUG_ON(btree_path_node(path, path->level) &&
  !btree_node_locked(path, path->level));

 /*
 * Note: path->nodes[path->level] may be temporarily NULL here - that
 * would indicate to other code that we got to the end of the btree,
 * here it indicates that relocking the root failed - it's critical that
 * btree_path_lock_root() comes next and that it can't fail
 */
 while (path->level > depth_want) {
  ret = btree_path_node(path, path->level)
   ? btree_path_down(trans, path, flags, trace_ip)
   : btree_path_lock_root(trans, path, depth_want, trace_ip);
  if (unlikely(ret)) {
   if (ret == 1) {
    /*
 * No nodes at this level - got to the end of
 * the btree:
 */
    ret = 0;
    goto out;
   }

   __bch2_btree_path_unlock(trans, path);
   path->level = depth_want;
   path->l[path->level].b = ERR_PTR(ret);
   goto out;
  }
 }

 if (unlikely(max_level > path->level)) {
  struct btree_path *linked;
  unsigned iter;

  trans_for_each_path_with_node(trans, path_l(path)->b, linked, iter)
   for (unsigned j = path->level + 1; j < max_level; j++)
    linked->l[j] = path->l[j];
 }

out_uptodate:
 path->uptodate = BTREE_ITER_UPTODATE;
 trace_btree_path_traverse_end(trans, path);
out:
 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted)
  panic("ret %s (%i) trans->restarted %s (%i)\n",
        bch2_err_str(ret), ret,
        bch2_err_str(trans->restarted), trans->restarted);
 bch2_btree_path_verify(trans, path);
 return ret;
}

static inline void btree_path_copy(struct btree_trans *trans, struct btree_path *dst,
       struct btree_path *src)
{
 unsigned i, offset = offsetof(struct btree_path, pos);

 memcpy((void *) dst + offset,
        (void *) src + offset,
        sizeof(struct btree_path) - offset);

 for (i = 0; i < BTREE_MAX_DEPTH; i++) {
  unsigned t = btree_node_locked_type(dst, i);

  if (t != BTREE_NODE_UNLOCKED)
   six_lock_increment(&dst->l[i].b->c.lock, t);
 }
}

static btree_path_idx_t btree_path_clone(struct btree_trans *trans, btree_path_idx_t src,
      bool intent, unsigned long ip)
{
 btree_path_idx_t new = btree_path_alloc(trans, src);
 btree_path_copy(trans, trans->paths + new, trans->paths + src);
 __btree_path_get(trans, trans->paths + new, intent);
#ifdef TRACK_PATH_ALLOCATED
 trans->paths[new].ip_allocated = ip;
#endif
 return new;
}

__flatten
btree_path_idx_t __bch2_btree_path_make_mut(struct btree_trans *trans,
   btree_path_idx_t path, bool intent, unsigned long ip)
{
 struct btree_path *old = trans->paths + path;
 __btree_path_put(trans, trans->paths + path, intent);
 path = btree_path_clone(trans, path, intent, ip);
 trace_btree_path_clone(trans, old, trans->paths + path);
 trans->paths[path].preserve = false;
 return path;
}

btree_path_idx_t __must_check
__bch2_btree_path_set_pos(struct btree_trans *trans,
     btree_path_idx_t path_idx, struct bpos new_pos,
     bool intent, unsigned long ip)
{
 int cmp = bpos_cmp(new_pos, trans->paths[path_idx].pos);

 bch2_trans_verify_not_unlocked_or_in_restart(trans);
 EBUG_ON(!trans->paths[path_idx].ref);

 trace_btree_path_set_pos(trans, trans->paths + path_idx, &new_pos);

 path_idx = bch2_btree_path_make_mut(trans, path_idx, intent, ip);

 struct btree_path *path = trans->paths + path_idx;
 path->pos  = new_pos;
 trans->paths_sorted = false;

 if (unlikely(path->cached)) {
  btree_node_unlock(trans, path, 0);
  path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_up);
  btree_path_set_dirty(trans, path, BTREE_ITER_NEED_TRAVERSE);
  goto out;
 }

 unsigned level = btree_path_up_until_good_node(trans, path, cmp);

 if (btree_path_node(path, level)) {
  struct btree_path_level *l = &path->l[level];

  BUG_ON(!btree_node_locked(path, level));
  /*
 * We might have to skip over many keys, or just a few: try
 * advancing the node iterator, and if we have to skip over too
 * many keys just reinit it (or if we're rewinding, since that
 * is expensive).
 */
  if (cmp < 0 ||
      !btree_path_advance_to_pos(path, l, 8))
   bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);

  /*
 * Iterators to interior nodes should always be pointed at the first non
 * whiteout:
 */
  if (unlikely(level))
   bch2_btree_node_iter_peek(&l->iter, l->b);
 }

 if (unlikely(level != path->level)) {
  btree_path_set_dirty(trans, path, BTREE_ITER_NEED_TRAVERSE);
  __bch2_btree_path_unlock(trans, path);
 }
out:
 bch2_btree_path_verify(trans, path);
 return path_idx;
}

/* Btree path: main interface: */

static struct btree_path *have_path_at_pos(struct btree_trans *trans, struct btree_path *path)
{
 struct btree_path *sib;

 sib = prev_btree_path(trans, path);
 if (sib && !btree_path_cmp(sib, path))
  return sib;

 sib = next_btree_path(trans, path);
 if (sib && !btree_path_cmp(sib, path))
  return sib;

 return NULL;
}

static struct btree_path *have_node_at_pos(struct btree_trans *trans, struct btree_path *path)
{
 struct btree_path *sib;

 sib = prev_btree_path(trans, path);
 if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
  return sib;

 sib = next_btree_path(trans, path);
 if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
  return sib;

 return NULL;
}

static inline void __bch2_path_free(struct btree_trans *trans, btree_path_idx_t path)
{
 __bch2_btree_path_unlock(trans, trans->paths + path);
 btree_path_list_remove(trans, trans->paths + path);
 __clear_bit(path, trans->paths_allocated);
}

static bool bch2_btree_path_can_relock(struct btree_trans *trans, struct btree_path *path)
{
 unsigned l = path->level;

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

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

  if (path->l[l].lock_seq != path->l[l].b->c.lock.seq)
   return false;

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

 return true;
}

void bch2_path_put(struct btree_trans *trans, btree_path_idx_t path_idx, bool intent)
{
 struct btree_path *path = trans->paths + path_idx, *dup = NULL;

 if (!__btree_path_put(trans, path, intent))
  return;

 if (!path->preserve && !path->should_be_locked)
  goto free;

 dup = path->preserve
  ? have_path_at_pos(trans, path)
  : have_node_at_pos(trans, path);
 if (!dup)
  return;

 /*
 * If we need this path locked, the duplicate also has te be locked
 * before we free this one:
 */
 if (path->should_be_locked &&
     !dup->should_be_locked &&
     !trans->restarted) {
  if (!(trans->locked
        ? bch2_btree_path_relock_norestart(trans, dup)
        : bch2_btree_path_can_relock(trans, dup)))
   return;

  dup->should_be_locked = true;
 }

 BUG_ON(path->should_be_locked &&
        !trans->restarted &&
        trans->locked &&
        !btree_node_locked(dup, dup->level));

 path->should_be_locked = false;
 dup->preserve |= path->preserve;
free:
 trace_btree_path_free(trans, path_idx, dup);
 __bch2_path_free(trans, path_idx);
}

void __noreturn bch2_trans_restart_error(struct btree_trans *trans, u32 restart_count)
{
 panic("trans->restart_count %u, should be %u, last restarted by %pS\n",
       trans->restart_count, restart_count,
       (void *) trans->last_begin_ip);
}

static void __noreturn bch2_trans_in_restart_error(struct btree_trans *trans)
{
#ifdef CONFIG_BCACHEFS_DEBUG
 struct printbuf buf = PRINTBUF;
 bch2_prt_backtrace(&buf, &trans->last_restarted_trace);
 panic("in transaction restart: %s, last restarted by\n%s",
       bch2_err_str(trans->restarted),
       buf.buf);
#else
 panic("in transaction restart: %s, last restarted by %pS\n",
       bch2_err_str(trans->restarted),
       (void *) trans->last_restarted_ip);
#endif
}

void __noreturn bch2_trans_unlocked_or_in_restart_error(struct btree_trans *trans)
{
 if (trans->restarted)
  bch2_trans_in_restart_error(trans);

 if (!trans->locked)
  panic("trans should be locked, unlocked by %pS\n",
        (void *) trans->last_unlock_ip);

 BUG();
}

noinline __cold
void bch2_trans_updates_to_text(struct printbuf *buf, struct btree_trans *trans)
{
 prt_printf(buf, "%u transaction updates for %s journal seq %llu\n",
     trans->nr_updates, trans->fn, trans->journal_res.seq);
 printbuf_indent_add(buf, 2);

 trans_for_each_update(trans, i) {
  struct bkey_s_c old = { &i->old_k, i->old_v };

  prt_str(buf, "update: btree=");
  bch2_btree_id_to_text(buf, i->btree_id);
  prt_printf(buf, " cached=%u %pS\n",
      i->cached,
      (void *) i->ip_allocated);

  prt_printf(buf, "  old ");
  bch2_bkey_val_to_text(buf, trans->c, old);
  prt_newline(buf);

  prt_printf(buf, "  new ");
  bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(i->k));
  prt_newline(buf);
 }

 for (struct jset_entry *e = btree_trans_journal_entries_start(trans);
      e != btree_trans_journal_entries_top(trans);
      e = vstruct_next(e)) {
  bch2_journal_entry_to_text(buf, trans->c, e);
  prt_newline(buf);
 }

 printbuf_indent_sub(buf, 2);
}

static void bch2_btree_path_to_text_short(struct printbuf *out, struct btree_trans *trans, btree_path_idx_t path_idx)
{
 struct btree_path *path = trans->paths + path_idx;

 prt_printf(out, "path: idx %3u ref %u:%u %c %c %c ",
     path_idx, path->ref, path->intent_ref,
     path->preserve ? 'P' : ' ',
     path->should_be_locked ? 'S' : ' ',
     path->cached ? 'C' : 'B');
 bch2_btree_id_level_to_text(out, path->btree_id, path->level);
 prt_str(out, " pos ");
 bch2_bpos_to_text(out, path->pos);

 if (!path->cached && btree_node_locked(path, path->level)) {
  prt_char(out, ' ');
  struct btree *b = path_l(path)->b;
  bch2_bpos_to_text(out, b->data->min_key);
  prt_char(out, '-');
  bch2_bpos_to_text(out, b->key.k.p);
 }

#ifdef TRACK_PATH_ALLOCATED
 prt_printf(out, " %pS", (void *) path->ip_allocated);
#endif
}

static const char *btree_node_locked_str(enum btree_node_locked_type t)
{
 switch (t) {
 case BTREE_NODE_UNLOCKED:
  return "unlocked";
 case BTREE_NODE_READ_LOCKED:
  return "read";
 case BTREE_NODE_INTENT_LOCKED:
  return "intent";
 case BTREE_NODE_WRITE_LOCKED:
  return "write";
 default:
  return NULL;
 }
}

void bch2_btree_path_to_text(struct printbuf *out, struct btree_trans *trans, btree_path_idx_t path_idx)
{
 bch2_btree_path_to_text_short(out, trans, path_idx);

 struct btree_path *path = trans->paths + path_idx;

 prt_printf(out, " uptodate %u locks_want %u", path->uptodate, path->locks_want);
 prt_newline(out);

 printbuf_indent_add(out, 2);
 for (unsigned l = 0; l < BTREE_MAX_DEPTH; l++) {
  prt_printf(out, "l=%u locks %s seq %u node ", l,
      btree_node_locked_str(btree_node_locked_type(path, l)),
      path->l[l].lock_seq);

  int ret = PTR_ERR_OR_ZERO(path->l[l].b);
  if (ret)
   prt_str(out, bch2_err_str(ret));
  else
   prt_printf(out, "%px", path->l[l].b);
  prt_newline(out);
 }
 printbuf_indent_sub(out, 2);
}

static noinline __cold
void __bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans,
    bool nosort)
{
 struct trans_for_each_path_inorder_iter iter;

 if (!nosort)
  btree_trans_sort_paths(trans);

 trans_for_each_path_idx_inorder(trans, iter) {
  bch2_btree_path_to_text_short(out, trans, iter.path_idx);
  prt_newline(out);
 }
}

noinline __cold
void bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans)
{
 __bch2_trans_paths_to_text(out, trans, false);
}

static noinline __cold
void __bch2_dump_trans_paths_updates(struct btree_trans *trans, bool nosort)
{
 struct printbuf buf = PRINTBUF;

 __bch2_trans_paths_to_text(&buf, trans, nosort);
 bch2_trans_updates_to_text(&buf, trans);

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

noinline __cold
void bch2_dump_trans_paths_updates(struct btree_trans *trans)
{
 __bch2_dump_trans_paths_updates(trans, false);
}

noinline __cold
static void bch2_trans_update_max_paths(struct btree_trans *trans)
{
 struct btree_transaction_stats *s = btree_trans_stats(trans);
 struct printbuf buf = PRINTBUF;
 size_t nr = bitmap_weight(trans->paths_allocated, trans->nr_paths);

 bch2_trans_paths_to_text(&buf, trans);

 if (!buf.allocation_failure) {
  mutex_lock(&s->lock);
  if (nr > s->nr_max_paths) {
   s->nr_max_paths = nr;
   swap(s->max_paths_text, buf.buf);
  }
  mutex_unlock(&s->lock);
 }

 printbuf_exit(&buf);

 trans->nr_paths_max = nr;
}

noinline __cold
int __bch2_btree_trans_too_many_iters(struct btree_trans *trans)
{
 if (trace_trans_restart_too_many_iters_enabled()) {
  struct printbuf buf = PRINTBUF;

  bch2_trans_paths_to_text(&buf, trans);
  trace_trans_restart_too_many_iters(trans, _THIS_IP_, buf.buf);
  printbuf_exit(&buf);
 }

 count_event(trans->c, trans_restart_too_many_iters);

 return btree_trans_restart(trans, BCH_ERR_transaction_restart_too_many_iters);
}

static noinline void btree_path_overflow(struct btree_trans *trans)
{
 bch2_dump_trans_paths_updates(trans);
 bch_err(trans->c, "trans path overflow");
}

static noinline void btree_paths_realloc(struct btree_trans *trans)
{
 unsigned nr = trans->nr_paths * 2;

 void *p = kvzalloc(BITS_TO_LONGS(nr) * sizeof(unsigned long) +
     sizeof(struct btree_trans_paths) +
     nr * sizeof(struct btree_path) +
     nr * sizeof(btree_path_idx_t) + 8 +
     nr * sizeof(struct btree_insert_entry), GFP_KERNEL|__GFP_NOFAIL);

 unsigned long *paths_allocated = p;
 memcpy(paths_allocated, trans->paths_allocated, BITS_TO_LONGS(trans->nr_paths) * sizeof(unsigned long));
 p += BITS_TO_LONGS(nr) * sizeof(unsigned long);

 p += sizeof(struct btree_trans_paths);
 struct btree_path *paths = p;
 *trans_paths_nr(paths) = nr;
 memcpy(paths, trans->paths, trans->nr_paths * sizeof(struct btree_path));
 p += nr * sizeof(struct btree_path);

 btree_path_idx_t *sorted = p;
 memcpy(sorted, trans->sorted, trans->nr_sorted * sizeof(btree_path_idx_t));
 p += nr * sizeof(btree_path_idx_t) + 8;

 struct btree_insert_entry *updates = p;
 memcpy(updates, trans->updates, trans->nr_paths * sizeof(struct btree_insert_entry));

 unsigned long *old = trans->paths_allocated;

 rcu_assign_pointer(trans->paths_allocated, paths_allocated);
 rcu_assign_pointer(trans->paths,  paths);
 rcu_assign_pointer(trans->sorted,  sorted);
 rcu_assign_pointer(trans->updates,  updates);

 trans->nr_paths  = nr;

 if (old != trans->_paths_allocated)
  kfree_rcu_mightsleep(old);
}

static inline btree_path_idx_t btree_path_alloc(struct btree_trans *trans,
      btree_path_idx_t pos)
{
 btree_path_idx_t idx = find_first_zero_bit(trans->paths_allocated, trans->nr_paths);

 if (unlikely(idx == trans->nr_paths)) {
  if (trans->nr_paths == BTREE_ITER_MAX) {
   btree_path_overflow(trans);
   return 0;
  }

  btree_paths_realloc(trans);
 }

 /*
 * Do this before marking the new path as allocated, since it won't be
 * initialized yet:
 */
 if (unlikely(idx > trans->nr_paths_max))
  bch2_trans_update_max_paths(trans);

 __set_bit(idx, trans->paths_allocated);

 struct btree_path *path = &trans->paths[idx];
 path->ref  = 0;
 path->intent_ref = 0;
 path->nodes_locked = 0;

 btree_path_list_add(trans, pos, idx);
 trans->paths_sorted = false;
 return idx;
}

btree_path_idx_t bch2_path_get(struct btree_trans *trans,
        enum btree_id btree_id, struct bpos pos,
        unsigned locks_want, unsigned level,
        unsigned flags, unsigned long ip)
{
 struct btree_path *path;
 bool cached = flags & BTREE_ITER_cached;
 bool intent = flags & BTREE_ITER_intent;
 struct trans_for_each_path_inorder_iter iter;
 btree_path_idx_t path_pos = 0, path_idx;

 bch2_trans_verify_not_unlocked_or_in_restart(trans);
 bch2_trans_verify_locks(trans);

 btree_trans_sort_paths(trans);

 if (intent)
  locks_want = max(locks_want, level + 1);
 locks_want = min(locks_want, BTREE_MAX_DEPTH);

 trans_for_each_path_inorder(trans, path, iter) {
  if (__btree_path_cmp(path,
         btree_id,
         cached,
         pos,
         level) > 0)
   break;

  path_pos = iter.path_idx;
 }

 if (path_pos &&
     trans->paths[path_pos].cached == cached &&
     trans->paths[path_pos].btree_id == btree_id &&
     trans->paths[path_pos].level == level &&
     bch2_btree_path_upgrade_norestart(trans, trans->paths + path_pos, locks_want)) {
  trace_btree_path_get(trans, trans->paths + path_pos, &pos);

  __btree_path_get(trans, trans->paths + path_pos, intent);
  path_idx = bch2_btree_path_set_pos(trans, path_pos, pos, intent, ip);
  path = trans->paths + path_idx;
 } else {
  path_idx = btree_path_alloc(trans, path_pos);
  path = trans->paths + path_idx;

  __btree_path_get(trans, path, intent);
  path->pos   = pos;
  path->btree_id   = btree_id;
  path->cached   = cached;
  path->uptodate   = BTREE_ITER_NEED_TRAVERSE;
  path->should_be_locked  = false;
  path->level   = level;
  path->locks_want  = locks_want;
  path->nodes_locked  = 0;
  for (unsigned i = 0; i < ARRAY_SIZE(path->l); i++)
   path->l[i].b  = ERR_PTR(-BCH_ERR_no_btree_node_init);
#ifdef TRACK_PATH_ALLOCATED
  path->ip_allocated  = ip;
#endif
  trans->paths_sorted  = false;

  trace_btree_path_alloc(trans, path);
 }

 if (!(flags & BTREE_ITER_nopreserve))
  path->preserve = true;

 /*
 * If the path has locks_want greater than requested, we don't downgrade
 * it here - on transaction restart because btree node split needs to
 * upgrade locks, we might be putting/getting the iterator again.
 * Downgrading iterators only happens via bch2_trans_downgrade(), after
 * a successful transaction commit.
 */

 return path_idx;
}

btree_path_idx_t bch2_path_get_unlocked_mut(struct btree_trans *trans,
         enum btree_id btree_id,
         unsigned level,
         struct bpos pos)
{
 btree_path_idx_t path_idx = bch2_path_get(trans, btree_id, pos, level + 1, level,
        BTREE_ITER_nopreserve|
        BTREE_ITER_intent, _RET_IP_);
 path_idx = bch2_btree_path_make_mut(trans, path_idx, true, _RET_IP_);

 struct btree_path *path = trans->paths + path_idx;
 bch2_btree_path_downgrade(trans, path);
 __bch2_btree_path_unlock(trans, path);
 return path_idx;
}

struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u)
{

 struct btree_path_level *l = path_l(path);
 struct bkey_packed *_k;
 struct bkey_s_c k;

 if (unlikely(!l->b))
  return bkey_s_c_null;

 EBUG_ON(path->uptodate != BTREE_ITER_UPTODATE);
 EBUG_ON(!btree_node_locked(path, path->level));

 if (!path->cached) {
  _k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
  k = _k ? bkey_disassemble(l->b, _k, u) : bkey_s_c_null;

  EBUG_ON(k.k && bkey_deleted(k.k) && bpos_eq(k.k->p, path->pos));

  if (!k.k || !bpos_eq(path->pos, k.k->p))
   goto hole;
 } else {
  struct bkey_cached *ck = (void *) path->l[0].b;
  if (!ck)
   return bkey_s_c_null;

  EBUG_ON(path->btree_id != ck->key.btree_id ||
   !bkey_eq(path->pos, ck->key.pos));

  *u = ck->k->k;
  k = (struct bkey_s_c) { u, &ck->k->v };
 }

 return k;
hole:
 bkey_init(u);
 u->p = path->pos;
 return (struct bkey_s_c) { u, NULL };
}

void bch2_set_btree_iter_dontneed(struct btree_trans *trans, struct btree_iter *iter)
{
 if (!iter->path || trans->restarted)
  return;

 struct btree_path *path = btree_iter_path(trans, iter);
 path->preserve  = false;
 if (path->ref == 1)
  path->should_be_locked = false;
}
/* Btree iterators: */

int __must_check
__bch2_btree_iter_traverse(struct btree_trans *trans, struct btree_iter *iter)
{
 return bch2_btree_path_traverse(trans, iter->path, iter->flags);
}

int __must_check
bch2_btree_iter_traverse(struct btree_trans *trans, struct btree_iter *iter)
{
 bch2_trans_verify_not_unlocked_or_in_restart(trans);

 iter->path = bch2_btree_path_set_pos(trans, iter->path,
     btree_iter_search_key(iter),
     iter->flags & BTREE_ITER_intent,
     btree_iter_ip_allocated(iter));

 int ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
 if (ret)
  return ret;

 struct btree_path *path = btree_iter_path(trans, iter);
 if (btree_path_node(path, path->level))
  btree_path_set_should_be_locked(trans, path);
 return 0;
}

/* Iterate across nodes (leaf and interior nodes) */

struct btree *bch2_btree_iter_peek_node(struct btree_trans *trans,
     struct btree_iter *iter)
{
 struct btree *b = NULL;
 int ret;

 EBUG_ON(trans->paths[iter->path].cached);
 bch2_btree_iter_verify(trans, iter);

 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
 if (ret)
  goto err;

 struct btree_path *path = btree_iter_path(trans, iter);
 b = btree_path_node(path, path->level);
 if (!b)
  goto out;

 BUG_ON(bpos_lt(b->key.k.p, iter->pos));

 bkey_init(&iter->k);
 iter->k.p = iter->pos = b->key.k.p;

 iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
     iter->flags & BTREE_ITER_intent,
     btree_iter_ip_allocated(iter));
 btree_path_set_should_be_locked(trans, btree_iter_path(trans, iter));
out:
 bch2_btree_iter_verify_entry_exit(iter);
 bch2_btree_iter_verify(trans, iter);

 return b;
err:
 b = ERR_PTR(ret);
 goto out;
}

/* Only kept for -tools */
struct btree *bch2_btree_iter_peek_node_and_restart(struct btree_trans *trans,
          struct btree_iter *iter)
{
 struct btree *b;

 while (b = bch2_btree_iter_peek_node(trans, iter),
        bch2_err_matches(PTR_ERR_OR_ZERO(b), BCH_ERR_transaction_restart))
  bch2_trans_begin(trans);

 return b;
}

struct btree *bch2_btree_iter_next_node(struct btree_trans *trans, struct btree_iter *iter)
{
 struct btree *b = NULL;
 int ret;

 EBUG_ON(trans->paths[iter->path].cached);
 bch2_trans_verify_not_unlocked_or_in_restart(trans);
 bch2_btree_iter_verify(trans, iter);

 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
 if (ret)
  goto err;


 struct btree_path *path = btree_iter_path(trans, iter);

 /* already at end? */
 if (!btree_path_node(path, path->level))
  return NULL;

 /* got to end? */
 if (!btree_path_node(path, path->level + 1)) {
  path->should_be_locked = false;
  btree_path_set_level_up(trans, path);
  return NULL;
 }

 /*
 * We don't correctly handle nodes with extra intent locks here:
 * downgrade so we don't violate locking invariants
 */
 bch2_btree_path_downgrade(trans, path);

 if (!bch2_btree_node_relock(trans, path, path->level + 1)) {
  trace_and_count(trans->c, trans_restart_relock_next_node, trans, _THIS_IP_, path);
  ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
  __bch2_btree_path_unlock(trans, path);
  path->l[path->level].b  = ERR_PTR(-BCH_ERR_no_btree_node_relock);
  path->l[path->level + 1].b = ERR_PTR(-BCH_ERR_no_btree_node_relock);
  btree_path_set_dirty(trans, path, BTREE_ITER_NEED_TRAVERSE);
  goto err;
 }

 b = btree_path_node(path, path->level + 1);

 if (bpos_eq(iter->pos, b->key.k.p)) {
  __btree_path_set_level_up(trans, path, path->level++);
 } else {
  if (btree_lock_want(path, path->level + 1) == BTREE_NODE_UNLOCKED)
   btree_node_unlock(trans, path, path->level + 1);

  /*
 * Haven't gotten to the end of the parent node: go back down to
 * the next child node
 */
  iter->path = bch2_btree_path_set_pos(trans, iter->path,
     bpos_successor(iter->pos),
     iter->flags & BTREE_ITER_intent,
     btree_iter_ip_allocated(iter));

  path = btree_iter_path(trans, iter);
  btree_path_set_level_down(trans, path, iter->min_depth);

  ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
  if (ret)
   goto err;

  path = btree_iter_path(trans, iter);
  b = path->l[path->level].b;
 }

 bkey_init(&iter->k);
 iter->k.p = iter->pos = b->key.k.p;

 iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
     iter->flags & BTREE_ITER_intent,
     btree_iter_ip_allocated(iter));
 btree_path_set_should_be_locked(trans, btree_iter_path(trans, iter));
 EBUG_ON(btree_iter_path(trans, iter)->uptodate);
out:
 bch2_btree_iter_verify_entry_exit(iter);
 bch2_btree_iter_verify(trans, iter);

 return b;
err:
 b = ERR_PTR(ret);
 goto out;
}

/* Iterate across keys (in leaf nodes only) */

inline bool bch2_btree_iter_advance(struct btree_trans *trans, struct btree_iter *iter)
{
 struct bpos pos = iter->k.p;
 bool ret = !(iter->flags & BTREE_ITER_all_snapshots
       ? bpos_eq(pos, SPOS_MAX)
       : bkey_eq(pos, SPOS_MAX));

 if (ret && !(iter->flags & BTREE_ITER_is_extents))
  pos = bkey_successor(iter, pos);
 bch2_btree_iter_set_pos(trans, iter, pos);
 return ret;
}

inline bool bch2_btree_iter_rewind(struct btree_trans *trans, struct btree_iter *iter)
{
 struct bpos pos = bkey_start_pos(&iter->k);
 bool ret = !(iter->flags & BTREE_ITER_all_snapshots
       ? bpos_eq(pos, POS_MIN)
       : bkey_eq(pos, POS_MIN));

 if (ret && !(iter->flags & BTREE_ITER_is_extents))
  pos = bkey_predecessor(iter, pos);
 bch2_btree_iter_set_pos(trans, iter, pos);
 return ret;
}

static noinline
void bch2_btree_trans_peek_prev_updates(struct btree_trans *trans, struct btree_iter *iter,
     struct bpos search_key, struct bkey_s_c *k)
{
 struct bpos end = path_l(btree_iter_path(trans, iter))->b->data->min_key;

 trans_for_each_update(trans, i)
  if (!i->key_cache_already_flushed &&
      i->btree_id == iter->btree_id &&
      bpos_le(i->k->k.p, search_key) &&
      bpos_ge(i->k->k.p, k->k ? k->k->p : end)) {
   iter->k = i->k->k;
   *k = bkey_i_to_s_c(i->k);
  }
}

static noinline
void bch2_btree_trans_peek_updates(struct btree_trans *trans, struct btree_iter *iter,
       struct bpos search_key,
       struct bkey_s_c *k)
{
 struct btree_path *path = btree_iter_path(trans, iter);
 struct bpos end = path_l(path)->b->key.k.p;

 trans_for_each_update(trans, i)
  if (!i->key_cache_already_flushed &&
      i->btree_id == iter->btree_id &&
      bpos_ge(i->k->k.p, search_key) &&
      bpos_le(i->k->k.p, k->k ? k->k->p : end)) {
   iter->k = i->k->k;
   *k = bkey_i_to_s_c(i->k);
  }
}

static noinline
void bch2_btree_trans_peek_slot_updates(struct btree_trans *trans, struct btree_iter *iter,
     struct bkey_s_c *k)
{
 trans_for_each_update(trans, i)
  if (!i->key_cache_already_flushed &&
      i->btree_id == iter->btree_id &&
      bpos_eq(i->k->k.p, iter->pos)) {
   iter->k = i->k->k;
   *k = bkey_i_to_s_c(i->k);
  }
}

static struct bkey_i *bch2_btree_journal_peek(struct btree_trans *trans,
           struct btree_iter *iter,
           struct bpos search_pos,
           struct bpos end_pos)
{
 struct btree_path *path = btree_iter_path(trans, iter);

 return bch2_journal_keys_peek_max(trans->c, iter->btree_id,
        path->level,
        search_pos,
        end_pos,
        &iter->journal_idx);
}

static noinline
struct bkey_s_c btree_trans_peek_slot_journal(struct btree_trans *trans,
           struct btree_iter *iter)
{
 struct btree_path *path = btree_iter_path(trans, iter);
 struct bkey_i *k = bch2_btree_journal_peek(trans, iter, path->pos, path->pos);

 if (k) {
  iter->k = k->k;
  return bkey_i_to_s_c(k);
 } else {
  return bkey_s_c_null;
 }
}

static noinline
void btree_trans_peek_journal(struct btree_trans *trans,
         struct btree_iter *iter,
         struct bpos search_key,
         struct bkey_s_c *k)
{
 struct btree_path *path = btree_iter_path(trans, iter);
 struct bkey_i *next_journal =
  bch2_btree_journal_peek(trans, iter, search_key,
    k->k ? k->k->p : path_l(path)->b->key.k.p);
 if (next_journal) {
  iter->k = next_journal->k;
  *k = bkey_i_to_s_c(next_journal);
 }
}

static struct bkey_i *bch2_btree_journal_peek_prev(struct btree_trans *trans,
           struct btree_iter *iter,
           struct bpos search_key,
           struct bpos end_pos)
{
 struct btree_path *path = btree_iter_path(trans, iter);

 return bch2_journal_keys_peek_prev_min(trans->c, iter->btree_id,
        path->level,
        search_key,
        end_pos,
        &iter->journal_idx);
}

static noinline
void btree_trans_peek_prev_journal(struct btree_trans *trans,
       struct btree_iter *iter,
       struct bpos search_key,
       struct bkey_s_c *k)
{
 struct btree_path *path = btree_iter_path(trans, iter);
 struct bkey_i *next_journal =
  bch2_btree_journal_peek_prev(trans, iter, search_key,
    k->k ? k->k->p : path_l(path)->b->data->min_key);

 if (next_journal) {
  iter->k = next_journal->k;
  *k = bkey_i_to_s_c(next_journal);
 }
}

/*
 * Checks btree key cache for key at iter->pos and returns it if present, or
 * bkey_s_c_null:
 */
static noinline
struct bkey_s_c btree_trans_peek_key_cache(struct btree_trans *trans, struct btree_iter *iter,
        struct bpos pos)
{
 struct bch_fs *c = trans->c;
 struct bkey u;
 struct bkey_s_c k;
 int ret;

 bch2_trans_verify_not_unlocked_or_in_restart(trans);

 if ((iter->flags & BTREE_ITER_key_cache_fill) &&
     bpos_eq(iter->pos, pos))
  return bkey_s_c_null;

 if (!bch2_btree_key_cache_find(c, iter->btree_id, pos))
  return bkey_s_c_null;

 if (!iter->key_cache_path)
  iter->key_cache_path = bch2_path_get(trans, iter->btree_id, pos,
           iter->flags & BTREE_ITER_intent, 0,
           iter->flags|BTREE_ITER_cached|
           BTREE_ITER_cached_nofill,
           _THIS_IP_);

 iter->key_cache_path = bch2_btree_path_set_pos(trans, iter->key_cache_path, pos,
     iter->flags & BTREE_ITER_intent,
     btree_iter_ip_allocated(iter));

 ret =   bch2_btree_path_traverse(trans, iter->key_cache_path,
      iter->flags|BTREE_ITER_cached) ?:
  bch2_btree_path_relock(trans, btree_iter_path(trans, iter), _THIS_IP_);
 if (unlikely(ret))
  return bkey_s_c_err(ret);

 k = bch2_btree_path_peek_slot(trans->paths + iter->key_cache_path, &u);
 if (!k.k)
  return k;

 if ((iter->flags & BTREE_ITER_all_snapshots) &&
     !bpos_eq(pos, k.k->p))
  return bkey_s_c_null;

 iter->k = u;
 k.k = &iter->k;
 btree_path_set_should_be_locked(trans, trans->paths + iter->key_cache_path);
 return k;
}

static struct bkey_s_c __bch2_btree_iter_peek(struct btree_trans *trans, struct btree_iter *iter,
           struct bpos search_key)
{
 struct bkey_s_c k, k2;
 int ret;

 EBUG_ON(btree_iter_path(trans, iter)->cached);
 bch2_btree_iter_verify(trans, iter);

 while (1) {
  iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
     iter->flags & BTREE_ITER_intent,
     btree_iter_ip_allocated(iter));

  ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
  if (unlikely(ret)) {
   /* ensure that iter->k is consistent with iter->pos: */
   bch2_btree_iter_set_pos(trans, iter, iter->pos);
   k = bkey_s_c_err(ret);
   break;
  }

  struct btree_path *path = btree_iter_path(trans, iter);
  struct btree_path_level *l = path_l(path);

  if (unlikely(!l->b)) {
   /* No btree nodes at requested level: */
   bch2_btree_iter_set_pos(trans, iter, SPOS_MAX);
   k = bkey_s_c_null;
   break;
  }

  btree_path_set_should_be_locked(trans, path);

  k = btree_path_level_peek_all(trans->c, l, &iter->k);

  if (unlikely(iter->flags & BTREE_ITER_with_key_cache) &&
      k.k &&
      (k2 = btree_trans_peek_key_cache(trans, iter, k.k->p)).k) {
   k = k2;
   if (bkey_err(k)) {
    bch2_btree_iter_set_pos(trans, iter, iter->pos);
    break;
   }
  }

  if (unlikely(iter->flags & BTREE_ITER_with_journal))
   btree_trans_peek_journal(trans, iter, search_key, &k);

  if (unlikely((iter->flags & BTREE_ITER_with_updates) &&
        trans->nr_updates))
   bch2_btree_trans_peek_updates(trans, iter, search_key, &k);

  if (k.k && bkey_deleted(k.k)) {
   /*
 * If we've got a whiteout, and it's after the search
 * key, advance the search key to the whiteout instead
 * of just after the whiteout - it might be a btree
 * whiteout, with a real key at the same position, since
 * in the btree deleted keys sort before non deleted.
 */
   search_key = !bpos_eq(search_key, k.k->p)
    ? k.k->p
    : bpos_successor(k.k->p);
   continue;
  }

  if (likely(k.k)) {
   break;
  } else if (likely(!bpos_eq(l->b->key.k.p, SPOS_MAX))) {
   /* Advance to next leaf node: */
   search_key = bpos_successor(l->b->key.k.p);
  } else {
   /* End of btree: */
   bch2_btree_iter_set_pos(trans, iter, SPOS_MAX);
   k = bkey_s_c_null;
   break;
  }
 }

 bch2_btree_iter_verify(trans, iter);

 if (trace___btree_iter_peek_enabled()) {
  CLASS(printbuf, buf)();

  int ret = bkey_err(k);
  if (ret)
   prt_str(&buf, bch2_err_str(ret));
  else if (k.k)
   bch2_bkey_val_to_text(&buf, trans->c, k);
  else
   prt_str(&buf, "(null)");
  trace___btree_iter_peek(trans->c, buf.buf);
 }

 return k;
}

/**
 * bch2_btree_iter_peek_max() - returns first key greater than or equal to
 * iterator's current position
 * @trans: btree transaction object
 * @iter: iterator to peek from
 * @end: search limit: returns keys less than or equal to @end
 *
 * Returns: key if found, or an error extractable with bkey_err().
 */
struct bkey_s_c bch2_btree_iter_peek_max(struct btree_trans *trans, struct btree_iter *iter,
      struct bpos end)
{
 struct bpos search_key = btree_iter_search_key(iter);
 struct bkey_s_c k;
 struct bpos iter_pos = iter->pos;
 int ret;

 bch2_trans_verify_not_unlocked_or_in_restart(trans);
 bch2_btree_iter_verify_entry_exit(iter);
 EBUG_ON((iter->flags & BTREE_ITER_filter_snapshots) && bkey_eq(end, POS_MAX));

 ret = trans_maybe_inject_restart(trans, _RET_IP_);
 if (unlikely(ret)) {
  k = bkey_s_c_err(ret);
  goto out_no_locked;
 }

 if (iter->update_path) {
  bch2_path_put(trans, iter->update_path, iter->flags & BTREE_ITER_intent);
  iter->update_path = 0;
 }

 while (1) {
  k = __bch2_btree_iter_peek(trans, iter, search_key);
  if (unlikely(!k.k))
   goto end;
  if (unlikely(bkey_err(k)))
   goto out_no_locked;

  if (iter->flags & BTREE_ITER_filter_snapshots) {
   /*
 * We need to check against @end before FILTER_SNAPSHOTS because
 * if we get to a different inode that requested we might be
 * seeing keys for a different snapshot tree that will all be
 * filtered out.
 *
 * But we can't do the full check here, because bkey_start_pos()
 * isn't monotonically increasing before FILTER_SNAPSHOTS, and
 * that's what we check against in extents mode:
 */
   if (unlikely(!(iter->flags & BTREE_ITER_is_extents)
         ? bkey_gt(k.k->p, end)
         : k.k->p.inode > end.inode))
    goto end;

   if (iter->update_path &&
       !bkey_eq(trans->paths[iter->update_path].pos, k.k->p)) {
    bch2_path_put(trans, iter->update_path,
           iter->flags & BTREE_ITER_intent);
    iter->update_path = 0;
   }

   if ((iter->flags & BTREE_ITER_intent) &&
       !(iter->flags & BTREE_ITER_is_extents) &&
       !iter->update_path) {
    struct bpos pos = k.k->p;

    if (pos.snapshot < iter->snapshot) {
     search_key = bpos_successor(k.k->p);
     continue;
    }

    pos.snapshot = iter->snapshot;

    /*
 * advance, same as on exit for iter->path, but only up
 * to snapshot
 */
    __btree_path_get(trans, trans->paths + iter->path, iter->flags & BTREE_ITER_intent);
    iter->update_path = iter->path;

    iter->update_path = bch2_btree_path_set_pos(trans,
       iter->update_path, pos,
       iter->flags & BTREE_ITER_intent,
       _THIS_IP_);
    ret = bch2_btree_path_traverse(trans, iter->update_path, iter->flags);
    if (unlikely(ret)) {
     k = bkey_s_c_err(ret);
     goto out_no_locked;
    }
   }

   /*
 * We can never have a key in a leaf node at POS_MAX, so
 * we don't have to check these successor() calls:
 */
   if (!bch2_snapshot_is_ancestor(trans->c,
             iter->snapshot,
             k.k->p.snapshot)) {
    search_key = bpos_successor(k.k->p);
    continue;
   }

   if (bkey_whiteout(k.k) &&
       !(iter->flags & BTREE_ITER_key_cache_fill)) {
    search_key = bkey_successor(iter, k.k->p);
    continue;
   }
  }

  /*
 * iter->pos should be mononotically increasing, and always be
 * equal to the key we just returned - except extents can
 * straddle iter->pos:
 */
  if (!(iter->flags & BTREE_ITER_is_extents))
   iter_pos = k.k->p;
  else
   iter_pos = bkey_max(iter->pos, bkey_start_pos(k.k));

  if (unlikely(iter->flags & BTREE_ITER_all_snapshots ? bpos_gt(iter_pos, end) :
        iter->flags & BTREE_ITER_is_extents ? bkey_ge(iter_pos, end) :
           bkey_gt(iter_pos, end)))
   goto end;

  break;
 }

 iter->pos = iter_pos;

 iter->path = bch2_btree_path_set_pos(trans, iter->path, k.k->p,
    iter->flags & BTREE_ITER_intent,
    btree_iter_ip_allocated(iter));

 btree_path_set_should_be_locked(trans, btree_iter_path(trans, iter));
out_no_locked:
 if (iter->update_path) {
  ret = bch2_btree_path_relock(trans, trans->paths + iter->update_path, _THIS_IP_);
  if (unlikely(ret))
   k = bkey_s_c_err(ret);
  else
   btree_path_set_should_be_locked(trans, trans->paths + iter->update_path);
 }

 if (!(iter->flags & BTREE_ITER_all_snapshots))
  iter->pos.snapshot = iter->snapshot;

 ret = bch2_btree_iter_verify_ret(trans, iter, k);
 if (unlikely(ret)) {
  bch2_btree_iter_set_pos(trans, iter, iter->pos);
  k = bkey_s_c_err(ret);
 }

 bch2_btree_iter_verify_entry_exit(iter);

 if (trace_btree_iter_peek_max_enabled()) {
  CLASS(printbuf, buf)();

  int ret = bkey_err(k);
  if (ret)
   prt_str(&buf, bch2_err_str(ret));
  else if (k.k)
   bch2_bkey_val_to_text(&buf, trans->c, k);
  else
   prt_str(&buf, "(null)");
  trace_btree_iter_peek_max(trans->c, buf.buf);
 }

 return k;
end:
 bch2_btree_iter_set_pos(trans, iter, end);
 k = bkey_s_c_null;
 goto out_no_locked;
}

/**
 * bch2_btree_iter_next() - returns first key greater than iterator's current
 * position
 * @trans: btree transaction object
 * @iter: iterator to peek from
 *
 * Returns: key if found, or an error extractable with bkey_err().
 */
struct bkey_s_c bch2_btree_iter_next(struct btree_trans *trans, struct btree_iter *iter)
{
 if (!bch2_btree_iter_advance(trans, iter))
  return bkey_s_c_null;

 return bch2_btree_iter_peek(trans, iter);
}

static struct bkey_s_c __bch2_btree_iter_peek_prev(struct btree_trans *trans, struct btree_iter *iter,
         struct bpos search_key)
{
 struct bkey_s_c k, k2;

 bch2_btree_iter_verify(trans, iter);

 while (1) {
  iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
     iter->flags & BTREE_ITER_intent,
     btree_iter_ip_allocated(iter));

  int ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
  if (unlikely(ret)) {
   /* ensure that iter->k is consistent with iter->pos: */
   bch2_btree_iter_set_pos(trans, iter, iter->pos);
   k = bkey_s_c_err(ret);
   break;
  }

  struct btree_path *path = btree_iter_path(trans, iter);
  struct btree_path_level *l = path_l(path);

  if (unlikely(!l->b)) {
   /* No btree nodes at requested level: */
   bch2_btree_iter_set_pos(trans, iter, SPOS_MAX);
   k = bkey_s_c_null;
   break;
  }

  btree_path_set_should_be_locked(trans, path);

  k = btree_path_level_peek_all(trans->c, l, &iter->k);
  if (!k.k || bpos_gt(k.k->p, search_key)) {
   k = btree_path_level_prev(trans, path, l, &iter->k);

   BUG_ON(k.k && bpos_gt(k.k->p, search_key));
  }

  if (unlikely(iter->flags & BTREE_ITER_with_key_cache) &&
      k.k &&
      (k2 = btree_trans_peek_key_cache(trans, iter, k.k->p)).k) {
   k = k2;
   if (bkey_err(k2)) {
    bch2_btree_iter_set_pos(trans, iter, iter->pos);
    break;
   }
  }

  if (unlikely(iter->flags & BTREE_ITER_with_journal))
   btree_trans_peek_prev_journal(trans, iter, search_key, &k);

  if (unlikely((iter->flags & BTREE_ITER_with_updates) &&
        trans->nr_updates))
   bch2_btree_trans_peek_prev_updates(trans, iter, search_key, &k);

  if (likely(k.k && !bkey_deleted(k.k))) {
   break;
  } else if (k.k) {
   search_key = bpos_predecessor(k.k->p);
  } else if (likely(!bpos_eq(path->l[0].b->data->min_key, POS_MIN))) {
   /* Advance to previous leaf node: */
   search_key = bpos_predecessor(path->l[0].b->data->min_key);
  } else {
   /* Start of btree: */
   bch2_btree_iter_set_pos(trans, iter, POS_MIN);
   k = bkey_s_c_null;
   break;
  }
 }

 bch2_btree_iter_verify(trans, iter);
 return k;
}

/**
 * bch2_btree_iter_peek_prev_min() - returns first key less than or equal to
 * iterator's current position
 * @trans: btree transaction object
 * @iter: iterator to peek from
 * @end: search limit: returns keys greater than or equal to @end
 *
 * Returns: key if found, or an error extractable with bkey_err().
 */
struct bkey_s_c bch2_btree_iter_peek_prev_min(struct btree_trans *trans, struct btree_iter *iter,
           struct bpos end)
{
 if ((iter->flags & (BTREE_ITER_is_extents|BTREE_ITER_filter_snapshots)) &&
    !bkey_eq(iter->pos, POS_MAX) &&
    !((iter->flags & BTREE_ITER_is_extents) &&
      iter->pos.offset == U64_MAX)) {

  /*
 * bkey_start_pos(), for extents, is not monotonically
 * increasing until after filtering for snapshots:
 *
 * Thus, for extents we need to search forward until we find a
 * real visible extents - easiest to just use peek_slot() (which
 * internally uses peek() for extents)
 */
  struct bkey_s_c k = bch2_btree_iter_peek_slot(trans, iter);
  if (bkey_err(k))
   return k;

  if (!bkey_deleted(k.k) &&
      (!(iter->flags & BTREE_ITER_is_extents) ||
       bkey_lt(bkey_start_pos(k.k), iter->pos)))
   return k;
 }

 struct bpos search_key = iter->pos;
 struct bkey_s_c k;
 btree_path_idx_t saved_path = 0;

 bch2_trans_verify_not_unlocked_or_in_restart(trans);
 bch2_btree_iter_verify_entry_exit(iter);
 EBUG_ON((iter->flags & BTREE_ITER_filter_snapshots) && iter->pos.inode != end.inode);

 int ret = trans_maybe_inject_restart(trans, _RET_IP_);
 if (unlikely(ret)) {
  k = bkey_s_c_err(ret);
  goto out_no_locked;
 }

 while (1) {
  k = __bch2_btree_iter_peek_prev(trans, iter, search_key);
  if (unlikely(!k.k))
   goto end;
  if (unlikely(bkey_err(k)))
   goto out_no_locked;

  if (iter->flags & BTREE_ITER_filter_snapshots) {
   struct btree_path *s = saved_path ? trans->paths + saved_path : NULL;
   if (s && bpos_lt(k.k->p, SPOS(s->pos.inode, s->pos.offset, iter->snapshot))) {
    /*
 * If we have a saved candidate, and we're past
 * the last possible snapshot overwrite, return
 * it:
 */
    bch2_path_put(trans, iter->path,
           iter->flags & BTREE_ITER_intent);
    iter->path = saved_path;
    saved_path = 0;
    k = bch2_btree_path_peek_slot(btree_iter_path(trans, iter), &iter->k);
    break;
   }

   /*
 * We need to check against @end before FILTER_SNAPSHOTS because
 * if we get to a different inode that requested we might be
 * seeing keys for a different snapshot tree that will all be
 * filtered out.
 */
   if (unlikely(bkey_lt(k.k->p, end)))
    goto end;

   if (!bch2_snapshot_is_ancestor(trans->c, iter->snapshot, k.k->p.snapshot)) {
    search_key = bpos_predecessor(k.k->p);
    continue;
   }

   if (k.k->p.snapshot != iter->snapshot) {
    /*
 * Have a key visible in iter->snapshot, but
 * might have overwrites: - save it and keep
 * searching. Unless it's a whiteout - then drop
 * our previous saved candidate:
 */
    if (saved_path) {
     bch2_path_put(trans, saved_path,
            iter->flags & BTREE_ITER_intent);
     saved_path = 0;
    }

    if (!bkey_whiteout(k.k)) {
     saved_path = btree_path_clone(trans, iter->path,
        iter->flags & BTREE_ITER_intent,
        _THIS_IP_);
     trace_btree_path_save_pos(trans,
          trans->paths + iter->path,
          trans->paths + saved_path);
    }

    search_key = bpos_predecessor(k.k->p);
    continue;
   }

   if (bkey_whiteout(k.k)) {
    search_key = bkey_predecessor(iter, k.k->p);
    search_key.snapshot = U32_MAX;
    continue;
   }
  }

  EBUG_ON(iter->flags & BTREE_ITER_all_snapshots  ? bpos_gt(k.k->p, iter->pos) :
   iter->flags & BTREE_ITER_is_extents  ? bkey_ge(bkey_start_pos(k.k), iter->pos) :
           bkey_gt(k.k->p, iter->pos));

  if (unlikely(iter->flags & BTREE_ITER_all_snapshots ? bpos_lt(k.k->p, end) :
        iter->flags & BTREE_ITER_is_extents ? bkey_le(k.k->p, end) :
           bkey_lt(k.k->p, end)))
   goto end;

  break;
 }

 /* Extents can straddle iter->pos: */
 iter->pos = bpos_min(iter->pos, k.k->p);;

 if (iter->flags & BTREE_ITER_filter_snapshots)
  iter->pos.snapshot = iter->snapshot;
out_no_locked:
 if (saved_path)
  bch2_path_put(trans, saved_path, iter->flags & BTREE_ITER_intent);

 bch2_btree_iter_verify_entry_exit(iter);
 bch2_btree_iter_verify(trans, iter);

 if (trace_btree_iter_peek_prev_min_enabled()) {
  CLASS(printbuf, buf)();

  int ret = bkey_err(k);
  if (ret)
   prt_str(&buf, bch2_err_str(ret));
  else if (k.k)
   bch2_bkey_val_to_text(&buf, trans->c, k);
  else
   prt_str(&buf, "(null)");
  trace_btree_iter_peek_prev_min(trans->c, buf.buf);
 }
 return k;
end:
 bch2_btree_iter_set_pos(trans, iter, end);
 k = bkey_s_c_null;
 goto out_no_locked;
}

/**
 * bch2_btree_iter_prev() - returns first key less than iterator's current
 * position
 * @trans: btree transaction object
 * @iter: iterator to peek from
 *
 * Returns: key if found, or an error extractable with bkey_err().
 */
struct bkey_s_c bch2_btree_iter_prev(struct btree_trans *trans, struct btree_iter *iter)
{
 if (!bch2_btree_iter_rewind(trans, iter))
  return bkey_s_c_null;

 return bch2_btree_iter_peek_prev(trans, iter);
}

struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_trans *trans, struct btree_iter *iter)
{
 struct bpos search_key;
 struct bkey_s_c k;
 int ret;

 bch2_trans_verify_not_unlocked_or_in_restart(trans);
 bch2_btree_iter_verify(trans, iter);
 bch2_btree_iter_verify_entry_exit(iter);
 EBUG_ON(btree_iter_path(trans, iter)->level && (iter->flags & BTREE_ITER_with_key_cache));

 ret = trans_maybe_inject_restart(trans, _RET_IP_);
 if (unlikely(ret)) {
  k = bkey_s_c_err(ret);
  goto out;
 }

 /* extents can't span inode numbers: */
 if ((iter->flags & BTREE_ITER_is_extents) &&
     unlikely(iter->pos.offset == KEY_OFFSET_MAX)) {
  if (iter->pos.inode == KEY_INODE_MAX) {
   k = bkey_s_c_null;
   goto out2;
  }

  bch2_btree_iter_set_pos(trans, iter, bpos_nosnap_successor(iter->pos));
 }

 search_key = btree_iter_search_key(iter);
 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
     iter->flags & BTREE_ITER_intent,
     btree_iter_ip_allocated(iter));

 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
 if (unlikely(ret)) {
  k = bkey_s_c_err(ret);
  goto out;
 }

 struct btree_path *path = btree_iter_path(trans, iter);
 if (unlikely(!btree_path_node(path, path->level))) {
  k = bkey_s_c_null;
  goto out2;
 }

 btree_path_set_should_be_locked(trans, path);

 if ((iter->flags & BTREE_ITER_cached) ||
     !(iter->flags & (BTREE_ITER_is_extents|BTREE_ITER_filter_snapshots))) {
  k = bkey_s_c_null;

  if (unlikely((iter->flags & BTREE_ITER_with_updates) &&
        trans->nr_updates)) {
   bch2_btree_trans_peek_slot_updates(trans, iter, &k);
   if (k.k)
    goto out;
  }

  if (unlikely(iter->flags & BTREE_ITER_with_journal) &&
      (k = btree_trans_peek_slot_journal(trans, iter)).k)
   goto out;

  if (unlikely(iter->flags & BTREE_ITER_with_key_cache) &&
      (k = btree_trans_peek_key_cache(trans, iter, iter->pos)).k) {
   if (!bkey_err(k))
    iter->k = *k.k;
   /* We're not returning a key from iter->path: */
   goto out;
  }

  k = bch2_btree_path_peek_slot(btree_iter_path(trans, iter), &iter->k);
  if (unlikely(!k.k))
   goto out;

  if (unlikely(k.k->type == KEY_TYPE_whiteout &&
        (iter->flags & BTREE_ITER_filter_snapshots) &&
        !(iter->flags & BTREE_ITER_key_cache_fill)))
   iter->k.type = KEY_TYPE_deleted;
 } else {
  struct bpos next;
  struct bpos end = iter->pos;

  if (iter->flags & BTREE_ITER_is_extents)
   end.offset = U64_MAX;

  EBUG_ON(btree_iter_path(trans, iter)->level);

  if (iter->flags & BTREE_ITER_intent) {
   struct btree_iter iter2;

   bch2_trans_copy_iter(trans, &iter2, iter);
   k = bch2_btree_iter_peek_max(trans, &iter2, end);

   if (k.k && !bkey_err(k)) {
    swap(iter->key_cache_path, iter2.key_cache_path);
    iter->k = iter2.k;
    k.k = &iter->k;
   }
   bch2_trans_iter_exit(trans, &iter2);
  } else {
   struct bpos pos = iter->pos;

   k = bch2_btree_iter_peek_max(trans, iter, end);
   if (unlikely(bkey_err(k)))
    bch2_btree_iter_set_pos(trans, iter, pos);
   else
    iter->pos = pos;
  }

  if (unlikely(bkey_err(k)))
   goto out;

  next = k.k ? bkey_start_pos(k.k) : POS_MAX;

  if (bkey_lt(iter->pos, next)) {
   bkey_init(&iter->k);
   iter->k.p = iter->pos;

   if (iter->flags & BTREE_ITER_is_extents) {
    bch2_key_resize(&iter->k,
      min_t(u64, KEY_SIZE_MAX,
            (next.inode == iter->pos.inode
             ? next.offset
             : KEY_OFFSET_MAX) -
            iter->pos.offset));
    EBUG_ON(!iter->k.size);
   }

   k = (struct bkey_s_c) { &iter->k, NULL };
  }
 }
out:
 bch2_btree_iter_verify_entry_exit(iter);
 bch2_btree_iter_verify(trans, iter);
 ret = bch2_btree_iter_verify_ret(trans, iter, k);
 if (unlikely(ret))
  k = bkey_s_c_err(ret);
out2:
 if (trace_btree_iter_peek_slot_enabled()) {
  CLASS(printbuf, buf)();

  int ret = bkey_err(k);
  if (ret)
   prt_str(&buf, bch2_err_str(ret));
  else if (k.k)
   bch2_bkey_val_to_text(&buf, trans->c, k);
  else
   prt_str(&buf, "(null)");
  trace_btree_iter_peek_slot(trans->c, buf.buf);
 }

 return k;
}

struct bkey_s_c bch2_btree_iter_next_slot(struct btree_trans *trans, struct btree_iter *iter)
{
 if (!bch2_btree_iter_advance(trans, iter))
  return bkey_s_c_null;

 return bch2_btree_iter_peek_slot(trans, iter);
}

struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_trans *trans, struct btree_iter *iter)
{
 if (!bch2_btree_iter_rewind(trans, iter))
  return bkey_s_c_null;

 return bch2_btree_iter_peek_slot(trans, iter);
}

/* Obsolete, but still used by rust wrapper in -tools */
struct bkey_s_c bch2_btree_iter_peek_and_restart_outlined(struct btree_trans *trans, struct btree_iter *iter)
{
 struct bkey_s_c k;

 while (btree_trans_too_many_iters(trans) ||
        (k = bch2_btree_iter_peek_type(trans, iter, iter->flags),
  bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart)))
  bch2_trans_begin(trans);

 return k;
}

/* new transactional stuff: */

#ifdef CONFIG_BCACHEFS_DEBUG
static void btree_trans_verify_sorted_refs(struct btree_trans *trans)
{
 struct btree_path *path;
 unsigned i;

 BUG_ON(trans->nr_sorted != bitmap_weight(trans->paths_allocated, trans->nr_paths) - 1);

 trans_for_each_path(trans, path, i) {
  BUG_ON(path->sorted_idx >= trans->nr_sorted);
  BUG_ON(trans->sorted[path->sorted_idx] != i);
 }

 for (i = 0; i < trans->nr_sorted; i++) {
  unsigned idx = trans->sorted[i];

  BUG_ON(!test_bit(idx, trans->paths_allocated));
  BUG_ON(trans->paths[idx].sorted_idx != i);
 }
}

static void btree_trans_verify_sorted(struct btree_trans *trans)
{
 struct btree_path *path, *prev = NULL;
 struct trans_for_each_path_inorder_iter iter;

 if (!static_branch_unlikely(&bch2_debug_check_iterators))
  return;

 trans_for_each_path_inorder(trans, path, iter) {
  if (prev && btree_path_cmp(prev, path) > 0) {
   __bch2_dump_trans_paths_updates(trans, true);
   panic("trans paths out of order!\n");
  }
  prev = path;
 }
}
#else
static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans) {}
static inline void btree_trans_verify_sorted(struct btree_trans *trans) {}
#endif

void __bch2_btree_trans_sort_paths(struct btree_trans *trans)
{
 int i, l = 0, r = trans->nr_sorted, inc = 1;
 bool swapped;

 btree_trans_verify_sorted_refs(trans);

 if (trans->paths_sorted)
  goto out;

 /*
 * Cocktail shaker sort: this is efficient because iterators will be
 * mostly sorted.
 */
 do {
  swapped = false;

  for (i = inc > 0 ? l : r - 2;
       i + 1 < r && i >= l;
       i += inc) {
   if (btree_path_cmp(trans->paths + trans->sorted[i],
        trans->paths + trans->sorted[i + 1]) > 0) {
    swap(trans->sorted[i], trans->sorted[i + 1]);
    trans->paths[trans->sorted[i]].sorted_idx = i;
    trans->paths[trans->sorted[i + 1]].sorted_idx = i + 1;
    swapped = true;
   }
  }

  if (inc > 0)
   --r;
  else
   l++;
  inc = -inc;
 } while (swapped);

 trans->paths_sorted = true;
out:
 btree_trans_verify_sorted(trans);
}

static inline void btree_path_list_remove(struct btree_trans *trans,
       struct btree_path *path)
{
 EBUG_ON(path->sorted_idx >= trans->nr_sorted);
#ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
 trans->nr_sorted--;
 memmove_u64s_down_small(trans->sorted + path->sorted_idx,
    trans->sorted + path->sorted_idx + 1,
    DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx,
          sizeof(u64) / sizeof(btree_path_idx_t)));
#else
 array_remove_item(trans->sorted, trans->nr_sorted, path->sorted_idx);
#endif
 for (unsigned i = path->sorted_idx; i < trans->nr_sorted; i++)
  trans->paths[trans->sorted[i]].sorted_idx = i;
}

static inline void btree_path_list_add(struct btree_trans *trans,
           btree_path_idx_t pos,
           btree_path_idx_t path_idx)
{
 struct btree_path *path = trans->paths + path_idx;

 path->sorted_idx = pos ? trans->paths[pos].sorted_idx + 1 : trans->nr_sorted;

#ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
 memmove_u64s_up_small(trans->sorted + path->sorted_idx + 1,
         trans->sorted + path->sorted_idx,
         DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx,
        sizeof(u64) / sizeof(btree_path_idx_t)));
 trans->nr_sorted++;
 trans->sorted[path->sorted_idx] = path_idx;
#else
 array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path_idx);
#endif

 for (unsigned i = path->sorted_idx; i < trans->nr_sorted; i++)
  trans->paths[trans->sorted[i]].sorted_idx = i;

 btree_trans_verify_sorted_refs(trans);
}

void bch2_trans_iter_exit(struct btree_trans *trans, struct btree_iter *iter)
{
 if (iter->update_path)
  bch2_path_put(trans, iter->update_path,
         iter->flags & BTREE_ITER_intent);
 if (iter->path)
  bch2_path_put(trans, iter->path,
         iter->flags & BTREE_ITER_intent);
 if (iter->key_cache_path)
  bch2_path_put(trans, iter->key_cache_path,
         iter->flags & BTREE_ITER_intent);
 iter->path  = 0;
 iter->update_path = 0;
 iter->key_cache_path = 0;
}

void bch2_trans_iter_init_outlined(struct btree_trans *trans,
     struct btree_iter *iter,
     enum btree_id btree_id, struct bpos pos,
     unsigned flags)
{
 bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
          bch2_btree_iter_flags(trans, btree_id, 0, flags),
          _RET_IP_);
}

void bch2_trans_node_iter_init(struct btree_trans *trans,
          struct btree_iter *iter,
          enum btree_id btree_id,
          struct bpos pos,
          unsigned locks_want,
          unsigned depth,
          unsigned flags)
{
 flags |= BTREE_ITER_not_extents;
 flags |= BTREE_ITER_snapshot_field;
 flags |= BTREE_ITER_all_snapshots;

 if (!depth && btree_id_cached(trans->c, btree_id))
  flags |= BTREE_ITER_with_key_cache;

 bch2_trans_iter_init_common(trans, iter, btree_id, pos, locks_want, depth,
          bch2_btree_iter_flags(trans, btree_id, depth, flags),
          _RET_IP_);

 iter->min_depth = depth;

 struct btree_path *path = btree_iter_path(trans, iter);
 BUG_ON(path->locks_want  < min(locks_want, BTREE_MAX_DEPTH));
 BUG_ON(path->level != depth);
 BUG_ON(iter->min_depth != depth);
}

void bch2_trans_copy_iter(struct btree_trans *trans,
     struct btree_iter *dst, struct btree_iter *src)
{
 *dst = *src;
#ifdef TRACK_PATH_ALLOCATED
 dst->ip_allocated = _RET_IP_;
#endif
 if (src->path)
  __btree_path_get(trans, trans->paths + src->path, src->flags & BTREE_ITER_intent);
 if (src->update_path)
  __btree_path_get(trans, trans->paths + src->update_path, src->flags & BTREE_ITER_intent);
 dst->key_cache_path = 0;
}

#ifdef CONFIG_BCACHEFS_TRANS_KMALLOC_TRACE
void bch2_trans_kmalloc_trace_to_text(struct printbuf *out,
          darray_trans_kmalloc_trace *trace)
{
 printbuf_tabstops_reset(out);
 printbuf_tabstop_push(out, 60);

 darray_for_each(*trace, i)
  prt_printf(out, "%pS\t%zu\n", (void *) i->ip, i->bytes);
}
#endif

void *__bch2_trans_kmalloc(struct btree_trans *trans, size_t size, unsigned long ip)
{
 struct bch_fs *c = trans->c;
 unsigned new_top = trans->mem_top + size;
 unsigned old_bytes = trans->mem_bytes;
 unsigned new_bytes = roundup_pow_of_two(new_top);
 int ret;
 void *new_mem;
 void *p;

 if (WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX)) {
#ifdef CONFIG_BCACHEFS_TRANS_KMALLOC_TRACE
  struct printbuf buf = PRINTBUF;
  bch2_log_msg_start(c, &buf);
  prt_printf(&buf, "bump allocator exceeded BTREE_TRANS_MEM_MAX (%u)\n",
      BTREE_TRANS_MEM_MAX);

  bch2_trans_kmalloc_trace_to_text(&buf, &trans->trans_kmalloc_trace);
  bch2_print_str(c, KERN_ERR, buf.buf);
  printbuf_exit(&buf);
#endif
 }

 ret = trans_maybe_inject_restart(trans, _RET_IP_);
 if (ret)
  return ERR_PTR(ret);

 struct btree_transaction_stats *s = btree_trans_stats(trans);
 if (new_bytes > s->max_mem) {
  mutex_lock(&s->lock);
#ifdef CONFIG_BCACHEFS_TRANS_KMALLOC_TRACE
  darray_resize(&s->trans_kmalloc_trace, trans->trans_kmalloc_trace.nr);
  s->trans_kmalloc_trace.nr = min(s->trans_kmalloc_trace.size,
      trans->trans_kmalloc_trace.nr);

  memcpy(s->trans_kmalloc_trace.data,
         trans->trans_kmalloc_trace.data,
         sizeof(s->trans_kmalloc_trace.data[0]) *
         s->trans_kmalloc_trace.nr);
#endif
  s->max_mem = new_bytes;
  mutex_unlock(&s->lock);
 }

 if (trans->used_mempool || new_bytes > BTREE_TRANS_MEM_MAX) {
  EBUG_ON(trans->mem_bytes >= new_bytes);
  return ERR_PTR(-BCH_ERR_ENOMEM_trans_kmalloc);
 }

 if (old_bytes) {
  trans->realloc_bytes_required = new_bytes;
  trace_and_count(c, trans_restart_mem_realloced, trans, _RET_IP_, new_bytes);
  return ERR_PTR(btree_trans_restart_ip(trans,
     BCH_ERR_transaction_restart_mem_realloced, _RET_IP_));
 }

 EBUG_ON(trans->mem);

 new_mem = kmalloc(new_bytes, GFP_NOWAIT|__GFP_NOWARN);
 if (unlikely(!new_mem)) {
  bch2_trans_unlock(trans);

  new_mem = kmalloc(new_bytes, GFP_KERNEL);
  if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) {
   new_mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL);
   new_bytes = BTREE_TRANS_MEM_MAX;
   trans->used_mempool = true;
  }

  EBUG_ON(!new_mem);

  trans->mem = new_mem;
  trans->mem_bytes = new_bytes;

  ret = bch2_trans_relock(trans);
  if (ret)
   return ERR_PTR(ret);
 }

 trans->mem = new_mem;
 trans->mem_bytes = new_bytes;

 p = trans->mem + trans->mem_top;
 trans->mem_top += size;
 memset(p, 0, size);
 return p;
}

static inline void check_srcu_held_too_long(struct btree_trans *trans)
{
 WARN(trans->srcu_held && time_after(jiffies, trans->srcu_lock_time + HZ * 10),
      "btree trans held srcu lock (delaying memory reclaim) for %lu seconds",
      (jiffies - trans->srcu_lock_time) / HZ);
}

void bch2_trans_srcu_unlock(struct btree_trans *trans)
{
 if (trans->srcu_held) {
  struct bch_fs *c = trans->c;
  struct btree_path *path;
  unsigned i;

  trans_for_each_path(trans, path, i)
   if (path->cached && !btree_node_locked(path, 0))
    path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_srcu_reset);

  check_srcu_held_too_long(trans);
  srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
  trans->srcu_held = false;
 }
}

static void bch2_trans_srcu_lock(struct btree_trans *trans)
{
 if (!trans->srcu_held) {
  trans->srcu_idx = srcu_read_lock(&trans->c->btree_trans_barrier);
  trans->srcu_lock_time = jiffies;
  trans->srcu_held = true;
 }
}

/**
 * bch2_trans_begin() - reset a transaction after a interrupted attempt
 * @trans: transaction to reset
 *
 * Returns: current restart counter, to be used with trans_was_restarted()
 *
 * While iterating over nodes or updating nodes a attempt to lock a btree node
 * may return BCH_ERR_transaction_restart when the trylock fails. When this
 * occurs bch2_trans_begin() should be called and the transaction retried.
 */
u32 bch2_trans_begin(struct btree_trans *trans)
{
 struct btree_path *path;
 unsigned i;
 u64 now;

 bch2_trans_reset_updates(trans);

 trans->restart_count++;
 trans->mem_top   = 0;

 if (trans->restarted == BCH_ERR_transaction_restart_mem_realloced) {
  EBUG_ON(!trans->mem || !trans->mem_bytes);
  unsigned new_bytes = trans->realloc_bytes_required;
  void *new_mem = krealloc(trans->mem, new_bytes, GFP_NOWAIT|__GFP_NOWARN);
  if (unlikely(!new_mem)) {
   bch2_trans_unlock(trans);
   new_mem = krealloc(trans->mem, new_bytes, GFP_KERNEL);

   EBUG_ON(new_bytes > BTREE_TRANS_MEM_MAX);

   if (!new_mem) {
    new_mem = mempool_alloc(&trans->c->btree_trans_mem_pool, GFP_KERNEL);
    new_bytes = BTREE_TRANS_MEM_MAX;
    trans->used_mempool = true;
    kfree(trans->mem);
   }
                }
  trans->mem = new_mem;
  trans->mem_bytes = new_bytes;
 }

 trans_for_each_path(trans, path, i) {
  path->should_be_locked = false;

  /*
 * If the transaction wasn't restarted, we're presuming to be
 * doing something new: dont keep iterators excpt the ones that
 * are in use - except for the subvolumes btree:
 */
  if (!trans->restarted && path->btree_id != BTREE_ID_subvolumes)
   path->preserve = false;

  /*
 * XXX: we probably shouldn't be doing this if the transaction
 * was restarted, but currently we still overflow transaction
 * iterators if we do that
 */
  if (!path->ref && !path->preserve)
   __bch2_path_free(trans, i);
  else
   path->preserve = false;
 }

 now = local_clock();

 if (!IS_ENABLED(CONFIG_BCACHEFS_NO_LATENCY_ACCT) &&
     time_after64(now, trans->last_begin_time + 10))
  __bch2_time_stats_update(&btree_trans_stats(trans)->duration,
      trans->last_begin_time, now);

 if (!trans->restarted &&
     (need_resched() ||
      time_after64(now, trans->last_begin_time + BTREE_TRANS_MAX_LOCK_HOLD_TIME_NS))) {
  bch2_trans_unlock(trans);
  cond_resched();
  now = local_clock();
 }
 trans->last_begin_time = now;

 if (unlikely(trans->srcu_held &&
       time_after(jiffies, trans->srcu_lock_time + msecs_to_jiffies(10))))
  bch2_trans_srcu_unlock(trans);

 trans->last_begin_ip = _RET_IP_;

#ifdef CONFIG_BCACHEFS_INJECT_TRANSACTION_RESTARTS
 if (trans->restarted) {
  trans->restart_count_this_trans++;
 } else {
  trans->restart_count_this_trans = 0;
 }
#endif

#ifdef CONFIG_BCACHEFS_TRANS_KMALLOC_TRACE
 trans->trans_kmalloc_trace.nr = 0;
#endif

 trans_set_locked(trans, false);

 if (trans->restarted) {
  bch2_btree_path_traverse_all(trans);
  trans->notrace_relock_fail = false;
 }

 bch2_trans_verify_not_unlocked_or_in_restart(trans);
 return trans->restart_count;
}

const char *bch2_btree_transaction_fns[BCH_TRANSACTIONS_NR] = { "(unknown)" };

unsigned bch2_trans_get_fn_idx(const char *fn)
{
 for (unsigned i = 0; i < ARRAY_SIZE(bch2_btree_transaction_fns); i++)
  if (!bch2_btree_transaction_fns[i] ||
      bch2_btree_transaction_fns[i] == fn) {
   bch2_btree_transaction_fns[i] = fn;
   return i;
  }

 pr_warn_once("BCH_TRANSACTIONS_NR not big enough!");
 return 0;
}

struct btree_trans *__bch2_trans_get(struct bch_fs *c, unsigned fn_idx)
 __acquires(&c->btree_trans_barrier)
{
 struct btree_trans *trans;

 if (IS_ENABLED(__KERNEL__)) {
  trans = this_cpu_xchg(c->btree_trans_bufs->trans, NULL);
  if (trans) {
   memset(trans, 0, offsetof(struct btree_trans, list));
   goto got_trans;
  }
 }

 trans = mempool_alloc(&c->btree_trans_pool, GFP_NOFS);
 memset(trans, 0, sizeof(*trans));

 seqmutex_lock(&c->btree_trans_lock);
 if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) {
  struct btree_trans *pos;
  pid_t pid = current->pid;

  trans->locking_wait.task = current;

  list_for_each_entry(pos, &c->btree_trans_list, list) {
   struct task_struct *pos_task = READ_ONCE(pos->locking_wait.task);
   /*
 * We'd much prefer to be stricter here and completely
 * disallow multiple btree_trans in the same thread -
 * but the data move path calls bch2_write when we
 * already have a btree_trans initialized.
 */
   BUG_ON(pos_task &&
          pid == pos_task->pid &&
          pos->locked);
  }
 }

 list_add(&trans->list, &c->btree_trans_list);
 seqmutex_unlock(&c->btree_trans_lock);
got_trans:
 trans->c  = c;
 trans->last_begin_time = local_clock();
 trans->fn_idx  = fn_idx;
 trans->locking_wait.task = current;
 trans->journal_replay_not_finished =
  unlikely(!test_bit(JOURNAL_replay_done, &c->journal.flags)) &&
  atomic_inc_not_zero(&c->journal_keys.ref);
 trans->nr_paths  = ARRAY_SIZE(trans->_paths);
 trans->paths_allocated = trans->_paths_allocated;
 trans->sorted  = trans->_sorted;
 trans->paths  = trans->_paths;
 trans->updates  = trans->_updates;

 *trans_paths_nr(trans->paths) = BTREE_ITER_INITIAL;

 trans->paths_allocated[0] = 1;

 static struct lock_class_key lockdep_key;
 lockdep_init_map(&trans->dep_map, "bcachefs_btree", &lockdep_key, 0);

 if (fn_idx < BCH_TRANSACTIONS_NR) {
  trans->fn = bch2_btree_transaction_fns[fn_idx];

  struct btree_transaction_stats *s = &c->btree_transaction_stats[fn_idx];

  if (s->max_mem) {
   unsigned expected_mem_bytes = roundup_pow_of_two(s->max_mem);

   trans->mem = kmalloc(expected_mem_bytes, GFP_KERNEL);
   if (likely(trans->mem))
    trans->mem_bytes = expected_mem_bytes;
  }

  trans->nr_paths_max = s->nr_max_paths;
 }

 trans->srcu_idx  = srcu_read_lock(&c->btree_trans_barrier);
 trans->srcu_lock_time = jiffies;
 trans->srcu_held = true;
 trans_set_locked(trans, false);

 closure_init_stack_release(&trans->ref);
 return trans;
}

#ifdef CONFIG_BCACHEFS_DEBUG

static bool btree_paths_leaked(struct btree_trans *trans)
{
 struct btree_path *path;
 unsigned i;

 trans_for_each_path(trans, path, i)
  if (path->ref)
   return true;
 return false;
}

static void check_btree_paths_leaked(struct btree_trans *trans)
{
 if (btree_paths_leaked(trans)) {
  struct bch_fs *c = trans->c;
  struct btree_path *path;
  unsigned i;

  struct printbuf buf = PRINTBUF;
  bch2_log_msg_start(c, &buf);

  prt_printf(&buf, "btree paths leaked from %s!\n", trans->fn);
  trans_for_each_path(trans, path, i)
   if (path->ref)
    prt_printf(&buf, "btree %s %pS\n",
        bch2_btree_id_str(path->btree_id),
        (void *) path->ip_allocated);

  bch2_fs_emergency_read_only2(c, &buf);
  bch2_print_str(c, KERN_ERR, buf.buf);
  printbuf_exit(&buf);
 }
}
#else
static inline void check_btree_paths_leaked(struct btree_trans *trans) {}
#endif

void bch2_trans_put(struct btree_trans *trans)
 __releases(&c->btree_trans_barrier)
{
 struct bch_fs *c = trans->c;

 if (trans->restarted)
  bch2_trans_in_restart_error(trans);

 bch2_trans_unlock(trans);

 trans_for_each_update(trans, i)
  __btree_path_put(trans, trans->paths + i->path, true);
 trans->nr_updates = 0;

 check_btree_paths_leaked(trans);

 if (trans->srcu_held) {
  check_srcu_held_too_long(trans);
  srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
 }

 if (unlikely(trans->journal_replay_not_finished))
  bch2_journal_keys_put(c);

 /*
 * trans->ref protects trans->locking_wait.task, btree_paths array; used
 * by cycle detector
 */
 closure_return_sync(&trans->ref);
 trans->locking_wait.task = NULL;

#ifdef CONFIG_BCACHEFS_DEBUG
 darray_exit(&trans->last_restarted_trace);
#endif
#ifdef CONFIG_BCACHEFS_TRANS_KMALLOC_TRACE
 darray_exit(&trans->trans_kmalloc_trace);
#endif

 unsigned long *paths_allocated = trans->paths_allocated;
 trans->paths_allocated = NULL;
 trans->paths  = NULL;

 if (paths_allocated != trans->_paths_allocated)
  kvfree_rcu_mightsleep(paths_allocated);

 if (trans->used_mempool)
  mempool_free(trans->mem, &c->btree_trans_mem_pool);
 else
  kfree(trans->mem);

 /* Userspace doesn't have a real percpu implementation: */
 if (IS_ENABLED(__KERNEL__))
  trans = this_cpu_xchg(c->btree_trans_bufs->trans, trans);

 if (trans) {
  seqmutex_lock(&c->btree_trans_lock);
  list_del(&trans->list);
  seqmutex_unlock(&c->btree_trans_lock);

  mempool_free(trans, &c->btree_trans_pool);
 }
}

bool bch2_current_has_btree_trans(struct bch_fs *c)
{
 seqmutex_lock(&c->btree_trans_lock);
 struct btree_trans *trans;
 bool ret = false;
 list_for_each_entry(trans, &c->btree_trans_list, list)
  if (trans->locking_wait.task == current &&
      trans->locked) {
   ret = true;
   break;
  }
 seqmutex_unlock(&c->btree_trans_lock);
 return ret;
}

static void __maybe_unused
bch2_btree_bkey_cached_common_to_text(struct printbuf *out,
          struct btree_bkey_cached_common *b)
{
 struct six_lock_count c = six_lock_counts(&b->lock);
 pid_t pid;

 scoped_guard(rcu) {
  struct task_struct *owner = READ_ONCE(b->lock.owner);
  pid = owner ? owner->pid : 0;
 }

 prt_printf(out, "\t%px %c ", b, b->cached ? 'c' : 'b');
 bch2_btree_id_to_text(out, b->btree_id);
 prt_printf(out, " l=%u:", b->level);
 bch2_bpos_to_text(out, btree_node_pos(b));

 prt_printf(out, "\t locks %u:%u:%u held by pid %u",
     c.n[0], c.n[1], c.n[2], pid);
}

void bch2_btree_trans_to_text(struct printbuf *out, struct btree_trans *trans)
{
 struct btree_bkey_cached_common *b;
 static char lock_types[] = { 'r', 'i', 'w' };
 struct task_struct *task = READ_ONCE(trans->locking_wait.task);
 unsigned l, idx;

 /* before rcu_read_lock(): */
 bch2_printbuf_make_room(out, 4096);

 if (!out->nr_tabstops) {
  printbuf_tabstop_push(out, 16);
  printbuf_tabstop_push(out, 32);
 }

 prt_printf(out, "%i %s\n", task ? task->pid : 0, trans->fn);

 /* trans->paths is rcu protected vs. freeing */
 guard(rcu)();
 out->atomic++;

 struct btree_path *paths = rcu_dereference(trans->paths);
 if (!paths)
  goto out;

 unsigned long *paths_allocated = trans_paths_allocated(paths);

 trans_for_each_path_idx_from(paths_allocated, *trans_paths_nr(paths), idx, 1) {
  struct btree_path *path = paths + idx;
  if (!path->nodes_locked)
   continue;

  prt_printf(out, "  path %u %c ",
      idx,
      path->cached ? 'c' : 'b');
  bch2_btree_id_to_text(out, path->btree_id);
  prt_printf(out, " l=%u:", path->level);
  bch2_bpos_to_text(out, path->pos);
  prt_newline(out);

  for (l = 0; l < BTREE_MAX_DEPTH; l++) {
   if (btree_node_locked(path, l) &&
       !IS_ERR_OR_NULL(b = (void *) READ_ONCE(path->l[l].b))) {
    prt_printf(out, "    %c l=%u ",
        lock_types[btree_node_locked_type(path, l)], l);
    bch2_btree_bkey_cached_common_to_text(out, b);
    prt_newline(out);
   }
  }
 }

 b = READ_ONCE(trans->locking);
 if (b) {
  prt_printf(out, "  blocked for %lluus on\n",
      div_u64(local_clock() - trans->locking_wait.start_time, 1000));
  prt_printf(out, "    %c", lock_types[trans->locking_wait.lock_want]);
  bch2_btree_bkey_cached_common_to_text(out, b);
  prt_newline(out);
 }
out:
 --out->atomic;
}

void bch2_fs_btree_iter_exit(struct bch_fs *c)
{
 struct btree_transaction_stats *s;
 struct btree_trans *trans;
 int cpu;

 if (c->btree_trans_bufs)
  for_each_possible_cpu(cpu) {
   struct btree_trans *trans =
    per_cpu_ptr(c->btree_trans_bufs, cpu)->trans;

   if (trans) {
    seqmutex_lock(&c->btree_trans_lock);
    list_del(&trans->list);
    seqmutex_unlock(&c->btree_trans_lock);
   }
   kfree(trans);
  }
 free_percpu(c->btree_trans_bufs);

 trans = list_first_entry_or_null(&c->btree_trans_list, struct btree_trans, list);
 if (trans)
  panic("%s leaked btree_trans\n", trans->fn);

 for (s = c->btree_transaction_stats;
      s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
      s++) {
#ifdef CONFIG_BCACHEFS_TRANS_KMALLOC_TRACE
  darray_exit(&s->trans_kmalloc_trace);
#endif
  kfree(s->max_paths_text);
  bch2_time_stats_exit(&s->lock_hold_times);
 }

 if (c->btree_trans_barrier_initialized) {
  synchronize_srcu_expedited(&c->btree_trans_barrier);
  cleanup_srcu_struct(&c->btree_trans_barrier);
 }
 mempool_exit(&c->btree_trans_mem_pool);
 mempool_exit(&c->btree_trans_pool);
}

void bch2_fs_btree_iter_init_early(struct bch_fs *c)
{
 struct btree_transaction_stats *s;

 for (s = c->btree_transaction_stats;
      s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
      s++) {
  bch2_time_stats_init(&s->duration);
  bch2_time_stats_init(&s->lock_hold_times);
  mutex_init(&s->lock);
 }

 INIT_LIST_HEAD(&c->btree_trans_list);
 seqmutex_init(&c->btree_trans_lock);
}

int bch2_fs_btree_iter_init(struct bch_fs *c)
{
 int ret;

 c->btree_trans_bufs = alloc_percpu(struct btree_trans_buf);
 if (!c->btree_trans_bufs)
  return -ENOMEM;

 ret   = mempool_init_kmalloc_pool(&c->btree_trans_pool, 1,
       sizeof(struct btree_trans)) ?:
  mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1,
       BTREE_TRANS_MEM_MAX) ?:
  init_srcu_struct(&c->btree_trans_barrier);
 if (ret)
  return ret;

 /*
 * static annotation (hackily done) for lock ordering of reclaim vs.
 * btree node locks:
 */
#ifdef CONFIG_LOCKDEP
 fs_reclaim_acquire(GFP_KERNEL);
 struct btree_trans *trans = bch2_trans_get(c);
 trans_set_locked(trans, false);
 bch2_trans_put(trans);
 fs_reclaim_release(GFP_KERNEL);
#endif

 c->btree_trans_barrier_initialized = true;
 return 0;

}