Quelle  btree_trans_commit.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0

#include "bcachefs.h"
#include "alloc_foreground.h"
#include "btree_gc.h"
#include "btree_io.h"
#include "btree_iter.h"
#include "btree_journal_iter.h"
#include "btree_key_cache.h"
#include "btree_update_interior.h"
#include "btree_write_buffer.h"
#include "buckets.h"
#include "disk_accounting.h"
#include "enumerated_ref.h"
#include "errcode.h"
#include "error.h"
#include "journal.h"
#include "journal_io.h"
#include "journal_reclaim.h"
#include "replicas.h"
#include "snapshot.h"

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

static const char * const trans_commit_flags_strs[] = {
#define x(n, ...) #n,
 BCH_TRANS_COMMIT_FLAGS()
#undef x
 NULL
};

void bch2_trans_commit_flags_to_text(struct printbuf *out, enum bch_trans_commit_flags flags)
{
 enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;

 prt_printf(out, "watermark=%s", bch2_watermarks[watermark]);

 flags >>= BCH_WATERMARK_BITS;
 if (flags) {
  prt_char(out, ' ');
  bch2_prt_bitflags(out, trans_commit_flags_strs, flags);
 }
}

static void verify_update_old_key(struct btree_trans *trans, struct btree_insert_entry *i)
{
#ifdef CONFIG_BCACHEFS_DEBUG
 struct bch_fs *c = trans->c;
 struct bkey u;
 struct bkey_s_c k = bch2_btree_path_peek_slot_exact(trans->paths + i->path, &u);

 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)
   k = bkey_i_to_s_c(j_k);
 }

 u = *k.k;
 u.needs_whiteout = i->old_k.needs_whiteout;

 BUG_ON(memcmp(&i->old_k, &u, sizeof(struct bkey)));
 BUG_ON(i->old_v != k.v);
#endif
}

static inline struct btree_path_level *insert_l(struct btree_trans *trans, struct btree_insert_entry *i)
{
 return (trans->paths + i->path)->l + i->level;
}

static inline bool same_leaf_as_prev(struct btree_trans *trans,
         struct btree_insert_entry *i)
{
 return i != trans->updates &&
  insert_l(trans, &i[0])->b == insert_l(trans, &i[-1])->b;
}

static inline bool same_leaf_as_next(struct btree_trans *trans,
         struct btree_insert_entry *i)
{
 return i + 1 < trans->updates + trans->nr_updates &&
  insert_l(trans, &i[0])->b == insert_l(trans, &i[1])->b;
}

inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
        struct btree_path *path,
        struct btree *b)
{
 struct bch_fs *c = trans->c;

 if (unlikely(btree_node_just_written(b)) &&
     bch2_btree_post_write_cleanup(c, b))
  bch2_trans_node_reinit_iter(trans, b);

 /*
 * If the last bset has been written, or if it's gotten too big - start
 * a new bset to insert into:
 */
 if (want_new_bset(c, b))
  bch2_btree_init_next(trans, b);
}

static noinline int trans_lock_write_fail(struct btree_trans *trans, struct btree_insert_entry *i)
{
 while (--i >= trans->updates) {
  if (same_leaf_as_prev(trans, i))
   continue;

  bch2_btree_node_unlock_write(trans, trans->paths + i->path, insert_l(trans, i)->b);
 }

 trace_and_count(trans->c, trans_restart_would_deadlock_write, trans);
 return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
}

static inline int bch2_trans_lock_write(struct btree_trans *trans)
{
 EBUG_ON(trans->write_locked);

 trans_for_each_update(trans, i) {
  if (same_leaf_as_prev(trans, i))
   continue;

  if (bch2_btree_node_lock_write(trans, trans->paths + i->path, &insert_l(trans, i)->b->c))
   return trans_lock_write_fail(trans, i);

  if (!i->cached)
   bch2_btree_node_prep_for_write(trans, trans->paths + i->path, insert_l(trans, i)->b);
 }

 trans->write_locked = true;
 return 0;
}

static inline void bch2_trans_unlock_updates_write(struct btree_trans *trans)
{
 if (likely(trans->write_locked)) {
  trans_for_each_update(trans, i)
   if (btree_node_locked_type(trans->paths + i->path, i->level) ==
       BTREE_NODE_WRITE_LOCKED)
    bch2_btree_node_unlock_write_inlined(trans,
      trans->paths + i->path, insert_l(trans, i)->b);
  trans->write_locked = false;
 }
}

/* Inserting into a given leaf node (last stage of insert): */

/* Handle overwrites and do insert, for non extents: */
bool bch2_btree_bset_insert_key(struct btree_trans *trans,
    struct btree_path *path,
    struct btree *b,
    struct btree_node_iter *node_iter,
    struct bkey_i *insert)
{
 struct bkey_packed *k;
 unsigned clobber_u64s = 0, new_u64s = 0;

 EBUG_ON(btree_node_just_written(b));
 EBUG_ON(bset_written(b, btree_bset_last(b)));
 EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
 EBUG_ON(bpos_lt(insert->k.p, b->data->min_key));
 EBUG_ON(bpos_gt(insert->k.p, b->data->max_key));
 EBUG_ON(insert->k.u64s > bch2_btree_keys_u64s_remaining(b));
 EBUG_ON(!b->c.level && !bpos_eq(insert->k.p, path->pos));
 kmsan_check_memory(insert, bkey_bytes(&insert->k));

 k = bch2_btree_node_iter_peek_all(node_iter, b);
 if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
  k = NULL;

 /* @k is the key being overwritten/deleted, if any: */
 EBUG_ON(k && bkey_deleted(k));

 /* Deleting, but not found? nothing to do: */
 if (bkey_deleted(&insert->k) && !k)
  return false;

 if (bkey_deleted(&insert->k)) {
  /* Deleting: */
  btree_account_key_drop(b, k);
  k->type = KEY_TYPE_deleted;

  if (k->needs_whiteout)
   push_whiteout(b, insert->k.p);
  k->needs_whiteout = false;

  if (k >= btree_bset_last(b)->start) {
   clobber_u64s = k->u64s;
   bch2_bset_delete(b, k, clobber_u64s);
   goto fix_iter;
  } else {
   bch2_btree_path_fix_key_modified(trans, b, k);
  }

  return true;
 }

 if (k) {
  /* Overwriting: */
  btree_account_key_drop(b, k);
  k->type = KEY_TYPE_deleted;

  insert->k.needs_whiteout = k->needs_whiteout;
  k->needs_whiteout = false;

  if (k >= btree_bset_last(b)->start) {
   clobber_u64s = k->u64s;
   goto overwrite;
  } else {
   bch2_btree_path_fix_key_modified(trans, b, k);
  }
 }

 k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
overwrite:
 bch2_bset_insert(b, k, insert, clobber_u64s);
 new_u64s = k->u64s;
fix_iter:
 if (clobber_u64s != new_u64s)
  bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
      clobber_u64s, new_u64s);
 return true;
}

static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
          unsigned i, u64 seq)
{
 struct bch_fs *c = container_of(j, struct bch_fs, journal);
 struct btree_write *w = container_of(pin, struct btree_write, journal);
 struct btree *b = container_of(w, struct btree, writes[i]);
 struct btree_trans *trans = bch2_trans_get(c);
 unsigned long old, new;
 unsigned idx = w - b->writes;

 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);

 old = READ_ONCE(b->flags);
 do {
  new = old;

  if (!(old & (1 << BTREE_NODE_dirty)) ||
      !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
      w->journal.seq != seq)
   break;

  new &= ~BTREE_WRITE_TYPE_MASK;
  new |= BTREE_WRITE_journal_reclaim;
  new |= 1 << BTREE_NODE_need_write;
 } while (!try_cmpxchg(&b->flags, &old, new));

 btree_node_write_if_need(trans, b, SIX_LOCK_read);
 six_unlock_read(&b->c.lock);

 bch2_trans_put(trans);
 return 0;
}

int bch2_btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
{
 return __btree_node_flush(j, pin, 0, seq);
}

int bch2_btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
{
 return __btree_node_flush(j, pin, 1, seq);
}

inline void bch2_btree_add_journal_pin(struct bch_fs *c,
           struct btree *b, u64 seq)
{
 struct btree_write *w = btree_current_write(b);

 bch2_journal_pin_add(&c->journal, seq, &w->journal,
        btree_node_write_idx(b) == 0
        ? bch2_btree_node_flush0
        : bch2_btree_node_flush1);
}

/**
 * bch2_btree_insert_key_leaf() - insert a key one key into a leaf node
 * @trans: btree transaction object
 * @path: path pointing to @insert's pos
 * @insert: key to insert
 * @journal_seq: sequence number of journal reservation
 */
inline void bch2_btree_insert_key_leaf(struct btree_trans *trans,
           struct btree_path *path,
           struct bkey_i *insert,
           u64 journal_seq)
{
 struct bch_fs *c = trans->c;
 struct btree *b = path_l(path)->b;
 struct bset_tree *t = bset_tree_last(b);
 struct bset *i = bset(b, t);
 int old_u64s = bset_u64s(t);
 int old_live_u64s = b->nr.live_u64s;
 int live_u64s_added, u64s_added;

 if (unlikely(!bch2_btree_bset_insert_key(trans, path, b,
     &path_l(path)->iter, insert)))
  return;

 i->journal_seq = cpu_to_le64(max(journal_seq, le64_to_cpu(i->journal_seq)));

 bch2_btree_add_journal_pin(c, b, journal_seq);

 if (unlikely(!btree_node_dirty(b))) {
  EBUG_ON(test_bit(BCH_FS_clean_shutdown, &c->flags));
  set_btree_node_dirty_acct(c, b);
 }

 live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 u64s_added = (int) bset_u64s(t) - old_u64s;

 if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
  b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
  b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);

 if (u64s_added > live_u64s_added &&
     bch2_maybe_compact_whiteouts(c, b))
  bch2_trans_node_reinit_iter(trans, b);
}

/* Cached btree updates: */

/* Normal update interface: */

static inline void btree_insert_entry_checks(struct btree_trans *trans,
          struct btree_insert_entry *i)
{
 struct btree_path *path = trans->paths + i->path;

 BUG_ON(!bpos_eq(i->k->k.p, path->pos));
 BUG_ON(i->cached != path->cached);
 BUG_ON(i->level  != path->level);
 BUG_ON(i->btree_id != path->btree_id);
 BUG_ON(i->bkey_type != __btree_node_type(path->level, path->btree_id));
 EBUG_ON(!i->level &&
  btree_type_has_snapshots(i->btree_id) &&
  !(i->flags & BTREE_UPDATE_internal_snapshot_node) &&
  test_bit(JOURNAL_replay_done, &trans->c->journal.flags) &&
  i->k->k.p.snapshot &&
  bch2_snapshot_is_internal_node(trans->c, i->k->k.p.snapshot) > 0);
}

static __always_inline int bch2_trans_journal_res_get(struct btree_trans *trans,
            unsigned flags)
{
 return bch2_journal_res_get(&trans->c->journal, &trans->journal_res,
        trans->journal_u64s, flags, trans);
}

#define JSET_ENTRY_LOG_U64s  4

static noinline void journal_transaction_name(struct btree_trans *trans)
{
 struct bch_fs *c = trans->c;
 struct journal *j = &c->journal;
 struct jset_entry *entry =
  bch2_journal_add_entry(j, &trans->journal_res,
           BCH_JSET_ENTRY_log, 0, 0,
           JSET_ENTRY_LOG_U64s);
 struct jset_entry_log *l =
  container_of(entry, struct jset_entry_log, entry);

 memcpy_and_pad(l->d, JSET_ENTRY_LOG_U64s * sizeof(u64),
         trans->fn, strlen(trans->fn), 0);
}

static inline int btree_key_can_insert(struct btree_trans *trans,
           struct btree *b, unsigned u64s)
{
 if (!bch2_btree_node_insert_fits(b, u64s))
  return bch_err_throw(trans->c, btree_insert_btree_node_full);

 return 0;
}

noinline static int
btree_key_can_insert_cached_slowpath(struct btree_trans *trans, unsigned flags,
         struct btree_path *path, unsigned new_u64s)
{
 struct bkey_cached *ck = (void *) path->l[0].b;
 struct bkey_i *new_k;
 int ret;

 bch2_trans_unlock_updates_write(trans);
 bch2_trans_unlock(trans);

 new_k = kmalloc(new_u64s * sizeof(u64), GFP_KERNEL);
 if (!new_k) {
  struct bch_fs *c = trans->c;
  bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
   bch2_btree_id_str(path->btree_id), new_u64s);
  return bch_err_throw(c, ENOMEM_btree_key_cache_insert);
 }

 ret =   bch2_trans_relock(trans) ?:
  bch2_trans_lock_write(trans);
 if (unlikely(ret)) {
  kfree(new_k);
  return ret;
 }

 memcpy(new_k, ck->k, ck->u64s * sizeof(u64));

 trans_for_each_update(trans, i)
  if (i->old_v == &ck->k->v)
   i->old_v = &new_k->v;

 kfree(ck->k);
 ck->u64s = new_u64s;
 ck->k  = new_k;
 return 0;
}

static int btree_key_can_insert_cached(struct btree_trans *trans, unsigned flags,
           struct btree_path *path, unsigned u64s)
{
 struct bch_fs *c = trans->c;
 struct bkey_cached *ck = (void *) path->l[0].b;
 unsigned new_u64s;
 struct bkey_i *new_k;
 unsigned watermark = flags & BCH_WATERMARK_MASK;

 EBUG_ON(path->level);

 if (watermark < BCH_WATERMARK_reclaim &&
     !test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
     bch2_btree_key_cache_must_wait(c))
  return bch_err_throw(c, btree_insert_need_journal_reclaim);

 /*
 * bch2_varint_decode can read past the end of the buffer by at most 7
 * bytes (it won't be used):
 */
 u64s += 1;

 if (u64s <= ck->u64s)
  return 0;

 new_u64s = roundup_pow_of_two(u64s);
 new_k  = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOWAIT|__GFP_NOWARN);
 if (unlikely(!new_k))
  return btree_key_can_insert_cached_slowpath(trans, flags, path, new_u64s);

 trans_for_each_update(trans, i)
  if (i->old_v == &ck->k->v)
   i->old_v = &new_k->v;

 ck->u64s = new_u64s;
 ck->k  = new_k;
 return 0;
}

/* Triggers: */

static int run_one_mem_trigger(struct btree_trans *trans,
          struct btree_insert_entry *i,
          unsigned flags)
{
 verify_update_old_key(trans, i);

 if (unlikely(flags & BTREE_TRIGGER_norun))
  return 0;

 struct bkey_s_c old = { &i->old_k, i->old_v };
 struct bkey_i *new = i->k;
 const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);

 if (old_ops->trigger == new_ops->trigger)
  return bch2_key_trigger(trans, i->btree_id, i->level,
    old, bkey_i_to_s(new),
    BTREE_TRIGGER_insert|BTREE_TRIGGER_overwrite|flags);
 else
  return bch2_key_trigger_new(trans, i->btree_id, i->level,
    bkey_i_to_s(new), flags) ?:
         bch2_key_trigger_old(trans, i->btree_id, i->level,
    old, flags);
}

static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i)
{
 verify_update_old_key(trans, i);

 if ((i->flags & BTREE_TRIGGER_norun) ||
     !btree_node_type_has_trans_triggers(i->bkey_type))
  return 0;

 /*
 * Transactional triggers create new btree_insert_entries, so we can't
 * pass them a pointer to a btree_insert_entry, that memory is going to
 * move:
 */
 struct bkey old_k = i->old_k;
 struct bkey_s_c old = { &old_k, i->old_v };
 const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
 unsigned flags = i->flags|BTREE_TRIGGER_transactional;

 if (!i->insert_trigger_run &&
     !i->overwrite_trigger_run &&
     old_ops->trigger == new_ops->trigger) {
  i->overwrite_trigger_run = true;
  i->insert_trigger_run = true;
  return bch2_key_trigger(trans, i->btree_id, i->level, old, bkey_i_to_s(i->k),
     BTREE_TRIGGER_insert|
     BTREE_TRIGGER_overwrite|flags) ?: 1;
 } else if (!i->overwrite_trigger_run) {
  i->overwrite_trigger_run = true;
  return bch2_key_trigger_old(trans, i->btree_id, i->level, old, flags) ?: 1;
 } else if (!i->insert_trigger_run) {
  i->insert_trigger_run = true;
  return bch2_key_trigger_new(trans, i->btree_id, i->level, bkey_i_to_s(i->k), flags) ?: 1;
 } else {
  return 0;
 }
}

static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
{
 unsigned sort_id_start = 0;

 while (sort_id_start < trans->nr_updates) {
  unsigned i, sort_id = trans->updates[sort_id_start].sort_order;
  bool trans_trigger_run;

  /*
 * For a given btree, this algorithm runs insert triggers before
 * overwrite triggers: this is so that when extents are being
 * moved (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop
 * references before they are re-added.
 *
 * Running triggers will append more updates to the list of
 * updates as we're walking it:
 */
  do {
   trans_trigger_run = false;

   for (i = sort_id_start;
        i < trans->nr_updates && trans->updates[i].sort_order <= sort_id;
        i++) {
    if (trans->updates[i].sort_order < sort_id) {
     sort_id_start = i;
     continue;
    }

    int ret = run_one_trans_trigger(trans, trans->updates + i);
    if (ret < 0)
     return ret;
    if (ret)
     trans_trigger_run = true;
   }
  } while (trans_trigger_run);

  sort_id_start = i;
 }

#ifdef CONFIG_BCACHEFS_DEBUG
 trans_for_each_update(trans, i)
  BUG_ON(!(i->flags & BTREE_TRIGGER_norun) &&
         btree_node_type_has_trans_triggers(i->bkey_type) &&
         (!i->insert_trigger_run || !i->overwrite_trigger_run));
#endif
 return 0;
}

static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
{
 trans_for_each_update(trans, i)
  if (btree_node_type_has_triggers(i->bkey_type) &&
      gc_visited(trans->c, gc_pos_btree(i->btree_id, i->level, i->k->k.p))) {
   int ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_gc);
   if (ret)
    return ret;
  }

 return 0;
}

static inline int
bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
          struct btree_insert_entry **stopped_at,
          unsigned long trace_ip)
{
 struct bch_fs *c = trans->c;
 struct btree_trans_commit_hook *h;
 unsigned u64s = 0;
 int ret = 0;

 bch2_trans_verify_not_unlocked_or_in_restart(trans);
#if 0
 /* todo: bring back dynamic fault injection */
 if (race_fault()) {
  trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
  return btree_trans_restart(trans, BCH_ERR_transaction_restart_fault_inject);
 }
#endif
 /*
 * Check if the insert will fit in the leaf node with the write lock
 * held, otherwise another thread could write the node changing the
 * amount of space available:
 */

 prefetch(&trans->c->journal.flags);

 trans_for_each_update(trans, i) {
  /* Multiple inserts might go to same leaf: */
  if (!same_leaf_as_prev(trans, i))
   u64s = 0;

  u64s += i->k->k.u64s;
  ret = !i->cached
   ? btree_key_can_insert(trans, insert_l(trans, i)->b, u64s)
   : btree_key_can_insert_cached(trans, flags, trans->paths + i->path, u64s);
  if (ret) {
   *stopped_at = i;
   return ret;
  }

  i->k->k.needs_whiteout = false;
 }

 /*
 * Don't get journal reservation until after we know insert will
 * succeed:
 */
 if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res))) {
  ret = bch2_trans_journal_res_get(trans,
    (flags & BCH_WATERMARK_MASK)|
    JOURNAL_RES_GET_NONBLOCK);
  if (ret)
   return ret;

  if (unlikely(trans->journal_transaction_names))
   journal_transaction_name(trans);
 }

 /*
 * Not allowed to fail after we've gotten our journal reservation - we
 * have to use it:
 */

 if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
     !(flags & BCH_TRANS_COMMIT_no_journal_res)) {
  if (static_branch_unlikely(&bch2_journal_seq_verify))
   trans_for_each_update(trans, i)
    i->k->k.bversion.lo = trans->journal_res.seq;
  else if (static_branch_unlikely(&bch2_inject_invalid_keys))
   trans_for_each_update(trans, i)
    i->k->k.bversion = MAX_VERSION;
 }

 h = trans->hooks;
 while (h) {
  ret = h->fn(trans, h);
  if (ret)
   return ret;
  h = h->next;
 }

 struct bkey_i *accounting;

 percpu_down_read(&c->mark_lock);
 for (accounting = btree_trans_subbuf_base(trans, &trans->accounting);
      accounting != btree_trans_subbuf_top(trans, &trans->accounting);
      accounting = bkey_next(accounting)) {
  ret = bch2_accounting_trans_commit_hook(trans,
     bkey_i_to_accounting(accounting), flags);
  if (ret)
   goto revert_fs_usage;
 }
 percpu_up_read(&c->mark_lock);

 /* XXX: we only want to run this if deltas are nonzero */
 bch2_trans_account_disk_usage_change(trans);

 trans_for_each_update(trans, i)
  if (btree_node_type_has_atomic_triggers(i->bkey_type)) {
   ret = run_one_mem_trigger(trans, i, BTREE_TRIGGER_atomic|i->flags);
   if (ret)
    goto fatal_err;
  }

 if (unlikely(c->gc_pos.phase)) {
  ret = bch2_trans_commit_run_gc_triggers(trans);
  if  (ret)
   goto fatal_err;
 }

 struct bkey_validate_context validate_context = { .from = BKEY_VALIDATE_commit };

 if (!(flags & BCH_TRANS_COMMIT_no_journal_res))
  validate_context.flags = BCH_VALIDATE_write|BCH_VALIDATE_commit;

 for (struct jset_entry *i = btree_trans_journal_entries_start(trans);
      i != btree_trans_journal_entries_top(trans);
      i = vstruct_next(i)) {
  ret = bch2_journal_entry_validate(c, NULL, i,
        bcachefs_metadata_version_current,
        CPU_BIG_ENDIAN, validate_context);
  if (unlikely(ret)) {
   bch2_trans_inconsistent(trans, "invalid journal entry on insert from %s\n",
      trans->fn);
   goto fatal_err;
  }
 }

 trans_for_each_update(trans, i) {
  validate_context.level = i->level;
  validate_context.btree = i->btree_id;

  ret = bch2_bkey_validate(c, bkey_i_to_s_c(i->k), validate_context);
  if (unlikely(ret)){
   bch2_trans_inconsistent(trans, "invalid bkey on insert from %s -> %ps\n",
      trans->fn, (void *) i->ip_allocated);
   goto fatal_err;
  }
  btree_insert_entry_checks(trans, i);
 }

 if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res))) {
  struct journal *j = &c->journal;
  struct jset_entry *entry;

  trans_for_each_update(trans, i) {
   if (i->key_cache_already_flushed)
    continue;

   if (i->flags & BTREE_UPDATE_nojournal)
    continue;

   verify_update_old_key(trans, i);

   if (trans->journal_transaction_names) {
    entry = bch2_journal_add_entry(j, &trans->journal_res,
             BCH_JSET_ENTRY_overwrite,
             i->btree_id, i->level,
             i->old_k.u64s);
    bkey_reassemble((struct bkey_i *) entry->start,
      (struct bkey_s_c) { &i->old_k, i->old_v });
   }

   entry = bch2_journal_add_entry(j, &trans->journal_res,
            BCH_JSET_ENTRY_btree_keys,
            i->btree_id, i->level,
            i->k->k.u64s);
   bkey_copy((struct bkey_i *) entry->start, i->k);
  }

  memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
      btree_trans_journal_entries_start(trans),
      trans->journal_entries.u64s);

  EBUG_ON(trans->journal_res.u64s < trans->journal_entries.u64s);

  trans->journal_res.offset += trans->journal_entries.u64s;
  trans->journal_res.u64s  -= trans->journal_entries.u64s;

  memcpy_u64s_small(bch2_journal_add_entry(j, &trans->journal_res,
      BCH_JSET_ENTRY_write_buffer_keys,
      BTREE_ID_accounting, 0,
      trans->accounting.u64s)->_data,
      btree_trans_subbuf_base(trans, &trans->accounting),
      trans->accounting.u64s);

  if (trans->journal_seq)
   *trans->journal_seq = trans->journal_res.seq;
 }

 trans_for_each_update(trans, i) {
  struct btree_path *path = trans->paths + i->path;

  if (!i->cached)
   bch2_btree_insert_key_leaf(trans, path, i->k, trans->journal_res.seq);
  else if (!i->key_cache_already_flushed)
   bch2_btree_insert_key_cached(trans, flags, i);
  else
   bch2_btree_key_cache_drop(trans, path);
 }

 return 0;
fatal_err:
 bch2_fs_fatal_error(c, "fatal error in transaction commit: %s", bch2_err_str(ret));
 percpu_down_read(&c->mark_lock);
revert_fs_usage:
 for (struct bkey_i *i = btree_trans_subbuf_base(trans, &trans->accounting);
      i != accounting;
      i = bkey_next(i))
  bch2_accounting_trans_commit_revert(trans, bkey_i_to_accounting(i), flags);
 percpu_up_read(&c->mark_lock);
 return ret;
}

static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
{
 /*
 * Accounting keys aren't deduped in the journal: we have to compare
 * each individual update against what's in the btree to see if it has
 * been applied yet, and accounting updates also don't overwrite,
 * they're deltas that accumulate.
 */
 trans_for_each_update(trans, i)
  if (i->k->k.type != KEY_TYPE_accounting)
   bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
}

static int bch2_trans_commit_journal_pin_flush(struct journal *j,
    struct journal_entry_pin *_pin, u64 seq)
{
 return 0;
}

/*
 * Get journal reservation, take write locks, and attempt to do btree update(s):
 */
static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags,
           struct btree_insert_entry **stopped_at,
           unsigned long trace_ip)
{
 struct bch_fs *c = trans->c;
 int ret = 0, u64s_delta = 0;

 for (unsigned idx = 0; idx < trans->nr_updates; idx++) {
  struct btree_insert_entry *i = trans->updates + idx;
  if (i->cached)
   continue;

  u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
  u64s_delta -= i->old_btree_u64s;

  if (!same_leaf_as_next(trans, i)) {
   if (u64s_delta <= 0) {
    ret = bch2_foreground_maybe_merge(trans, i->path,
       i->level, flags);
    if (unlikely(ret))
     return ret;
   }

   u64s_delta = 0;
  }
 }

 ret = bch2_trans_lock_write(trans);
 if (unlikely(ret))
  return ret;

 ret = bch2_trans_commit_write_locked(trans, flags, stopped_at, trace_ip);

 if (!ret && unlikely(trans->journal_replay_not_finished))
  bch2_drop_overwrites_from_journal(trans);

 bch2_trans_unlock_updates_write(trans);

 if (!ret && trans->journal_pin)
  bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
         trans->journal_pin,
         bch2_trans_commit_journal_pin_flush);

 /*
 * Drop journal reservation after dropping write locks, since dropping
 * the journal reservation may kick off a journal write:
 */
 if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res)))
  bch2_journal_res_put(&c->journal, &trans->journal_res);

 return ret;
}

static int journal_reclaim_wait_done(struct bch_fs *c)
{
 int ret = bch2_journal_error(&c->journal) ?:
  bch2_btree_key_cache_wait_done(c);

 if (!ret)
  journal_reclaim_kick(&c->journal);
 return ret;
}

static noinline
int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
       struct btree_insert_entry *i,
       int ret, unsigned long trace_ip)
{
 struct bch_fs *c = trans->c;
 enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;

 if (bch2_err_matches(ret, BCH_ERR_journal_res_blocked)) {
  /*
 * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
 * flag
 */
  if ((flags & BCH_TRANS_COMMIT_journal_reclaim) &&
      watermark < BCH_WATERMARK_reclaim) {
   ret = bch_err_throw(c, journal_reclaim_would_deadlock);
   goto out;
  }

  ret = drop_locks_do(trans,
   bch2_trans_journal_res_get(trans,
     (flags & BCH_WATERMARK_MASK)|
     JOURNAL_RES_GET_CHECK));
  goto out;
 }

 switch (ret) {
 case -BCH_ERR_btree_insert_btree_node_full:
  ret = bch2_btree_split_leaf(trans, i->path, flags);
  if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
   trace_and_count(c, trans_restart_btree_node_split, trans,
     trace_ip, trans->paths + i->path);
  break;
 case -BCH_ERR_btree_insert_need_mark_replicas:
  ret = drop_locks_do(trans,
   bch2_accounting_update_sb(trans));
  break;
 case -BCH_ERR_btree_insert_need_journal_reclaim:
  bch2_trans_unlock(trans);

  trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
  track_event_change(&c->times[BCH_TIME_blocked_key_cache_flush], true);

  wait_event_freezable(c->journal.reclaim_wait,
         (ret = journal_reclaim_wait_done(c)));

  track_event_change(&c->times[BCH_TIME_blocked_key_cache_flush], false);

  if (ret < 0)
   break;

  ret = bch2_trans_relock(trans);
  break;
 default:
  BUG_ON(ret >= 0);
  break;
 }
out:
 BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);

 bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
    (flags & BCH_TRANS_COMMIT_no_enospc), c,
  "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));

 return ret;
}

/*
 * This is for updates done in the early part of fsck - btree_gc - before we've
 * gone RW. we only add the new key to the list of keys for journal replay to
 * do.
 */
static noinline int
do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
{
 struct bch_fs *c = trans->c;

 BUG_ON(current != c->recovery_task);

 trans_for_each_update(trans, i) {
  int ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
  if (ret)
   return ret;
 }

 for (struct jset_entry *i = btree_trans_journal_entries_start(trans);
      i != btree_trans_journal_entries_top(trans);
      i = vstruct_next(i)) {
  if (i->type == BCH_JSET_ENTRY_btree_keys ||
      i->type == BCH_JSET_ENTRY_write_buffer_keys) {
   jset_entry_for_each_key(i, k) {
    int ret = bch2_journal_key_insert(c, i->btree_id, i->level, k);
    if (ret)
     return ret;
   }
  }

  if (i->type == BCH_JSET_ENTRY_btree_root) {
   guard(mutex)(&c->btree_root_lock);

   struct btree_root *r = bch2_btree_id_root(c, i->btree_id);

   bkey_copy(&r->key, i->start);
   r->level = i->level;
   r->alive = true;
  }
 }

 for (struct bkey_i *i = btree_trans_subbuf_base(trans, &trans->accounting);
      i != btree_trans_subbuf_top(trans, &trans->accounting);
      i = bkey_next(i)) {
  int ret = bch2_journal_key_insert(c, BTREE_ID_accounting, 0, i);
  if (ret)
   return ret;
 }

 return 0;
}

int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
{
 struct btree_insert_entry *errored_at = NULL;
 struct bch_fs *c = trans->c;
 unsigned journal_u64s = 0;
 int ret = 0;

 bch2_trans_verify_not_unlocked_or_in_restart(trans);

 ret = trans_maybe_inject_restart(trans, _RET_IP_);
 if (unlikely(ret))
  goto out_reset;

 if (!trans->nr_updates &&
     !trans->journal_entries.u64s &&
     !trans->accounting.u64s)
  goto out_reset;

 ret = bch2_trans_commit_run_triggers(trans);
 if (ret)
  goto out_reset;

 if (!(flags & BCH_TRANS_COMMIT_no_check_rw) &&
     unlikely(!enumerated_ref_tryget(&c->writes, BCH_WRITE_REF_trans))) {
  if (unlikely(!test_bit(BCH_FS_may_go_rw, &c->flags)))
   ret = do_bch2_trans_commit_to_journal_replay(trans);
  else
   ret = bch_err_throw(c, erofs_trans_commit);
  goto out_reset;
 }

 EBUG_ON(test_bit(BCH_FS_clean_shutdown, &c->flags));

 journal_u64s = jset_u64s(trans->accounting.u64s);
 trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
 if (trans->journal_transaction_names)
  journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);

 trans_for_each_update(trans, i) {
  struct btree_path *path = trans->paths + i->path;

  EBUG_ON(!path->should_be_locked);

  ret = bch2_btree_path_upgrade(trans, path, i->level + 1);
  if (unlikely(ret))
   goto out;

  EBUG_ON(!btree_node_intent_locked(path, i->level));

  if (i->key_cache_already_flushed)
   continue;

  if (i->flags & BTREE_UPDATE_nojournal)
   continue;

  /* we're going to journal the key being updated: */
  journal_u64s += jset_u64s(i->k->k.u64s);

  /* and we're also going to log the overwrite: */
  if (trans->journal_transaction_names)
   journal_u64s += jset_u64s(i->old_k.u64s);
 }

 if (trans->extra_disk_res) {
  ret = bch2_disk_reservation_add(c, trans->disk_res,
    trans->extra_disk_res,
    (flags & BCH_TRANS_COMMIT_no_enospc)
    ? BCH_DISK_RESERVATION_NOFAIL : 0);
  if (ret)
   goto err;
 }
retry:
 errored_at = NULL;
 bch2_trans_verify_not_unlocked_or_in_restart(trans);
 if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res)))
  memset(&trans->journal_res, 0, sizeof(trans->journal_res));
 memset(&trans->fs_usage_delta, 0, sizeof(trans->fs_usage_delta));

 trans->journal_u64s = journal_u64s + trans->journal_entries.u64s;

 ret = do_bch2_trans_commit(trans, flags, &errored_at, _RET_IP_);

 /* make sure we didn't drop or screw up locks: */
 bch2_trans_verify_locks(trans);

 if (ret)
  goto err;

 trace_and_count(c, transaction_commit, trans, _RET_IP_);
out:
 if (likely(!(flags & BCH_TRANS_COMMIT_no_check_rw)))
  enumerated_ref_put(&c->writes, BCH_WRITE_REF_trans);
out_reset:
 if (!ret)
  bch2_trans_downgrade(trans);
 bch2_trans_reset_updates(trans);

 return ret;
err:
 ret = bch2_trans_commit_error(trans, flags, errored_at, ret, _RET_IP_);
 if (ret)
  goto out;

 /*
 * We might have done another transaction commit in the error path -
 * i.e. btree write buffer flush - which will have made use of
 * trans->journal_res, but with BCH_TRANS_COMMIT_no_journal_res that is
 * how the journal sequence number to pin is passed in - so we must
 * restart:
 */
 if (flags & BCH_TRANS_COMMIT_no_journal_res) {
  ret = bch_err_throw(c, transaction_restart_nested);
  goto out;
 }

 goto retry;
}