bcachefs: Accumulate accounting keys in journal replay

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.overstreet@linux.dev>

1 files changed, 70 insertions, 2 deletions

diff --git a/fs/bcachefs/recovery.c b/fs/bcachefs/recovery.c
index d336a7c69edd..0091af3beeef 100644
--- a/fs/bcachefs/recovery.c
+++ b/fs/bcachefs/recovery.c
@@ -10,6 +10,7 @@
 #include "btree_io.h"
 #include "buckets.h"
 #include "dirent.h"
+#include "disk_accounting.h"
 #include "errcode.h"
 #include "error.h"
 #include "fs-common.h"
@@ -135,6 +136,47 @@ 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);
+
+	/* Has this delta already been applied to the btree? */
+	if (bversion_cmp(old.k->version, k->k->k.version) >= 0) {
+		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)
 {
@@ -185,6 +227,11 @@ static int bch2_journal_replay_key(struct btree_trans *trans,
 	if (k->overwritten)
 		goto out;
 
+	if (k->k->k.type == KEY_TYPE_accounting) {
+		ret = bch2_trans_update_buffered(trans, BTREE_ID_accounting, k->k);
+		goto out;
+	}
+
 	ret = bch2_trans_update(trans, &iter, k->k, update_flags);
 out:
 	bch2_trans_iter_exit(trans, &iter);
@@ -223,6 +270,27 @@ int bch2_journal_replay(struct bch_fs *c)
 	trans = bch2_trans_get(c);
 
 	/*
+	 * 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_journal_reclaim|
+				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;
+
+		k->overwritten = true;
+	}
+
+	/*
 	 * 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.
@@ -244,7 +312,7 @@ 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)
@@ -281,7 +349,7 @@ 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);
 	}
 
 	/*