bcachefs: Data update path won't accidentaly grow replicas

Previously, there was a bug where if an extent had greater durability
than required (because we needed to move a durability=1 pointer and
ended up putting it on a durability 2 device), we would submit a write
for replicas=2 - the durability of the pointer being rewritten - instead
of the number of replicas required to bring it back up to the
data_replicas option.

This, plus the allocation path sometimes allocating on a greater
durability device than requested, meant that extents could continue
having more and more replicas added as they were being rewritten.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

Author: Kent Overstreet

fs/bcachefs/data_update.c

@@ -356,7 +356,7 @@ void bch2_data_update_exit(struct data_update *update)
 	bch2_bio_free_pages_pool(c, &update->op.wbio.bio);
 }
 
-void bch2_update_unwritten_extent(struct btree_trans *trans,
+static void bch2_update_unwritten_extent(struct btree_trans *trans,
 				  struct data_update *update)
 {
 	struct bch_fs *c = update->op.c;
@@ -436,7 +436,51 @@ void bch2_update_unwritten_extent(struct btree_trans *trans,
 	}
 }
 
+int bch2_extent_drop_ptrs(struct btree_trans *trans,
+			  struct btree_iter *iter,
+			  struct bkey_s_c k,
+			  struct data_update_opts data_opts)
+{
+	struct bch_fs *c = trans->c;
+	struct bkey_i *n;
+	int ret;
+
+	n = bch2_bkey_make_mut_noupdate(trans, k);
+	ret = PTR_ERR_OR_ZERO(n);
+	if (ret)
+		return ret;
+
+	while (data_opts.kill_ptrs) {
+		unsigned i = 0, drop = __fls(data_opts.kill_ptrs);
+		struct bch_extent_ptr *ptr;
+
+		bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, i++ == drop);
+		data_opts.kill_ptrs ^= 1U << drop;
+	}
+
+	/*
+	 * If the new extent no longer has any pointers, bch2_extent_normalize()
+	 * will do the appropriate thing with it (turning it into a
+	 * KEY_TYPE_error key, or just a discard if it was a cached extent)
+	 */
+	bch2_extent_normalize(c, bkey_i_to_s(n));
+
+	/*
+	 * Since we're not inserting through an extent iterator
+	 * (BTREE_ITER_ALL_SNAPSHOTS iterators aren't extent iterators),
+	 * we aren't using the extent overwrite path to delete, we're
+	 * just using the normal key deletion path:
+	 */
+	if (bkey_deleted(&n->k))
+		n->k.size = 0;
+
+	return bch2_trans_relock(trans) ?:
+		bch2_trans_update(trans, iter, n, BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
+		bch2_trans_commit(trans, NULL, NULL, BTREE_INSERT_NOFAIL);
+}
+
 int bch2_data_update_init(struct btree_trans *trans,
+			  struct btree_iter *iter,
 			  struct moving_context *ctxt,
 			  struct data_update *m,
 			  struct write_point_specifier wp,
@@ -452,7 +496,7 @@ int bch2_data_update_init(struct btree_trans *trans,
 	const struct bch_extent_ptr *ptr;
 	unsigned i, reserve_sectors = k.k->size * data_opts.extra_replicas;
 	unsigned ptrs_locked = 0;
-	int ret;
+	int ret = 0;
 
 	bch2_bkey_buf_init(&m->k);
 	bch2_bkey_buf_reassemble(&m->k, c, k);
@@ -478,6 +522,8 @@ int bch2_data_update_init(struct btree_trans *trans,
 	bkey_for_each_ptr(ptrs, ptr)
 		percpu_ref_get(&bch_dev_bkey_exists(c, ptr->dev)->ref);
 
+	unsigned durability_have = 0, durability_removing = 0;
+
 	i = 0;
 	bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 		bool locked;
@@ -489,8 +535,11 @@ int bch2_data_update_init(struct btree_trans *trans,
 				reserve_sectors += k.k->size;
 
 			m->op.nr_replicas += bch2_extent_ptr_desired_durability(c, &p);
-		} else if (!p.ptr.cached) {
+			durability_removing += bch2_extent_ptr_desired_durability(c, &p);
+		} else if (!p.ptr.cached &&
+			   !((1U << i) & m->data_opts.kill_ptrs)) {
 			bch2_dev_list_add_dev(&m->op.devs_have, p.ptr.dev);
+			durability_have += bch2_extent_ptr_durability(c, &p);
 		}
 
 		/*
@@ -529,6 +578,29 @@ int bch2_data_update_init(struct btree_trans *trans,
 		i++;
 	}
 
+	/*
+	 * If current extent durability is less than io_opts.data_replicas,
+	 * we're not trying to rereplicate the extent up to data_replicas here -
+	 * unless extra_replicas was specified
+	 *
+	 * Increasing replication is an explicit operation triggered by
+	 * rereplicate, currently, so that users don't get an unexpected -ENOSPC
+	 */
+	if (durability_have >= io_opts.data_replicas) {
+		m->data_opts.kill_ptrs |= m->data_opts.rewrite_ptrs;
+		m->data_opts.rewrite_ptrs = 0;
+		/* if iter == NULL, it's just a promote */
+		if (iter)
+			ret = bch2_extent_drop_ptrs(trans, iter, k, data_opts);
+		goto done;
+	}
+
+	m->op.nr_replicas = min(durability_removing, io_opts.data_replicas - durability_have) +
+		m->data_opts.extra_replicas;
+	m->op.nr_replicas_required = m->op.nr_replicas;
+
+	BUG_ON(!m->op.nr_replicas);
+
 	if (reserve_sectors) {
 		ret = bch2_disk_reservation_add(c, &m->op.res, reserve_sectors,
 				m->data_opts.extra_replicas
@@ -538,14 +610,11 @@ int bch2_data_update_init(struct btree_trans *trans,
 			goto err;
 	}
 
-	m->op.nr_replicas += m->data_opts.extra_replicas;
-	m->op.nr_replicas_required = m->op.nr_replicas;
-
-	BUG_ON(!m->op.nr_replicas);
+	if (bkey_extent_is_unwritten(k)) {
+		bch2_update_unwritten_extent(trans, m);
+		goto done;
+	}
 
-	/* Special handling required: */
-	if (bkey_extent_is_unwritten(k))
-		return -BCH_ERR_unwritten_extent_update;
 	return 0;
 err:
 	i = 0;
@@ -560,6 +629,9 @@ int bch2_data_update_init(struct btree_trans *trans,
 	bch2_bkey_buf_exit(&m->k, c);
 	bch2_bio_free_pages_pool(c, &m->op.wbio.bio);
 	return ret;
+done:
+	bch2_data_update_exit(m);
+	return ret ?: -BCH_ERR_data_update_done;
 }
 
 void bch2_data_update_opts_normalize(struct bkey_s_c k, struct data_update_opts *opts)

fs/bcachefs/data_update.h

@@ -32,9 +32,14 @@ int bch2_data_update_index_update(struct bch_write_op *);
 void bch2_data_update_read_done(struct data_update *,
 				struct bch_extent_crc_unpacked);
 
+int bch2_extent_drop_ptrs(struct btree_trans *,
+			  struct btree_iter *,
+			  struct bkey_s_c,
+			  struct data_update_opts);
+
 void bch2_data_update_exit(struct data_update *);
-void bch2_update_unwritten_extent(struct btree_trans *, struct data_update *);
-int bch2_data_update_init(struct btree_trans *, struct moving_context *,
+int bch2_data_update_init(struct btree_trans *, struct btree_iter *,
+			  struct moving_context *,
 			  struct data_update *,
 			  struct write_point_specifier,
 			  struct bch_io_opts, struct data_update_opts,

fs/bcachefs/errcode.h

@@ -162,7 +162,7 @@
 	x(BCH_ERR_fsck,			fsck_repair_unimplemented)		\
 	x(BCH_ERR_fsck,			fsck_repair_impossible)			\
 	x(0,				restart_recovery)			\
-	x(0,				unwritten_extent_update)		\
+	x(0,				data_update_done)			\
 	x(EINVAL,			device_state_not_allowed)		\
 	x(EINVAL,			member_info_missing)			\
 	x(EINVAL,			mismatched_block_size)			\

fs/bcachefs/io_read.c

@@ -209,7 +209,7 @@ static struct promote_op *__promote_alloc(struct btree_trans *trans,
 	bio = &op->write.op.wbio.bio;
 	bio_init(bio, NULL, bio->bi_inline_vecs, pages, 0);
 
-	ret = bch2_data_update_init(trans, NULL, &op->write,
+	ret = bch2_data_update_init(trans, NULL, NULL, &op->write,
 			writepoint_hashed((unsigned long) current),
 			opts,
 			(struct data_update_opts) {

fs/bcachefs/move.c

@@ -229,49 +229,6 @@ void bch2_move_stats_init(struct bch_move_stats *stats, char *name)
 	scnprintf(stats->name, sizeof(stats->name), "%s", name);
 }
 
-static int bch2_extent_drop_ptrs(struct btree_trans *trans,
-				 struct btree_iter *iter,
-				 struct bkey_s_c k,
-				 struct data_update_opts data_opts)
-{
-	struct bch_fs *c = trans->c;
-	struct bkey_i *n;
-	int ret;
-
-	n = bch2_bkey_make_mut_noupdate(trans, k);
-	ret = PTR_ERR_OR_ZERO(n);
-	if (ret)
-		return ret;
-
-	while (data_opts.kill_ptrs) {
-		unsigned i = 0, drop = __fls(data_opts.kill_ptrs);
-		struct bch_extent_ptr *ptr;
-
-		bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, i++ == drop);
-		data_opts.kill_ptrs ^= 1U << drop;
-	}
-
-	/*
-	 * If the new extent no longer has any pointers, bch2_extent_normalize()
-	 * will do the appropriate thing with it (turning it into a
-	 * KEY_TYPE_error key, or just a discard if it was a cached extent)
-	 */
-	bch2_extent_normalize(c, bkey_i_to_s(n));
-
-	/*
-	 * Since we're not inserting through an extent iterator
-	 * (BTREE_ITER_ALL_SNAPSHOTS iterators aren't extent iterators),
-	 * we aren't using the extent overwrite path to delete, we're
-	 * just using the normal key deletion path:
-	 */
-	if (bkey_deleted(&n->k))
-		n->k.size = 0;
-
-	return bch2_trans_relock(trans) ?:
-		bch2_trans_update(trans, iter, n, BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
-		bch2_trans_commit(trans, NULL, NULL, BTREE_INSERT_NOFAIL);
-}
-
 int bch2_move_extent(struct moving_context *ctxt,
 		     struct move_bucket_in_flight *bucket_in_flight,
 		     struct btree_iter *iter,
@@ -341,19 +298,11 @@ int bch2_move_extent(struct moving_context *ctxt,
 	io->rbio.bio.bi_iter.bi_sector	= bkey_start_offset(k.k);
 	io->rbio.bio.bi_end_io		= move_read_endio;
 
-	ret = bch2_data_update_init(trans, ctxt, &io->write, ctxt->wp,
+	ret = bch2_data_update_init(trans, iter, ctxt, &io->write, ctxt->wp,
 				    io_opts, data_opts, iter->btree_id, k);
-	if (ret && ret != -BCH_ERR_unwritten_extent_update)
+	if (ret)
 		goto err_free_pages;
 
-	if (ret == -BCH_ERR_unwritten_extent_update) {
-		bch2_update_unwritten_extent(trans, &io->write);
-		move_free(io);
-		return 0;
-	}
-
-	BUG_ON(ret);
-
 	io->write.op.end_io = move_write_done;
 
 	if (ctxt->rate)
@@ -397,6 +346,9 @@ int bch2_move_extent(struct moving_context *ctxt,
 err_free:
 	kfree(io);
 err:
+	if (ret == -BCH_ERR_data_update_done)
+		return 0;
+
 	this_cpu_inc(c->counters[BCH_COUNTER_move_extent_alloc_mem_fail]);
 	trace_move_extent_alloc_mem_fail2(c, k);
 	return ret;