bcachefs: btree node scan no longer uses btree cache

Previously, btree node scan used the btree node cache to check if btree
nodes were readable, but this is subject to interference from threads
scanning different devices trying to read the same node - and more
critically, nodes that we already attempted and failed to read before
kicking off scan.

Instead, we now allocate a 'struct btree' that does not live in the
btree node cache, and call bch2_btree_node_read_done() directly.

Cc: Nikita Ofitserov <[email protected]>
Reviewed-by: Nikita Ofitserov <[email protected]>
Reported-and-tested-by: Edoardo Codeglia <[email protected]>
Signed-off-by: Kent Overstreet <[email protected]>

Author: Kent Overstreet

fs/bcachefs/btree_node_scan.c

@@ -75,39 +75,6 @@ static inline u64 bkey_journal_seq(struct bkey_s_c k)
 	}
 }
 
-static bool found_btree_node_is_readable(struct btree_trans *trans,
-					 struct found_btree_node *f)
-{
-	struct { __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX); } tmp;
-
-	found_btree_node_to_key(&tmp.k, f);
-
-	struct btree *b = bch2_btree_node_get_noiter(trans, &tmp.k, f->btree_id, f->level, false);
-	bool ret = !IS_ERR_OR_NULL(b);
-	if (!ret)
-		return ret;
-
-	f->sectors_written = b->written;
-	f->journal_seq = le64_to_cpu(b->data->keys.journal_seq);
-
-	struct bkey_s_c k;
-	struct bkey unpacked;
-	struct btree_node_iter iter;
-	for_each_btree_node_key_unpack(b, k, &iter, &unpacked)
-		f->journal_seq = max(f->journal_seq, bkey_journal_seq(k));
-
-	six_unlock_read(&b->c.lock);
-
-	/*
-	 * We might update this node's range; if that happens, we need the node
-	 * to be re-read so the read path can trim keys that are no longer in
-	 * this node
-	 */
-	if (b != btree_node_root(trans->c, b))
-		bch2_btree_node_evict(trans, &tmp.k);
-	return ret;
-}
-
 static int found_btree_node_cmp_cookie(const void *_l, const void *_r)
 {
 	const struct found_btree_node *l = _l;
@@ -159,17 +126,17 @@ static const struct min_heap_callbacks found_btree_node_heap_cbs = {
 };
 
 static void try_read_btree_node(struct find_btree_nodes *f, struct bch_dev *ca,
-				struct bio *bio, struct btree_node *bn, u64 offset)
+				struct btree *b, struct bio *bio, u64 offset)
 {
 	struct bch_fs *c = container_of(f, struct bch_fs, found_btree_nodes);
+	struct btree_node *bn = b->data;
 
 	bio_reset(bio, ca->disk_sb.bdev, REQ_OP_READ);
 	bio->bi_iter.bi_sector	= offset;
-	bch2_bio_map(bio, bn, PAGE_SIZE);
+	bch2_bio_map(bio, b->data, c->opts.block_size);
 
 	u64 submit_time = local_clock();
 	submit_bio_wait(bio);
-
 	bch2_account_io_completion(ca, BCH_MEMBER_ERROR_read, submit_time, !bio->bi_status);
 
 	if (bio->bi_status) {
@@ -201,6 +168,14 @@ static void try_read_btree_node(struct find_btree_nodes *f, struct bch_dev *ca,
 	if (BTREE_NODE_ID(bn) >= BTREE_ID_NR_MAX)
 		return;
 
+	bio_reset(bio, ca->disk_sb.bdev, REQ_OP_READ);
+	bio->bi_iter.bi_sector	= offset;
+	bch2_bio_map(bio, b->data, c->opts.btree_node_size);
+
+	submit_time = local_clock();
+	submit_bio_wait(bio);
+	bch2_account_io_completion(ca, BCH_MEMBER_ERROR_read, submit_time, !bio->bi_status);
+
 	rcu_read_lock();
 	struct found_btree_node n = {
 		.btree_id	= BTREE_NODE_ID(bn),
@@ -217,7 +192,20 @@ static void try_read_btree_node(struct find_btree_nodes *f, struct bch_dev *ca,
 	};
 	rcu_read_unlock();
 
-	if (bch2_trans_run(c, found_btree_node_is_readable(trans, &n))) {
+	found_btree_node_to_key(&b->key, &n);
+
+	CLASS(printbuf, buf)();
+	if (!bch2_btree_node_read_done(c, ca, b, NULL, &buf)) {
+		/* read_done will swap out b->data for another buffer */
+		bn = b->data;
+		/*
+		 * Grab journal_seq here because we want the max journal_seq of
+		 * any bset; read_done sorts down to a single set and picks the
+		 * max journal_seq
+		 */
+		n.journal_seq		= le64_to_cpu(bn->keys.journal_seq),
+		n.sectors_written	= b->written;
+
 		mutex_lock(&f->lock);
 		if (BSET_BIG_ENDIAN(&bn->keys) != CPU_BIG_ENDIAN) {
 			bch_err(c, "try_read_btree_node() can't handle endian conversion");
@@ -237,12 +225,20 @@ static int read_btree_nodes_worker(void *p)
 	struct find_btree_nodes_worker *w = p;
 	struct bch_fs *c = container_of(w->f, struct bch_fs, found_btree_nodes);
 	struct bch_dev *ca = w->ca;
-	void *buf = (void *) __get_free_page(GFP_KERNEL);
-	struct bio *bio = bio_alloc(NULL, 1, 0, GFP_KERNEL);
 	unsigned long last_print = jiffies;
+	struct btree *b = NULL;
+	struct bio *bio = NULL;
+
+	b = __bch2_btree_node_mem_alloc(c);
+	if (!b) {
+		bch_err(c, "read_btree_nodes_worker: error allocating buf");
+		w->f->ret = -ENOMEM;
+		goto err;
+	}
 
-	if (!buf || !bio) {
-		bch_err(c, "read_btree_nodes_worker: error allocating bio/buf");
+	bio = bio_alloc(NULL, buf_pages(b->data, c->opts.btree_node_size), 0, GFP_KERNEL);
+	if (!bio) {
+		bch_err(c, "read_btree_nodes_worker: error allocating bio");
 		w->f->ret = -ENOMEM;
 		goto err;
 	}
@@ -266,11 +262,13 @@ static int read_btree_nodes_worker(void *p)
 			    !bch2_dev_btree_bitmap_marked_sectors(ca, sector, btree_sectors(c)))
 				continue;
 
-			try_read_btree_node(w->f, ca, bio, buf, sector);
+			try_read_btree_node(w->f, ca, b, bio, sector);
 		}
 err:
+	if (b)
+		__btree_node_data_free(b);
+	kfree(b);
 	bio_put(bio);
-	free_page((unsigned long) buf);
 	enumerated_ref_put(&ca->io_ref[READ], BCH_DEV_READ_REF_btree_node_scan);
 	closure_put(w->cl);
 	kfree(w);