Quelle  btree_write_buffer.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0

#include "bcachefs.h"
#include "bkey_buf.h"
#include "btree_locking.h"
#include "btree_update.h"
#include "btree_update_interior.h"
#include "btree_write_buffer.h"
#include "disk_accounting.h"
#include "enumerated_ref.h"
#include "error.h"
#include "extents.h"
#include "journal.h"
#include "journal_io.h"
#include "journal_reclaim.h"

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

static int bch2_btree_write_buffer_journal_flush(struct journal *,
    struct journal_entry_pin *, u64);

static inline bool __wb_key_ref_cmp(const struct wb_key_ref *l, const struct wb_key_ref *r)
{
 return (cmp_int(l->hi, r->hi) ?:
  cmp_int(l->mi, r->mi) ?:
  cmp_int(l->lo, r->lo)) >= 0;
}

static inline bool wb_key_ref_cmp(const struct wb_key_ref *l, const struct wb_key_ref *r)
{
#ifdef CONFIG_X86_64
 int cmp;

 asm("mov (%[l]), %%rax;"
     "sub (%[r]), %%rax;"
     "mov 8(%[l]), %%rax;"
     "sbb 8(%[r]), %%rax;"
     "mov 16(%[l]), %%rax;"
     "sbb 16(%[r]), %%rax;"
     : "=@ccae" (cmp)
     : [l] "r" (l), [r] "r" (r)
     : "rax", "cc");

 EBUG_ON(cmp != __wb_key_ref_cmp(l, r));
 return cmp;
#else
 return __wb_key_ref_cmp(l, r);
#endif
}

static int wb_key_seq_cmp(const void *_l, const void *_r)
{
 const struct btree_write_buffered_key *l = _l;
 const struct btree_write_buffered_key *r = _r;

 return cmp_int(l->journal_seq, r->journal_seq);
}

/* Compare excluding idx, the low 24 bits: */
static inline bool wb_key_eq(const void *_l, const void *_r)
{
 const struct wb_key_ref *l = _l;
 const struct wb_key_ref *r = _r;

 return !((l->hi ^ r->hi)|
   (l->mi ^ r->mi)|
   ((l->lo >> 24) ^ (r->lo >> 24)));
}

static noinline void wb_sort(struct wb_key_ref *base, size_t num)
{
 size_t n = num, a = num / 2;

 if (!a)  /* num < 2 || size == 0 */
  return;

 for (;;) {
  size_t b, c, d;

  if (a)   /* Building heap: sift down --a */
   --a;
  else if (--n)  /* Sorting: Extract root to --n */
   swap(base[0], base[n]);
  else   /* Sort complete */
   break;

  /*
 * Sift element at "a" down into heap.  This is the
 * "bottom-up" variant, which significantly reduces
 * calls to cmp_func(): we find the sift-down path all
 * the way to the leaves (one compare per level), then
 * backtrack to find where to insert the target element.
 *
 * Because elements tend to sift down close to the leaves,
 * this uses fewer compares than doing two per level
 * on the way down.  (A bit more than half as many on
 * average, 3/4 worst-case.)
 */
  for (b = a; c = 2*b + 1, (d = c + 1) < n;)
   b = wb_key_ref_cmp(base + c, base + d) ? c : d;
  if (d == n)  /* Special case last leaf with no sibling */
   b = c;

  /* Now backtrack from "b" to the correct location for "a" */
  while (b != a && wb_key_ref_cmp(base + a, base + b))
   b = (b - 1) / 2;
  c = b;   /* Where "a" belongs */
  while (b != a) { /* Shift it into place */
   b = (b - 1) / 2;
   swap(base[b], base[c]);
  }
 }
}

static noinline int wb_flush_one_slowpath(struct btree_trans *trans,
       struct btree_iter *iter,
       struct btree_write_buffered_key *wb)
{
 struct btree_path *path = btree_iter_path(trans, iter);

 bch2_btree_node_unlock_write(trans, path, path->l[0].b);

 trans->journal_res.seq = wb->journal_seq;

 return bch2_trans_update(trans, iter, &wb->k,
     BTREE_UPDATE_internal_snapshot_node) ?:
  bch2_trans_commit(trans, NULL, NULL,
      BCH_TRANS_COMMIT_no_enospc|
      BCH_TRANS_COMMIT_no_check_rw|
      BCH_TRANS_COMMIT_no_journal_res|
      BCH_TRANS_COMMIT_journal_reclaim);
}

static inline int wb_flush_one(struct btree_trans *trans, struct btree_iter *iter,
          struct btree_write_buffered_key *wb,
          bool *write_locked,
          bool *accounting_accumulated,
          size_t *fast)
{
 struct btree_path *path;
 int ret;

 EBUG_ON(!wb->journal_seq);
 EBUG_ON(!trans->c->btree_write_buffer.flushing.pin.seq);
 EBUG_ON(trans->c->btree_write_buffer.flushing.pin.seq > wb->journal_seq);

 ret = bch2_btree_iter_traverse(trans, iter);
 if (ret)
  return ret;

 if (!*accounting_accumulated && wb->k.k.type == KEY_TYPE_accounting) {
  struct bkey u;
  struct bkey_s_c k = bch2_btree_path_peek_slot_exact(btree_iter_path(trans, iter), &u);

  if (k.k->type == KEY_TYPE_accounting)
   bch2_accounting_accumulate(bkey_i_to_accounting(&wb->k),
         bkey_s_c_to_accounting(k));
 }
 *accounting_accumulated = true;

 /*
 * We can't clone a path that has write locks: unshare it now, before
 * set_pos and traverse():
 */
 if (btree_iter_path(trans, iter)->ref > 1)
  iter->path = __bch2_btree_path_make_mut(trans, iter->path, true, _THIS_IP_);

 path = btree_iter_path(trans, iter);

 if (!*write_locked) {
  ret = bch2_btree_node_lock_write(trans, path, &path->l[0].b->c);
  if (ret)
   return ret;

  bch2_btree_node_prep_for_write(trans, path, path->l[0].b);
  *write_locked = true;
 }

 if (unlikely(!bch2_btree_node_insert_fits(path->l[0].b, wb->k.k.u64s))) {
  *write_locked = false;
  return wb_flush_one_slowpath(trans, iter, wb);
 }

 EBUG_ON(!bpos_eq(wb->k.k.p, path->pos));

 bch2_btree_insert_key_leaf(trans, path, &wb->k, wb->journal_seq);
 (*fast)++;
 return 0;
}

/*
 * Update a btree with a write buffered key using the journal seq of the
 * original write buffer insert.
 *
 * It is not safe to rejournal the key once it has been inserted into the write
 * buffer because that may break recovery ordering. For example, the key may
 * have already been modified in the active write buffer in a seq that comes
 * before the current transaction. If we were to journal this key again and
 * crash, recovery would process updates in the wrong order.
 */
static int
btree_write_buffered_insert(struct btree_trans *trans,
     struct btree_write_buffered_key *wb)
{
 struct btree_iter iter;
 int ret;

 bch2_trans_iter_init(trans, &iter, wb->btree, bkey_start_pos(&wb->k.k),
        BTREE_ITER_cached|BTREE_ITER_intent);

 trans->journal_res.seq = wb->journal_seq;

 ret   = bch2_btree_iter_traverse(trans, &iter) ?:
  bch2_trans_update(trans, &iter, &wb->k,
      BTREE_UPDATE_internal_snapshot_node);
 bch2_trans_iter_exit(trans, &iter);
 return ret;
}

static void move_keys_from_inc_to_flushing(struct btree_write_buffer *wb)
{
 struct bch_fs *c = container_of(wb, struct bch_fs, btree_write_buffer);
 struct journal *j = &c->journal;

 if (!wb->inc.keys.nr)
  return;

 bch2_journal_pin_add(j, wb->inc.keys.data[0].journal_seq, &wb->flushing.pin,
        bch2_btree_write_buffer_journal_flush);

 darray_resize(&wb->flushing.keys, min_t(size_t, 1U << 20, wb->flushing.keys.nr + wb->inc.keys.nr));
 darray_resize(&wb->sorted, wb->flushing.keys.size);

 if (!wb->flushing.keys.nr && wb->sorted.size >= wb->inc.keys.nr) {
  swap(wb->flushing.keys, wb->inc.keys);
  goto out;
 }

 size_t nr = min(darray_room(wb->flushing.keys),
   wb->sorted.size - wb->flushing.keys.nr);
 nr = min(nr, wb->inc.keys.nr);

 memcpy(&darray_top(wb->flushing.keys),
        wb->inc.keys.data,
        sizeof(wb->inc.keys.data[0]) * nr);

 memmove(wb->inc.keys.data,
  wb->inc.keys.data + nr,
        sizeof(wb->inc.keys.data[0]) * (wb->inc.keys.nr - nr));

 wb->flushing.keys.nr += nr;
 wb->inc.keys.nr  -= nr;
out:
 if (!wb->inc.keys.nr)
  bch2_journal_pin_drop(j, &wb->inc.pin);
 else
  bch2_journal_pin_update(j, wb->inc.keys.data[0].journal_seq, &wb->inc.pin,
     bch2_btree_write_buffer_journal_flush);

 if (j->watermark) {
  spin_lock(&j->lock);
  bch2_journal_set_watermark(j);
  spin_unlock(&j->lock);
 }

 BUG_ON(wb->sorted.size < wb->flushing.keys.nr);
}

int bch2_btree_write_buffer_insert_err(struct bch_fs *c,
           enum btree_id btree, struct bkey_i *k)
{
 struct printbuf buf = PRINTBUF;

 prt_printf(&buf, "attempting to do write buffer update on non wb btree=");
 bch2_btree_id_to_text(&buf, btree);
 prt_str(&buf, "\n");
 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(k));

 bch2_fs_inconsistent(c, "%s", buf.buf);
 printbuf_exit(&buf);
 return -EROFS;
}

static int bch2_btree_write_buffer_flush_locked(struct btree_trans *trans)
{
 struct bch_fs *c = trans->c;
 struct journal *j = &c->journal;
 struct btree_write_buffer *wb = &c->btree_write_buffer;
 struct btree_iter iter = {};
 size_t overwritten = 0, fast = 0, slowpath = 0, could_not_insert = 0;
 bool write_locked = false;
 bool accounting_replay_done = test_bit(BCH_FS_accounting_replay_done, &c->flags);
 int ret = 0;

 ret = bch2_journal_error(&c->journal);
 if (ret)
  return ret;

 bch2_trans_unlock(trans);
 bch2_trans_begin(trans);

 mutex_lock(&wb->inc.lock);
 move_keys_from_inc_to_flushing(wb);
 mutex_unlock(&wb->inc.lock);

 for (size_t i = 0; i < wb->flushing.keys.nr; i++) {
  wb->sorted.data[i].idx = i;
  wb->sorted.data[i].btree = wb->flushing.keys.data[i].btree;
  memcpy(&wb->sorted.data[i].pos, &wb->flushing.keys.data[i].k.k.p, sizeof(struct bpos));
 }
 wb->sorted.nr = wb->flushing.keys.nr;

 /*
 * We first sort so that we can detect and skip redundant updates, and
 * then we attempt to flush in sorted btree order, as this is most
 * efficient.
 *
 * However, since we're not flushing in the order they appear in the
 * journal we won't be able to drop our journal pin until everything is
 * flushed - which means this could deadlock the journal if we weren't
 * passing BCH_TRANS_COMMIT_journal_reclaim. This causes the update to fail
 * if it would block taking a journal reservation.
 *
 * If that happens, simply skip the key so we can optimistically insert
 * as many keys as possible in the fast path.
 */
 wb_sort(wb->sorted.data, wb->sorted.nr);

 darray_for_each(wb->sorted, i) {
  struct btree_write_buffered_key *k = &wb->flushing.keys.data[i->idx];

  if (unlikely(!btree_type_uses_write_buffer(k->btree))) {
   ret = bch2_btree_write_buffer_insert_err(trans->c, k->btree, &k->k);
   goto err;
  }

  for (struct wb_key_ref *n = i + 1; n < min(i + 4, &darray_top(wb->sorted)); n++)
   prefetch(&wb->flushing.keys.data[n->idx]);

  BUG_ON(!k->journal_seq);

  if (!accounting_replay_done &&
      k->k.k.type == KEY_TYPE_accounting) {
   slowpath++;
   continue;
  }

  if (i + 1 < &darray_top(wb->sorted) &&
      wb_key_eq(i, i + 1)) {
   struct btree_write_buffered_key *n = &wb->flushing.keys.data[i[1].idx];

   if (k->k.k.type == KEY_TYPE_accounting &&
       n->k.k.type == KEY_TYPE_accounting)
    bch2_accounting_accumulate(bkey_i_to_accounting(&n->k),
          bkey_i_to_s_c_accounting(&k->k));

   overwritten++;
   n->journal_seq = min_t(u64, n->journal_seq, k->journal_seq);
   k->journal_seq = 0;
   continue;
  }

  if (write_locked) {
   struct btree_path *path = btree_iter_path(trans, &iter);

   if (path->btree_id != i->btree ||
       bpos_gt(k->k.k.p, path->l[0].b->key.k.p)) {
    bch2_btree_node_unlock_write(trans, path, path->l[0].b);
    write_locked = false;

    ret = lockrestart_do(trans,
     bch2_btree_iter_traverse(trans, &iter) ?:
     bch2_foreground_maybe_merge(trans, iter.path, 0,
       BCH_WATERMARK_reclaim|
       BCH_TRANS_COMMIT_journal_reclaim|
       BCH_TRANS_COMMIT_no_check_rw|
       BCH_TRANS_COMMIT_no_enospc));
    if (ret)
     goto err;
   }
  }

  if (!iter.path || iter.btree_id != k->btree) {
   bch2_trans_iter_exit(trans, &iter);
   bch2_trans_iter_init(trans, &iter, k->btree, k->k.k.p,
          BTREE_ITER_intent|BTREE_ITER_all_snapshots);
  }

  bch2_btree_iter_set_pos(trans, &iter, k->k.k.p);
  btree_iter_path(trans, &iter)->preserve = false;

  bool accounting_accumulated = false;
  do {
   if (race_fault()) {
    ret = bch_err_throw(c, journal_reclaim_would_deadlock);
    break;
   }

   ret = wb_flush_one(trans, &iter, k, &write_locked,
        &accounting_accumulated, &fast);
   if (!write_locked)
    bch2_trans_begin(trans);
  } while (bch2_err_matches(ret, BCH_ERR_transaction_restart));

  if (!ret) {
   k->journal_seq = 0;
  } else if (ret == -BCH_ERR_journal_reclaim_would_deadlock) {
   slowpath++;
   ret = 0;
  } else
   break;
 }

 if (write_locked) {
  struct btree_path *path = btree_iter_path(trans, &iter);
  bch2_btree_node_unlock_write(trans, path, path->l[0].b);
 }
 bch2_trans_iter_exit(trans, &iter);

 if (ret)
  goto err;

 if (slowpath) {
  /*
 * Flush in the order they were present in the journal, so that
 * we can release journal pins:
 * The fastpath zapped the seq of keys that were successfully flushed so
 * we can skip those here.
 */
  trace_and_count(c, write_buffer_flush_slowpath, trans, slowpath, wb->flushing.keys.nr);

  sort_nonatomic(wb->flushing.keys.data,
          wb->flushing.keys.nr,
          sizeof(wb->flushing.keys.data[0]),
          wb_key_seq_cmp, NULL);

  darray_for_each(wb->flushing.keys, i) {
   if (!i->journal_seq)
    continue;

   if (!accounting_replay_done &&
       i->k.k.type == KEY_TYPE_accounting) {
    could_not_insert++;
    continue;
   }

   if (!could_not_insert)
    bch2_journal_pin_update(j, i->journal_seq, &wb->flushing.pin,
       bch2_btree_write_buffer_journal_flush);

   bch2_trans_begin(trans);

   ret = commit_do(trans, NULL, NULL,
     BCH_WATERMARK_reclaim|
     BCH_TRANS_COMMIT_journal_reclaim|
     BCH_TRANS_COMMIT_no_check_rw|
     BCH_TRANS_COMMIT_no_enospc|
     BCH_TRANS_COMMIT_no_journal_res ,
     btree_write_buffered_insert(trans, i));
   if (ret)
    goto err;

   i->journal_seq = 0;
  }

  /*
 * If journal replay hasn't finished with accounting keys we
 * can't flush accounting keys at all - condense them and leave
 * them for next time.
 *
 * Q: Can the write buffer overflow?
 * A Shouldn't be any actual risk. It's just new accounting
 * updates that the write buffer can't flush, and those are only
 * going to be generated by interior btree node updates as
 * journal replay has to split/rewrite nodes to make room for
 * its updates.
 *
 * And for those new acounting updates, updates to the same
 * counters get accumulated as they're flushed from the journal
 * to the write buffer - see the patch for eytzingcer tree
 * accumulated. So we could only overflow if the number of
 * distinct counters touched somehow was very large.
 */
  if (could_not_insert) {
   struct btree_write_buffered_key *dst = wb->flushing.keys.data;

   darray_for_each(wb->flushing.keys, i)
    if (i->journal_seq)
     *dst++ = *i;
   wb->flushing.keys.nr = dst - wb->flushing.keys.data;
  }
 }
err:
 if (ret || !could_not_insert) {
  bch2_journal_pin_drop(j, &wb->flushing.pin);
  wb->flushing.keys.nr = 0;
 }

 bch2_fs_fatal_err_on(ret, c, "%s", bch2_err_str(ret));
 trace_write_buffer_flush(trans, wb->flushing.keys.nr, overwritten, fast, 0);
 return ret;
}

static int bch2_journal_keys_to_write_buffer(struct bch_fs *c, struct journal_buf *buf)
{
 struct journal_keys_to_wb dst;
 int ret = 0;

 bch2_journal_keys_to_write_buffer_start(c, &dst, le64_to_cpu(buf->data->seq));

 for_each_jset_entry_type(entry, buf->data, BCH_JSET_ENTRY_write_buffer_keys) {
  jset_entry_for_each_key(entry, k) {
   ret = bch2_journal_key_to_wb(c, &dst, entry->btree_id, k);
   if (ret)
    goto out;
  }

  entry->type = BCH_JSET_ENTRY_btree_keys;
 }
out:
 ret = bch2_journal_keys_to_write_buffer_end(c, &dst) ?: ret;
 return ret;
}

static int fetch_wb_keys_from_journal(struct bch_fs *c, u64 max_seq)
{
 struct journal *j = &c->journal;
 struct journal_buf *buf;
 bool blocked;
 int ret = 0;

 while (!ret && (buf = bch2_next_write_buffer_flush_journal_buf(j, max_seq, &blocked))) {
  ret = bch2_journal_keys_to_write_buffer(c, buf);

  if (!blocked && !ret) {
   spin_lock(&j->lock);
   buf->need_flush_to_write_buffer = false;
   spin_unlock(&j->lock);
  }

  mutex_unlock(&j->buf_lock);

  if (blocked) {
   bch2_journal_unblock(j);
   break;
  }
 }

 return ret;
}

static int btree_write_buffer_flush_seq(struct btree_trans *trans, u64 max_seq,
     bool *did_work)
{
 struct bch_fs *c = trans->c;
 struct btree_write_buffer *wb = &c->btree_write_buffer;
 int ret = 0, fetch_from_journal_err;

 do {
  bch2_trans_unlock(trans);

  fetch_from_journal_err = fetch_wb_keys_from_journal(c, max_seq);

  *did_work |= wb->inc.keys.nr || wb->flushing.keys.nr;

  /*
 * On memory allocation failure, bch2_btree_write_buffer_flush_locked()
 * is not guaranteed to empty wb->inc:
 */
  mutex_lock(&wb->flushing.lock);
  ret = bch2_btree_write_buffer_flush_locked(trans);
  mutex_unlock(&wb->flushing.lock);
 } while (!ret &&
   (fetch_from_journal_err ||
    (wb->inc.pin.seq && wb->inc.pin.seq <= max_seq) ||
    (wb->flushing.pin.seq && wb->flushing.pin.seq <= max_seq)));

 return ret;
}

static int bch2_btree_write_buffer_journal_flush(struct journal *j,
    struct journal_entry_pin *_pin, u64 seq)
{
 struct bch_fs *c = container_of(j, struct bch_fs, journal);
 bool did_work = false;

 return bch2_trans_run(c, btree_write_buffer_flush_seq(trans, seq, &did_work));
}

int bch2_btree_write_buffer_flush_sync(struct btree_trans *trans)
{
 struct bch_fs *c = trans->c;
 bool did_work = false;

 trace_and_count(c, write_buffer_flush_sync, trans, _RET_IP_);

 return btree_write_buffer_flush_seq(trans, journal_cur_seq(&c->journal), &did_work);
}

/*
 * The write buffer requires flushing when going RO: keys in the journal for the
 * write buffer don't have a journal pin yet
 */
bool bch2_btree_write_buffer_flush_going_ro(struct bch_fs *c)
{
 if (bch2_journal_error(&c->journal))
  return false;

 bool did_work = false;
 bch2_trans_run(c, btree_write_buffer_flush_seq(trans,
    journal_cur_seq(&c->journal), &did_work));
 return did_work;
}

int bch2_btree_write_buffer_flush_nocheck_rw(struct btree_trans *trans)
{
 struct bch_fs *c = trans->c;
 struct btree_write_buffer *wb = &c->btree_write_buffer;
 int ret = 0;

 if (mutex_trylock(&wb->flushing.lock)) {
  ret = bch2_btree_write_buffer_flush_locked(trans);
  mutex_unlock(&wb->flushing.lock);
 }

 return ret;
}

int bch2_btree_write_buffer_tryflush(struct btree_trans *trans)
{
 struct bch_fs *c = trans->c;

 if (!enumerated_ref_tryget(&c->writes, BCH_WRITE_REF_btree_write_buffer))
  return bch_err_throw(c, erofs_no_writes);

 int ret = bch2_btree_write_buffer_flush_nocheck_rw(trans);
 enumerated_ref_put(&c->writes, BCH_WRITE_REF_btree_write_buffer);
 return ret;
}

/*
 * In check and repair code, when checking references to write buffer btrees we
 * need to issue a flush before we have a definitive error: this issues a flush
 * if this is a key we haven't yet checked.
 */
int bch2_btree_write_buffer_maybe_flush(struct btree_trans *trans,
     struct bkey_s_c referring_k,
     struct bkey_buf *last_flushed)
{
 struct bch_fs *c = trans->c;
 struct bkey_buf tmp;
 int ret = 0;

 bch2_bkey_buf_init(&tmp);

 if (!bkey_and_val_eq(referring_k, bkey_i_to_s_c(last_flushed->k))) {
  if (trace_write_buffer_maybe_flush_enabled()) {
   struct printbuf buf = PRINTBUF;

   bch2_bkey_val_to_text(&buf, c, referring_k);
   trace_write_buffer_maybe_flush(trans, _RET_IP_, buf.buf);
   printbuf_exit(&buf);
  }

  bch2_bkey_buf_reassemble(&tmp, c, referring_k);

  if (bkey_is_btree_ptr(referring_k.k)) {
   bch2_trans_unlock(trans);
   bch2_btree_interior_updates_flush(c);
  }

  ret = bch2_btree_write_buffer_flush_sync(trans);
  if (ret)
   goto err;

  bch2_bkey_buf_copy(last_flushed, c, tmp.k);

  /* can we avoid the unconditional restart? */
  trace_and_count(c, trans_restart_write_buffer_flush, trans, _RET_IP_);
  ret = bch_err_throw(c, transaction_restart_write_buffer_flush);
 }
err:
 bch2_bkey_buf_exit(&tmp, c);
 return ret;
}

static void bch2_btree_write_buffer_flush_work(struct work_struct *work)
{
 struct bch_fs *c = container_of(work, struct bch_fs, btree_write_buffer.flush_work);
 struct btree_write_buffer *wb = &c->btree_write_buffer;
 int ret;

 mutex_lock(&wb->flushing.lock);
 do {
  ret = bch2_trans_run(c, bch2_btree_write_buffer_flush_locked(trans));
 } while (!ret && bch2_btree_write_buffer_should_flush(c));
 mutex_unlock(&wb->flushing.lock);

 enumerated_ref_put(&c->writes, BCH_WRITE_REF_btree_write_buffer);
}

static void wb_accounting_sort(struct btree_write_buffer *wb)
{
 eytzinger0_sort(wb->accounting.data, wb->accounting.nr,
   sizeof(wb->accounting.data[0]),
   wb_key_cmp, NULL);
}

int bch2_accounting_key_to_wb_slowpath(struct bch_fs *c, enum btree_id btree,
           struct bkey_i_accounting *k)
{
 struct btree_write_buffer *wb = &c->btree_write_buffer;
 struct btree_write_buffered_key new = { .btree = btree };

 bkey_copy(&new.k, &k->k_i);

 int ret = darray_push(&wb->accounting, new);
 if (ret)
  return ret;

 wb_accounting_sort(wb);
 return 0;
}

int bch2_journal_key_to_wb_slowpath(struct bch_fs *c,
        struct journal_keys_to_wb *dst,
        enum btree_id btree, struct bkey_i *k)
{
 struct btree_write_buffer *wb = &c->btree_write_buffer;
 int ret;
retry:
 ret = darray_make_room_gfp(&dst->wb->keys, 1, GFP_KERNEL);
 if (!ret && dst->wb == &wb->flushing)
  ret = darray_resize(&wb->sorted, wb->flushing.keys.size);

 if (unlikely(ret)) {
  if (dst->wb == &c->btree_write_buffer.flushing) {
   mutex_unlock(&dst->wb->lock);
   dst->wb = &c->btree_write_buffer.inc;
   bch2_journal_pin_add(&c->journal, dst->seq, &dst->wb->pin,
          bch2_btree_write_buffer_journal_flush);
   goto retry;
  }

  return ret;
 }

 dst->room = darray_room(dst->wb->keys);
 if (dst->wb == &wb->flushing)
  dst->room = min(dst->room, wb->sorted.size - wb->flushing.keys.nr);
 BUG_ON(!dst->room);
 BUG_ON(!dst->seq);

 struct btree_write_buffered_key *wb_k = &darray_top(dst->wb->keys);
 wb_k->journal_seq = dst->seq;
 wb_k->btree  = btree;
 bkey_copy(&wb_k->k, k);
 dst->wb->keys.nr++;
 dst->room--;
 return 0;
}

void bch2_journal_keys_to_write_buffer_start(struct bch_fs *c, struct journal_keys_to_wb *dst, u64 seq)
{
 struct btree_write_buffer *wb = &c->btree_write_buffer;

 if (mutex_trylock(&wb->flushing.lock)) {
  mutex_lock(&wb->inc.lock);
  move_keys_from_inc_to_flushing(wb);

  /*
 * Attempt to skip wb->inc, and add keys directly to
 * wb->flushing, saving us a copy later:
 */

  if (!wb->inc.keys.nr) {
   dst->wb = &wb->flushing;
  } else {
   mutex_unlock(&wb->flushing.lock);
   dst->wb = &wb->inc;
  }
 } else {
  mutex_lock(&wb->inc.lock);
  dst->wb = &wb->inc;
 }

 dst->room = darray_room(dst->wb->keys);
 if (dst->wb == &wb->flushing)
  dst->room = min(dst->room, wb->sorted.size - wb->flushing.keys.nr);
 dst->seq = seq;

 bch2_journal_pin_add(&c->journal, seq, &dst->wb->pin,
        bch2_btree_write_buffer_journal_flush);

 darray_for_each(wb->accounting, i)
  memset(&i->k.v, 0, bkey_val_bytes(&i->k.k));
}

int bch2_journal_keys_to_write_buffer_end(struct bch_fs *c, struct journal_keys_to_wb *dst)
{
 struct btree_write_buffer *wb = &c->btree_write_buffer;
 unsigned live_accounting_keys = 0;
 int ret = 0;

 darray_for_each(wb->accounting, i)
  if (!bch2_accounting_key_is_zero(bkey_i_to_s_c_accounting(&i->k))) {
   i->journal_seq = dst->seq;
   live_accounting_keys++;
   ret = __bch2_journal_key_to_wb(c, dst, i->btree, &i->k);
   if (ret)
    break;
  }

 if (live_accounting_keys * 2 < wb->accounting.nr) {
  struct btree_write_buffered_key *dst = wb->accounting.data;

  darray_for_each(wb->accounting, src)
   if (!bch2_accounting_key_is_zero(bkey_i_to_s_c_accounting(&src->k)))
    *dst++ = *src;
  wb->accounting.nr = dst - wb->accounting.data;
  wb_accounting_sort(wb);
 }

 if (!dst->wb->keys.nr)
  bch2_journal_pin_drop(&c->journal, &dst->wb->pin);

 if (bch2_btree_write_buffer_should_flush(c) &&
     __enumerated_ref_tryget(&c->writes, BCH_WRITE_REF_btree_write_buffer) &&
     !queue_work(system_unbound_wq, &c->btree_write_buffer.flush_work))
  enumerated_ref_put(&c->writes, BCH_WRITE_REF_btree_write_buffer);

 if (dst->wb == &wb->flushing)
  mutex_unlock(&wb->flushing.lock);
 mutex_unlock(&wb->inc.lock);

 return ret;
}

static int wb_keys_resize(struct btree_write_buffer_keys *wb, size_t new_size)
{
 if (wb->keys.size >= new_size)
  return 0;

 if (!mutex_trylock(&wb->lock))
  return -EINTR;

 int ret = darray_resize(&wb->keys, new_size);
 mutex_unlock(&wb->lock);
 return ret;
}

int bch2_btree_write_buffer_resize(struct bch_fs *c, size_t new_size)
{
 struct btree_write_buffer *wb = &c->btree_write_buffer;

 return wb_keys_resize(&wb->flushing, new_size) ?:
  wb_keys_resize(&wb->inc, new_size);
}

void bch2_fs_btree_write_buffer_exit(struct bch_fs *c)
{
 struct btree_write_buffer *wb = &c->btree_write_buffer;

 BUG_ON((wb->inc.keys.nr || wb->flushing.keys.nr) &&
        !bch2_journal_error(&c->journal));

 darray_exit(&wb->accounting);
 darray_exit(&wb->sorted);
 darray_exit(&wb->flushing.keys);
 darray_exit(&wb->inc.keys);
}

void bch2_fs_btree_write_buffer_init_early(struct bch_fs *c)
{
 struct btree_write_buffer *wb = &c->btree_write_buffer;

 mutex_init(&wb->inc.lock);
 mutex_init(&wb->flushing.lock);
 INIT_WORK(&wb->flush_work, bch2_btree_write_buffer_flush_work);
}

int bch2_fs_btree_write_buffer_init(struct bch_fs *c)
{
 struct btree_write_buffer *wb = &c->btree_write_buffer;

 /* Will be resized by journal as needed: */
 unsigned initial_size = 1 << 16;

 return  darray_make_room(&wb->inc.keys, initial_size) ?:
  darray_make_room(&wb->flushing.keys, initial_size) ?:
  darray_make_room(&wb->sorted, initial_size);
}