On Sat, Feb 24, 2024 at 09:38:04PM -0500, Kent Overstreet wrote:
> Until accounting keys hit the btree, they are deltas, not new versions
> of the existing key; this means we have to teach journal replay to
> accumulate them.
>
> Additionally, the journal doesn't track precisely which entries have
> been flushed to the btree; it only tracks a range of entries that may
> possibly still need to be flushed.
>
> That means we need to compare accounting keys against the version in the
> btree and only flush updates that are newer.
>
> There's another wrinkle with the write buffer: if the write buffer
> starts flushing accounting keys before journal replay has finished
> flushing accounting keys, journal replay will see the version number
> from the new updates and updates from the journal will be lost.
>
> To avoid this, journal replay has to flush accounting keys first, and
> we'll be adding a flag so that write buffer flush knows to hold
> accounting keys until then.
>
> Signed-off-by: Kent Overstreet <kent.overstr...@linux.dev>
> ---
> fs/bcachefs/btree_journal_iter.c | 23 +++-------
> fs/bcachefs/btree_journal_iter.h | 15 +++++++
> fs/bcachefs/btree_trans_commit.c | 9 +++-
> fs/bcachefs/btree_update.h | 14 +++++-
> fs/bcachefs/recovery.c | 76 +++++++++++++++++++++++++++++++-
> 5 files changed, 117 insertions(+), 20 deletions(-)
>
...
> diff --git a/fs/bcachefs/recovery.c b/fs/bcachefs/recovery.c
> index 96e7a1ec7091..6829d80bd181 100644
> --- a/fs/bcachefs/recovery.c
> +++ b/fs/bcachefs/recovery.c
> @@ -11,6 +11,7 @@
> #include "btree_io.h"
> #include "buckets.h"
> #include "dirent.h"
> +#include "disk_accounting.h"
> #include "ec.h"
> #include "errcode.h"
> #include "error.h"
> @@ -87,6 +88,56 @@ static void replay_now_at(struct journal *j, u64 seq)
> bch2_journal_pin_put(j, j->replay_journal_seq++);
> }
>
> +static int bch2_journal_replay_accounting_key(struct btree_trans *trans,
> + struct journal_key *k)
> +{
> + struct journal_keys *keys = &trans->c->journal_keys;
> +
> + struct btree_iter iter;
> + bch2_trans_node_iter_init(trans, &iter, k->btree_id, k->k->k.p,
> + BTREE_MAX_DEPTH, k->level,
> + BTREE_ITER_INTENT);
> + int ret = bch2_btree_iter_traverse(&iter);
> + if (ret)
> + goto out;
> +
> + struct bkey u;
> + struct bkey_s_c old = bch2_btree_path_peek_slot(btree_iter_path(trans,
> &iter), &u);
> +
> + if (bversion_cmp(old.k->version, k->k->k.version) >= 0) {
> + ret = 0;
> + goto out;
> + }
So I assume this is what correlates back to the need to not flush the
write buffer until replay completes, otherwise we could unintentionally
skip subsequent key updates. Is that the case?
If so, it would be nice to have some comments here that explain this.
I.e., I don't quite have a big enough picture to know where or how this
is prevented to ensure that the version updates down the key
accumulation helpers don't conflict with this particular check, so
something that helps connect the dots enough to somebody who doesn't
already know how this is all supposed to work would be useful.
Brian
> +
> + if (k + 1 < &darray_top(*keys) &&
> + !journal_key_cmp(k, k + 1)) {
> + BUG_ON(bversion_cmp(k[0].k->k.version, k[1].k->k.version) > 0);
> +
> + bch2_accounting_accumulate(bkey_i_to_accounting(k[1].k),
> + bkey_i_to_s_c_accounting(k[0].k));
> + ret = 0;
> + goto out;
> + }
> +
> + struct bkey_i *new = k->k;
> + if (old.k->type == KEY_TYPE_accounting) {
> + new = bch2_bkey_make_mut_noupdate(trans, bkey_i_to_s_c(k->k));
> + ret = PTR_ERR_OR_ZERO(new);
> + if (ret)
> + goto out;
> +
> + bch2_accounting_accumulate(bkey_i_to_accounting(new),
> + bkey_s_c_to_accounting(old));
> + }
> +
> + trans->journal_res.seq = k->journal_seq;
> +
> + ret = bch2_trans_update(trans, &iter, new, BTREE_TRIGGER_NORUN);
> +out:
> + bch2_trans_iter_exit(trans, &iter);
> + return ret;
> +}
> +
> static int bch2_journal_replay_key(struct btree_trans *trans,
> struct journal_key *k)
> {
> @@ -159,12 +210,33 @@ static int bch2_journal_replay(struct bch_fs *c)
>
> BUG_ON(!atomic_read(&keys->ref));
>
> + /*
> + * Replay accounting keys first: we can't allow the write buffer to
> + * flush accounting keys until we're done
> + */
> + darray_for_each(*keys, k) {
> + if (!(k->k->k.type == KEY_TYPE_accounting && !k->allocated))
> + continue;
> +
> + cond_resched();
> +
> + ret = commit_do(trans, NULL, NULL,
> + BCH_TRANS_COMMIT_no_enospc|
> + BCH_TRANS_COMMIT_no_journal_res,
> + bch2_journal_replay_accounting_key(trans, k));
> + if (bch2_fs_fatal_err_on(ret, c, "error replaying accounting;
> %s", bch2_err_str(ret)))
> + goto err;
> + }
> +
> /*
> * First, attempt to replay keys in sorted order. This is more
> * efficient - better locality of btree access - but some might fail if
> * that would cause a journal deadlock.
> */
> darray_for_each(*keys, k) {
> + if (k->k->k.type == KEY_TYPE_accounting && !k->allocated)
> + continue;
> +
> cond_resched();
>
> /* Skip fastpath if we're low on space in the journal */
> @@ -174,7 +246,7 @@ static int bch2_journal_replay(struct bch_fs *c)
> BCH_TRANS_COMMIT_journal_reclaim|
> (!k->allocated ?
> BCH_TRANS_COMMIT_no_journal_res : 0),
> bch2_journal_replay_key(trans, k));
> - BUG_ON(!ret && !k->overwritten);
> + BUG_ON(!ret && !k->overwritten && k->k->k.type !=
> KEY_TYPE_accounting);
> if (ret) {
> ret = darray_push(&keys_sorted, k);
> if (ret)
> @@ -208,7 +280,7 @@ static int bch2_journal_replay(struct bch_fs *c)
> if (ret)
> goto err;
>
> - BUG_ON(!k->overwritten);
> + BUG_ON(k->btree_id != BTREE_ID_accounting && !k->overwritten);
> }
>
> /*
> --
> 2.43.0
>
