bcachefs: bch2_trans_inconsistent()

Add a new error macro that also dumps transaction updates in addition to
doing an emergency shutdown - when a transaction update discovers or is
causing a fs inconsistency, it's helpful to see what updates it was
doing.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>

fs/bcachefs/btree_iter.c

@@ -1794,19 +1794,44 @@ void bch2_path_put(struct btree_trans *trans, struct btree_path *path, bool inte
 }
 
 noinline __cold
-void bch2_dump_trans_paths_updates(struct btree_trans *trans)
+void bch2_dump_trans_updates(struct btree_trans *trans)
 {
-	struct btree_path *path;
 	struct btree_insert_entry *i;
 	struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
+
+	bch_err(trans->c, "transaction updates:");
+
+	trans_for_each_update(trans, i) {
+		struct bkey_s_c old = { &i->old_k, i->old_v };
+
+		printbuf_reset(&buf1);
+		printbuf_reset(&buf2);
+		bch2_bkey_val_to_text(&buf1, trans->c, old);
+		bch2_bkey_val_to_text(&buf2, trans->c, bkey_i_to_s_c(i->k));
+
+		printk(KERN_ERR "update: btree %s %pS\n  old %s\n  new %s",
+		       bch2_btree_ids[i->btree_id],
+		       (void *) i->ip_allocated,
+		       buf1.buf, buf2.buf);
+	}
+
+	printbuf_exit(&buf2);
+	printbuf_exit(&buf1);
+}
+
+noinline __cold
+void bch2_dump_trans_paths_updates(struct btree_trans *trans)
+{
+	struct btree_path *path;
+	struct printbuf buf = PRINTBUF;
 	unsigned idx;
 
 	btree_trans_sort_paths(trans);
 
 	trans_for_each_path_inorder(trans, path, idx) {
-		printbuf_reset(&buf1);
+		printbuf_reset(&buf);
 
-		bch2_bpos_to_text(&buf1, path->pos);
+		bch2_bpos_to_text(&buf, path->pos);
 
 		printk(KERN_ERR "path: idx %u ref %u:%u%s%s btree=%s l=%u pos %s locks %u %pS\n",
 		       path->idx, path->ref, path->intent_ref,
@@ -1814,7 +1839,7 @@ void bch2_dump_trans_paths_updates(struct btree_trans *trans)
 		       path->preserve ? " P" : "",
 		       bch2_btree_ids[path->btree_id],
 		       path->level,
-		       buf1.buf,
+		       buf.buf,
 		       path->nodes_locked,
 #ifdef CONFIG_BCACHEFS_DEBUG
 		       (void *) path->ip_allocated
@@ -1824,23 +1849,9 @@ void bch2_dump_trans_paths_updates(struct btree_trans *trans)
 		       );
 	}
 
-	trans_for_each_update(trans, i) {
-		struct bkey u;
-		struct bkey_s_c old = bch2_btree_path_peek_slot(i->path, &u);
+	printbuf_exit(&buf);
 
-		printbuf_reset(&buf1);
-		printbuf_reset(&buf2);
-		bch2_bkey_val_to_text(&buf1, trans->c, old);
-		bch2_bkey_val_to_text(&buf2, trans->c, bkey_i_to_s_c(i->k));
-
-		printk(KERN_ERR "update: btree %s %pS\n  old %s\n  new %s",
-		       bch2_btree_ids[i->btree_id],
-		       (void *) i->ip_allocated,
-		       buf1.buf, buf2.buf);
-	}
-
-	printbuf_exit(&buf2);
-	printbuf_exit(&buf1);
+	bch2_dump_trans_updates(trans);
 }
 
 static struct btree_path *btree_path_alloc(struct btree_trans *trans,

fs/bcachefs/btree_iter.h

@@ -425,6 +425,7 @@ __bch2_btree_iter_peek_and_restart(struct btree_trans *trans,
 
 /* new multiple iterator interface: */
 
+void bch2_dump_trans_updates(struct btree_trans *);
 void bch2_dump_trans_paths_updates(struct btree_trans *);
 void __bch2_trans_init(struct btree_trans *, struct bch_fs *,
 		       unsigned, size_t, const char *);

fs/bcachefs/buckets.c

@@ -510,11 +510,16 @@ static int bch2_mark_alloc(struct btree_trans *trans,
 	struct bch_fs *c = trans->c;
 	struct bkey_alloc_unpacked old_u = bch2_alloc_unpack(old);
 	struct bkey_alloc_unpacked new_u = bch2_alloc_unpack(new);
-	struct bch_dev *ca;
+	struct bch_dev *ca = bch_dev_bkey_exists(c, new_u.dev);
 	struct bucket *g;
 	struct bucket_mark old_m, m;
 	int ret = 0;
 
+	if (bch2_trans_inconsistent_on(new_u.bucket < ca->mi.first_bucket ||
+				       new_u.bucket >= ca->mi.nbuckets, trans,
+				       "alloc key outside range of device's buckets"))
+		return -EIO;
+
 	/*
 	 * alloc btree is read in by bch2_alloc_read, not gc:
 	 */
@@ -554,11 +559,6 @@ static int bch2_mark_alloc(struct btree_trans *trans,
 		}
 	}
 
-	ca = bch_dev_bkey_exists(c, new_u.dev);
-
-	if (new_u.bucket >= ca->mi.nbuckets)
-		return 0;
-
 	percpu_down_read(&c->mark_lock);
 	if (!gc && new_u.gen != old_u.gen)
 		*bucket_gen(ca, new_u.bucket) = new_u.gen;
@@ -1466,7 +1466,6 @@ static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
 			struct extent_ptr_decoded p,
 			s64 sectors, enum bch_data_type data_type)
 {
-	struct bch_fs *c = trans->c;
 	struct btree_iter iter;
 	struct bkey_s_c k;
 	struct bkey_i_stripe *s;
@@ -1482,16 +1481,15 @@ static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
 		goto err;
 
 	if (k.k->type != KEY_TYPE_stripe) {
-		bch2_fs_inconsistent(c,
+		bch2_trans_inconsistent(trans,
 			"pointer to nonexistent stripe %llu",
 			(u64) p.ec.idx);
-		bch2_inconsistent_error(c);
 		ret = -EIO;
 		goto err;
 	}
 
 	if (!bch2_ptr_matches_stripe(bkey_s_c_to_stripe(k).v, p)) {
-		bch2_fs_inconsistent(c,
+		bch2_trans_inconsistent(trans,
 			"stripe pointer doesn't match stripe %llu",
 			(u64) p.ec.idx);
 		ret = -EIO;
@@ -1605,8 +1603,8 @@ static int bch2_trans_mark_stripe_bucket(struct btree_trans *trans,
 		goto err;
 
 	if (!deleting) {
-		if (bch2_fs_inconsistent_on(u.stripe ||
-					    u.stripe_redundancy, c,
+		if (bch2_trans_inconsistent_on(u.stripe ||
+					    u.stripe_redundancy, trans,
 				"bucket %llu:%llu gen %u data type %s dirty_sectors %u: multiple stripes using same bucket (%u, %llu)",
 				iter.pos.inode, iter.pos.offset, u.gen,
 				bch2_data_types[u.data_type],
@@ -1616,7 +1614,7 @@ static int bch2_trans_mark_stripe_bucket(struct btree_trans *trans,
 			goto err;
 		}
 
-		if (bch2_fs_inconsistent_on(data_type && u.dirty_sectors, c,
+		if (bch2_trans_inconsistent_on(data_type && u.dirty_sectors, trans,
 				"bucket %llu:%llu gen %u data type %s dirty_sectors %u: data already in stripe bucket %llu",
 				iter.pos.inode, iter.pos.offset, u.gen,
 				bch2_data_types[u.data_type],
@@ -1629,8 +1627,8 @@ static int bch2_trans_mark_stripe_bucket(struct btree_trans *trans,
 		u.stripe		= s.k->p.offset;
 		u.stripe_redundancy	= s.v->nr_redundant;
 	} else {
-		if (bch2_fs_inconsistent_on(u.stripe != s.k->p.offset ||
-					    u.stripe_redundancy != s.v->nr_redundant, c,
+		if (bch2_trans_inconsistent_on(u.stripe != s.k->p.offset ||
+					    u.stripe_redundancy != s.v->nr_redundant, trans,
 				"bucket %llu:%llu gen %u: not marked as stripe when deleting stripe %llu (got %u)",
 				iter.pos.inode, iter.pos.offset, u.gen,
 				s.k->p.offset, u.stripe)) {
@@ -1791,7 +1789,7 @@ static int __bch2_trans_mark_reflink_p(struct btree_trans *trans,
 	refcount = bkey_refcount(n);
 	if (!refcount) {
 		bch2_bkey_val_to_text(&buf, c, p.s_c);
-		bch2_fs_inconsistent(c,
+		bch2_trans_inconsistent(trans,
 			"nonexistent indirect extent at %llu while marking\n  %s",
 			*idx, buf.buf);
 		ret = -EIO;
@@ -1800,7 +1798,7 @@ static int __bch2_trans_mark_reflink_p(struct btree_trans *trans,
 
 	if (!*refcount && (flags & BTREE_TRIGGER_OVERWRITE)) {
 		bch2_bkey_val_to_text(&buf, c, p.s_c);
-		bch2_fs_inconsistent(c,
+		bch2_trans_inconsistent(trans,
 			"indirect extent refcount underflow at %llu while marking\n  %s",
 			*idx, buf.buf);
 		ret = -EIO;

fs/bcachefs/error.h

@@ -66,6 +66,26 @@ do {									\
 	_ret;								\
 })
 
+/*
+ * When a transaction update discovers or is causing a fs inconsistency, it's
+ * helpful to also dump the pending updates:
+ */
+#define bch2_trans_inconsistent(trans, ...)				\
+({									\
+	bch_err(trans->c, __VA_ARGS__);					\
+	bch2_inconsistent_error(trans->c);				\
+	bch2_dump_trans_updates(trans);					\
+})
+
+#define bch2_trans_inconsistent_on(cond, trans, ...)			\
+({									\
+	bool _ret = unlikely(!!(cond));					\
+									\
+	if (_ret)							\
+		bch2_trans_inconsistent(trans, __VA_ARGS__);		\
+	_ret;								\
+})
+
 /*
  * Fsck errors: inconsistency errors we detect at mount time, and should ideally
  * be able to repair: