bcachefs: Allocate fs_usage in do_btree_insert_at()

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

fs/bcachefs/bcachefs.h

@@ -635,7 +635,10 @@ struct bch_fs {
 	struct percpu_rw_semaphore	mark_lock;
 
 	struct bch_fs_usage __percpu	*usage[2];
-	struct bch_fs_usage __percpu	*usage_scratch;
+
+	/* single element mempool: */
+	struct mutex		usage_scratch_lock;
+	struct bch_fs_usage	*usage_scratch;
 
 	/*
 	 * When we invalidate buckets, we use both the priority and the amount

fs/bcachefs/btree_update_interior.c

@@ -1076,8 +1076,7 @@ static void bch2_btree_set_root_inmem(struct btree_update *as, struct btree *b)
 
 	mutex_lock(&c->btree_interior_update_lock);
 	percpu_down_read(&c->mark_lock);
-	preempt_disable();
-	fs_usage = bch2_fs_usage_get_scratch(c);
+	fs_usage = bch2_fs_usage_scratch_get(c);
 
 	bch2_mark_key_locked(c, bkey_i_to_s_c(&b->key),
 		      true, 0,
@@ -1090,7 +1089,7 @@ static void bch2_btree_set_root_inmem(struct btree_update *as, struct btree *b)
 					   fs_usage);
 	bch2_fs_usage_apply(c, fs_usage, &as->reserve->disk_res);
 
-	preempt_enable();
+	bch2_fs_usage_scratch_put(c, fs_usage);
 	percpu_up_read(&c->mark_lock);
 	mutex_unlock(&c->btree_interior_update_lock);
 }
@@ -1171,8 +1170,7 @@ static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b
 
 	mutex_lock(&c->btree_interior_update_lock);
 	percpu_down_read(&c->mark_lock);
-	preempt_disable();
-	fs_usage = bch2_fs_usage_get_scratch(c);
+	fs_usage = bch2_fs_usage_scratch_get(c);
 
 	bch2_mark_key_locked(c, bkey_i_to_s_c(insert),
 			     true, 0,
@@ -1193,7 +1191,7 @@ static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b
 
 	bch2_fs_usage_apply(c, fs_usage, &as->reserve->disk_res);
 
-	preempt_enable();
+	bch2_fs_usage_scratch_put(c, fs_usage);
 	percpu_up_read(&c->mark_lock);
 	mutex_unlock(&c->btree_interior_update_lock);
 
@@ -1987,7 +1985,7 @@ static void __bch2_btree_node_update_key(struct bch_fs *c,
 
 		mutex_lock(&c->btree_interior_update_lock);
 		percpu_down_read(&c->mark_lock);
-		fs_usage = bch2_fs_usage_get_scratch(c);
+		fs_usage = bch2_fs_usage_scratch_get(c);
 
 		bch2_mark_key_locked(c, bkey_i_to_s_c(&new_key->k_i),
 			      true, 0,
@@ -1998,6 +1996,7 @@ static void __bch2_btree_node_update_key(struct bch_fs *c,
 					   fs_usage);
 		bch2_fs_usage_apply(c, fs_usage, &as->reserve->disk_res);
 
+		bch2_fs_usage_scratch_put(c, fs_usage);
 		percpu_up_read(&c->mark_lock);
 		mutex_unlock(&c->btree_interior_update_lock);
 

fs/bcachefs/btree_update_leaf.c

@@ -269,8 +269,6 @@ static void btree_insert_key_leaf(struct btree_trans *trans,
 	int old_live_u64s = b->nr.live_u64s;
 	int live_u64s_added, u64s_added;
 
-	bch2_mark_update(trans, insert);
-
 	if (!btree_node_is_extents(b))
 		bch2_insert_fixup_key(trans, insert);
 	else
@@ -499,11 +497,6 @@ btree_key_can_insert(struct btree_trans *trans,
 	if (unlikely(btree_node_fake(b)))
 		return BTREE_INSERT_BTREE_NODE_FULL;
 
-	if (!bch2_bkey_replicas_marked(c,
-			bkey_i_to_s_c(insert->k),
-			true))
-		return BTREE_INSERT_NEED_MARK_REPLICAS;
-
 	ret = !btree_node_is_extents(b)
 		? BTREE_INSERT_OK
 		: bch2_extent_can_insert(trans, insert, u64s);
@@ -555,6 +548,7 @@ static inline int do_btree_insert_at(struct btree_trans *trans,
 				     struct btree_insert_entry **stopped_at)
 {
 	struct bch_fs *c = trans->c;
+	struct bch_fs_usage *fs_usage = NULL;
 	struct btree_insert_entry *i;
 	struct btree_iter *linked;
 	int ret;
@@ -562,12 +556,29 @@ static inline int do_btree_insert_at(struct btree_trans *trans,
 	trans_for_each_update_iter(trans, i)
 		BUG_ON(i->iter->uptodate >= BTREE_ITER_NEED_RELOCK);
 
+	trans_for_each_update_iter(trans, i) {
+		if (i->deferred ||
+		    !btree_node_type_needs_gc(i->iter->btree_id))
+			continue;
+
+		if (!fs_usage) {
+			percpu_down_read(&c->mark_lock);
+			fs_usage = bch2_fs_usage_scratch_get(c);
+		}
+
+		if (!bch2_bkey_replicas_marked_locked(c,
+				bkey_i_to_s_c(i->k), true)) {
+			ret = BTREE_INSERT_NEED_MARK_REPLICAS;
+			goto out;
+		}
+	}
+
 	btree_trans_lock_write(c, trans);
 
 	if (race_fault()) {
 		ret = -EINTR;
 		trans_restart(" (race)");
-		goto out;
+		goto out_unlock;
 	}
 
 	/*
@@ -577,7 +588,7 @@ static inline int do_btree_insert_at(struct btree_trans *trans,
 	 */
 	ret = btree_trans_check_can_insert(trans, stopped_at);
 	if (ret)
-		goto out;
+		goto out_unlock;
 
 	/*
 	 * Don't get journal reservation until after we know insert will
@@ -585,7 +596,7 @@ static inline int do_btree_insert_at(struct btree_trans *trans,
 	 */
 	ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_NONBLOCK);
 	if (ret)
-		goto out;
+		goto out_unlock;
 
 	if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 		if (journal_seq_verify(c))
@@ -610,14 +621,25 @@ static inline int do_btree_insert_at(struct btree_trans *trans,
 		}
 	}
 
+	trans_for_each_update_iter(trans, i)
+		bch2_mark_update(trans, i, fs_usage);
+	if (fs_usage)
+		bch2_trans_fs_usage_apply(trans, fs_usage);
+
 	trans_for_each_update(trans, i)
 		do_btree_insert_one(trans, i);
-out:
+out_unlock:
 	BUG_ON(ret &&
 	       (trans->flags & BTREE_INSERT_JOURNAL_RESERVED) &&
 	       trans->journal_res.ref);
 
 	btree_trans_unlock_write(trans);
+out:
+	if (fs_usage) {
+		bch2_fs_usage_scratch_put(c, fs_usage);
+		percpu_up_read(&c->mark_lock);
+	}
+
 	bch2_journal_res_put(&c->journal, &trans->journal_res);
 
 	return ret;

fs/bcachefs/buckets.c

@@ -144,6 +144,37 @@ void bch2_fs_usage_initialize(struct bch_fs *c)
 	percpu_up_write(&c->mark_lock);
 }
 
+void bch2_fs_usage_scratch_put(struct bch_fs *c, struct bch_fs_usage *fs_usage)
+{
+	if (fs_usage == c->usage_scratch)
+		mutex_unlock(&c->usage_scratch_lock);
+	else
+		kfree(fs_usage);
+}
+
+struct bch_fs_usage *bch2_fs_usage_scratch_get(struct bch_fs *c)
+{
+	struct bch_fs_usage *ret;
+	unsigned bytes = fs_usage_u64s(c) * sizeof(u64);
+
+	ret = kzalloc(bytes, GFP_NOWAIT);
+	if (ret)
+		return ret;
+
+	if (mutex_trylock(&c->usage_scratch_lock))
+		goto out_pool;
+
+	ret = kzalloc(bytes, GFP_NOFS);
+	if (ret)
+		return ret;
+
+	mutex_lock(&c->usage_scratch_lock);
+out_pool:
+	ret = c->usage_scratch;
+	memset(ret, 0, bytes);
+	return ret;
+}
+
 struct bch_dev_usage bch2_dev_usage_read(struct bch_fs *c, struct bch_dev *ca)
 {
 	struct bch_dev_usage ret;
@@ -906,31 +937,39 @@ static int __bch2_mark_key(struct bch_fs *c, struct bkey_s_c k,
 			   unsigned journal_seq, unsigned flags,
 			   bool gc)
 {
+	int ret = 0;
+
+	preempt_disable();
+
 	if (!fs_usage || gc)
 		fs_usage = this_cpu_ptr(c->usage[gc]);
 
 	switch (k.k->type) {
 	case KEY_TYPE_alloc:
-		return bch2_mark_alloc(c, k, inserting,
+		ret = bch2_mark_alloc(c, k, inserting,
 				fs_usage, journal_seq, flags, gc);
+		break;
 	case KEY_TYPE_btree_ptr:
-		return bch2_mark_extent(c, k, inserting
+		ret = bch2_mark_extent(c, k, inserting
 				?  c->opts.btree_node_size
 				: -c->opts.btree_node_size,
 				BCH_DATA_BTREE,
 				fs_usage, journal_seq, flags, gc);
+		break;
 	case KEY_TYPE_extent:
-		return bch2_mark_extent(c, k, sectors, BCH_DATA_USER,
+		ret = bch2_mark_extent(c, k, sectors, BCH_DATA_USER,
 				fs_usage, journal_seq, flags, gc);
+		break;
 	case KEY_TYPE_stripe:
-		return bch2_mark_stripe(c, k, inserting,
+		ret = bch2_mark_stripe(c, k, inserting,
 				fs_usage, journal_seq, flags, gc);
+		break;
 	case KEY_TYPE_inode:
 		if (inserting)
 			fs_usage->nr_inodes++;
 		else
 			fs_usage->nr_inodes--;
-		return 0;
+		break;
 	case KEY_TYPE_reservation: {
 		unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
 
@@ -940,11 +979,13 @@ static int __bch2_mark_key(struct bch_fs *c, struct bkey_s_c k,
 
 		fs_usage->reserved				+= sectors;
 		fs_usage->persistent_reserved[replicas - 1]	+= sectors;
-		return 0;
+		break;
 	}
-	default:
-		return 0;
 	}
+
+	preempt_enable();
+
+	return ret;
 }
 
 int bch2_mark_key_locked(struct bch_fs *c,
@@ -976,25 +1017,19 @@ int bch2_mark_key(struct bch_fs *c, struct bkey_s_c k,
 }
 
 void bch2_mark_update(struct btree_trans *trans,
-		      struct btree_insert_entry *insert)
+		      struct btree_insert_entry *insert,
+		      struct bch_fs_usage *fs_usage)
 {
 	struct bch_fs		*c = trans->c;
 	struct btree_iter	*iter = insert->iter;
 	struct btree		*b = iter->l[0].b;
 	struct btree_node_iter	node_iter = iter->l[0].iter;
-	struct bch_fs_usage	*fs_usage;
 	struct gc_pos		pos = gc_pos_btree_node(b);
 	struct bkey_packed	*_k;
-	u64 disk_res_sectors = trans->disk_res ? trans->disk_res->sectors : 0;
-	static int warned_disk_usage = 0;
 
 	if (!btree_node_type_needs_gc(iter->btree_id))
 		return;
 
-	percpu_down_read(&c->mark_lock);
-	preempt_disable();
-	fs_usage = bch2_fs_usage_get_scratch(c);
-
 	if (!(trans->flags & BTREE_INSERT_NOMARK))
 		bch2_mark_key_locked(c, bkey_i_to_s_c(insert->k), true,
 			bpos_min(insert->k->k.p, b->key.k.p).offset -
@@ -1047,16 +1082,32 @@ void bch2_mark_update(struct btree_trans *trans,
 
 		bch2_btree_node_iter_advance(&node_iter, b);
 	}
+}
 
-	if (bch2_fs_usage_apply(c, fs_usage, trans->disk_res) &&
-	    !warned_disk_usage &&
-	    !xchg(&warned_disk_usage, 1)) {
+void bch2_trans_fs_usage_apply(struct btree_trans *trans,
+			       struct bch_fs_usage *fs_usage)
+{
+	struct bch_fs *c = trans->c;
+	struct btree_insert_entry *i;
+	static int warned_disk_usage = 0;
+	u64 disk_res_sectors = trans->disk_res ? trans->disk_res->sectors : 0;
 	char buf[200];
 
+	if (!bch2_fs_usage_apply(c, fs_usage, trans->disk_res) ||
+	    warned_disk_usage ||
+	    xchg(&warned_disk_usage, 1))
+		return;
+
 	pr_err("disk usage increased more than %llu sectors reserved", disk_res_sectors);
 
+	trans_for_each_update_iter(trans, i) {
+		struct btree_iter	*iter = i->iter;
+		struct btree		*b = iter->l[0].b;
+		struct btree_node_iter	node_iter = iter->l[0].iter;
+		struct bkey_packed	*_k;
+
 		pr_err("while inserting");
-		bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(insert->k));
+		bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(i->k));
 		pr_err("%s", buf);
 		pr_err("overlapping with");
 
@@ -1069,8 +1120,8 @@ void bch2_mark_update(struct btree_trans *trans,
 			k = bkey_disassemble(b, _k, &unpacked);
 
 			if (btree_node_is_extents(b)
-			    ? bkey_cmp(insert->k->k.p, bkey_start_pos(k.k)) <= 0
-			    : bkey_cmp(insert->k->k.p, k.k->p))
+			    ? bkey_cmp(i->k->k.p, bkey_start_pos(k.k)) <= 0
+			    : bkey_cmp(i->k->k.p, k.k->p))
 				break;
 
 			bch2_bkey_val_to_text(&PBUF(buf), c, k);
@@ -1079,9 +1130,6 @@ void bch2_mark_update(struct btree_trans *trans,
 			bch2_btree_node_iter_advance(&node_iter, b);
 		}
 	}
-
-	preempt_enable();
-	percpu_up_read(&c->mark_lock);
 }
 
 /* Disk reservations: */

fs/bcachefs/buckets.h

@@ -219,13 +219,8 @@ static inline unsigned fs_usage_u64s(struct bch_fs *c)
 		READ_ONCE(c->replicas.nr);
 }
 
-static inline struct bch_fs_usage *bch2_fs_usage_get_scratch(struct bch_fs *c)
-{
-	struct bch_fs_usage *ret = this_cpu_ptr(c->usage_scratch);
-
-	memset(ret, 0, fs_usage_u64s(c) * sizeof(u64));
-	return ret;
-}
+void bch2_fs_usage_scratch_put(struct bch_fs *, struct bch_fs_usage *);
+struct bch_fs_usage *bch2_fs_usage_scratch_get(struct bch_fs *);
 
 struct bch_fs_usage *bch2_fs_usage_read(struct bch_fs *);
 
@@ -256,10 +251,13 @@ int bch2_mark_key_locked(struct bch_fs *, struct bkey_s_c,
 int bch2_mark_key(struct bch_fs *, struct bkey_s_c,
 		  bool, s64, struct gc_pos,
 		  struct bch_fs_usage *, u64, unsigned);
-void bch2_mark_update(struct btree_trans *, struct btree_insert_entry *);
 int bch2_fs_usage_apply(struct bch_fs *, struct bch_fs_usage *,
 			struct disk_reservation *);
 
+void bch2_mark_update(struct btree_trans *, struct btree_insert_entry *,
+		      struct bch_fs_usage *);
+void bch2_trans_fs_usage_apply(struct btree_trans *, struct bch_fs_usage *);
+
 /* disk reservations: */
 
 void __bch2_disk_reservation_put(struct bch_fs *, struct disk_reservation *);

fs/bcachefs/extents.c

@@ -1190,11 +1190,12 @@ void bch2_insert_fixup_extent(struct btree_trans *trans,
 
 		if (s.deleting)
 			tmp.k.k.type = KEY_TYPE_discard;
-
+#if 0
+		/* disabled due to lock recursion - mark_lock: */
 		if (debug_check_bkeys(c))
 			bch2_bkey_debugcheck(c, iter->l[0].b,
 					     bkey_i_to_s_c(&tmp.k));
-
+#endif
 		EBUG_ON(bkey_deleted(&tmp.k.k) || !tmp.k.k.size);
 
 		extent_bset_insert(c, iter, &tmp.k);

fs/bcachefs/replicas.c

@@ -207,22 +207,29 @@ static bool __replicas_has_entry(struct bch_replicas_cpu *r,
 	return __replicas_entry_idx(r, search) >= 0;
 }
 
-bool bch2_replicas_marked(struct bch_fs *c,
+static bool bch2_replicas_marked_locked(struct bch_fs *c,
 			  struct bch_replicas_entry *search,
 			  bool check_gc_replicas)
 {
-	bool marked;
-
 	if (!search->nr_devs)
 		return true;
 
 	verify_replicas_entry_sorted(search);
 
-	percpu_down_read(&c->mark_lock);
-	marked = __replicas_has_entry(&c->replicas, search) &&
+	return __replicas_has_entry(&c->replicas, search) &&
 		(!check_gc_replicas ||
 		 likely((!c->replicas_gc.entries)) ||
 		 __replicas_has_entry(&c->replicas_gc, search));
+}
+
+bool bch2_replicas_marked(struct bch_fs *c,
+			  struct bch_replicas_entry *search,
+			  bool check_gc_replicas)
+{
+	bool marked;
+
+	percpu_down_read(&c->mark_lock);
+	marked = bch2_replicas_marked_locked(c, search, check_gc_replicas);
 	percpu_up_read(&c->mark_lock);
 
 	return marked;
@@ -263,7 +270,7 @@ static int replicas_table_update(struct bch_fs *c,
 				 struct bch_replicas_cpu *new_r)
 {
 	struct bch_fs_usage __percpu *new_usage[2] = { NULL, NULL };
-	struct bch_fs_usage __percpu *new_scratch = NULL;
+	struct bch_fs_usage *new_scratch = NULL;
 	unsigned bytes = sizeof(struct bch_fs_usage) +
 		sizeof(u64) * new_r->nr;
 	int ret = -ENOMEM;
@@ -273,8 +280,7 @@ static int replicas_table_update(struct bch_fs *c,
 	    (c->usage[1] &&
 	     !(new_usage[1] = __alloc_percpu_gfp(bytes, sizeof(u64),
 						 GFP_NOIO))) ||
-	    !(new_scratch  = __alloc_percpu_gfp(bytes, sizeof(u64),
-						GFP_NOIO)))
+	    !(new_scratch  = kmalloc(bytes, GFP_NOIO)))
 		goto err;
 
 	if (c->usage[0])
@@ -290,7 +296,7 @@ static int replicas_table_update(struct bch_fs *c,
 	swap(c->replicas,	*new_r);
 	ret = 0;
 err:
-	free_percpu(new_scratch);
+	kfree(new_scratch);
 	free_percpu(new_usage[1]);
 	free_percpu(new_usage[0]);
 	return ret;
@@ -390,7 +396,7 @@ int bch2_mark_replicas(struct bch_fs *c,
 		: bch2_mark_replicas_slowpath(c, r);
 }
 
-bool bch2_bkey_replicas_marked(struct bch_fs *c,
+bool bch2_bkey_replicas_marked_locked(struct bch_fs *c,
 				      struct bkey_s_c k,
 				      bool check_gc_replicas)
 {
@@ -401,13 +407,27 @@ bool bch2_bkey_replicas_marked(struct bch_fs *c,
 	for (i = 0; i < cached.nr; i++) {
 		bch2_replicas_entry_cached(&search.e, cached.devs[i]);
 
-		if (!bch2_replicas_marked(c, &search.e, check_gc_replicas))
+		if (!bch2_replicas_marked_locked(c, &search.e,
+						 check_gc_replicas))
 			return false;
 	}
 
 	bkey_to_replicas(&search.e, k);
 
-	return bch2_replicas_marked(c, &search.e, check_gc_replicas);
+	return bch2_replicas_marked_locked(c, &search.e, check_gc_replicas);
+}
+
+bool bch2_bkey_replicas_marked(struct bch_fs *c,
+			       struct bkey_s_c k,
+			       bool check_gc_replicas)
+{
+	bool marked;
+
+	percpu_down_read(&c->mark_lock);
+	marked = bch2_bkey_replicas_marked_locked(c, k, check_gc_replicas);
+	percpu_up_read(&c->mark_lock);
+
+	return marked;
 }
 
 int bch2_mark_bkey_replicas(struct bch_fs *c, struct bkey_s_c k)

fs/bcachefs/replicas.h

@@ -26,6 +26,8 @@ bool bch2_replicas_marked(struct bch_fs *,
 int bch2_mark_replicas(struct bch_fs *,
 		       struct bch_replicas_entry *);
 
+bool bch2_bkey_replicas_marked_locked(struct bch_fs *,
+				      struct bkey_s_c, bool);
 bool bch2_bkey_replicas_marked(struct bch_fs *,
 			       struct bkey_s_c, bool);
 int bch2_mark_bkey_replicas(struct bch_fs *, struct bkey_s_c);

fs/bcachefs/super.c

@@ -404,7 +404,7 @@ static void bch2_fs_free(struct bch_fs *c)
 	bch2_io_clock_exit(&c->io_clock[READ]);
 	bch2_fs_compress_exit(c);
 	percpu_free_rwsem(&c->mark_lock);
-	free_percpu(c->usage_scratch);
+	kfree(c->usage_scratch);
 	free_percpu(c->usage[0]);
 	free_percpu(c->pcpu);
 	mempool_exit(&c->btree_iters_pool);
@@ -572,6 +572,8 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 	mutex_init(&c->btree_reserve_cache_lock);
 	mutex_init(&c->btree_interior_update_lock);
 
+	mutex_init(&c->usage_scratch_lock);
+
 	mutex_init(&c->bio_bounce_pages_lock);
 
 	bio_list_init(&c->btree_write_error_list);