bcachefs: Go rw lazily

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

fs/bcachefs/alloc_background.c

@@ -345,6 +345,7 @@ int bch2_alloc_replay_key(struct bch_fs *c, struct bkey_i *k)
 
 	ret = bch2_trans_commit(&trans, NULL, NULL,
 				BTREE_INSERT_NOFAIL|
+				BTREE_INSERT_LAZY_RW|
 				BTREE_INSERT_JOURNAL_REPLAY|
 				BTREE_INSERT_NOMARK);
 err:
@@ -1626,7 +1627,7 @@ static bool bch2_fs_allocator_start_fast(struct bch_fs *c)
 	return ret;
 }
 
-static int __bch2_fs_allocator_start(struct bch_fs *c)
+int bch2_fs_allocator_start(struct bch_fs *c)
 {
 	struct bch_dev *ca;
 	unsigned dev_iter;
@@ -1635,6 +1636,10 @@ static int __bch2_fs_allocator_start(struct bch_fs *c)
 	long bu;
 	int ret = 0;
 
+	if (!test_alloc_startup(c) &&
+	    bch2_fs_allocator_start_fast(c))
+		return 0;
+
 	pr_debug("not enough empty buckets; scanning for reclaimable buckets");
 
 	/*
@@ -1709,31 +1714,6 @@ static int __bch2_fs_allocator_start(struct bch_fs *c)
 	return ret;
 }
 
-int bch2_fs_allocator_start(struct bch_fs *c)
-{
-	struct bch_dev *ca;
-	unsigned i;
-	int ret;
-
-	ret = bch2_fs_allocator_start_fast(c) ? 0 :
-		__bch2_fs_allocator_start(c);
-	if (ret)
-		return ret;
-
-	set_bit(BCH_FS_ALLOCATOR_STARTED, &c->flags);
-
-	for_each_rw_member(ca, c, i) {
-		ret = bch2_dev_allocator_start(ca);
-		if (ret) {
-			percpu_ref_put(&ca->io_ref);
-			return ret;
-		}
-	}
-
-	set_bit(BCH_FS_ALLOCATOR_RUNNING, &c->flags);
-	return 0;
-}
-
 void bch2_fs_allocator_background_init(struct bch_fs *c)
 {
 	spin_lock_init(&c->freelist_lock);

fs/bcachefs/bcachefs.h

@@ -486,6 +486,7 @@ enum {
 	BCH_FS_INITIAL_GC_DONE,
 	BCH_FS_FSCK_DONE,
 	BCH_FS_STARTED,
+	BCH_FS_RW,
 
 	/* shutdown: */
 	BCH_FS_EMERGENCY_RO,
@@ -510,13 +511,6 @@ struct btree_debug {
 	struct dentry		*failed;
 };
 
-enum bch_fs_state {
-	BCH_FS_STARTING		= 0,
-	BCH_FS_STOPPING,
-	BCH_FS_RO,
-	BCH_FS_RW,
-};
-
 struct bch_fs_pcpu {
 	u64			sectors_available;
 };
@@ -538,7 +532,6 @@ struct bch_fs {
 
 	/* ro/rw, add/remove devices: */
 	struct mutex		state_lock;
-	enum bch_fs_state	state;
 
 	/* Counts outstanding writes, for clean transition to read-only */
 	struct percpu_ref	writes;
@@ -800,11 +793,6 @@ static inline void bch2_set_ra_pages(struct bch_fs *c, unsigned ra_pages)
 #endif
 }
 
-static inline bool bch2_fs_running(struct bch_fs *c)
-{
-	return c->state == BCH_FS_RO || c->state == BCH_FS_RW;
-}
-
 static inline unsigned bucket_bytes(const struct bch_dev *ca)
 {
 	return ca->mi.bucket_size << 9;

fs/bcachefs/btree_update.h

@@ -38,6 +38,7 @@ enum {
 	__BTREE_INSERT_NOUNLOCK,
 	__BTREE_INSERT_NOFAIL,
 	__BTREE_INSERT_NOCHECK_RW,
+	__BTREE_INSERT_LAZY_RW,
 	__BTREE_INSERT_USE_RESERVE,
 	__BTREE_INSERT_USE_ALLOC_RESERVE,
 	__BTREE_INSERT_JOURNAL_REPLAY,
@@ -64,6 +65,7 @@ enum {
 #define BTREE_INSERT_NOFAIL		(1 << __BTREE_INSERT_NOFAIL)
 
 #define BTREE_INSERT_NOCHECK_RW		(1 << __BTREE_INSERT_NOCHECK_RW)
+#define BTREE_INSERT_LAZY_RW		(1 << __BTREE_INSERT_LAZY_RW)
 
 /* for copygc, or when merging btree nodes */
 #define BTREE_INSERT_USE_RESERVE	(1 << __BTREE_INSERT_USE_RESERVE)

fs/bcachefs/btree_update_leaf.c

@@ -849,8 +849,23 @@ int bch2_trans_commit(struct btree_trans *trans,
 		btree_insert_entry_checks(trans, i);
 
 	if (unlikely(!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
-		     !percpu_ref_tryget(&c->writes)))
-		return -EROFS;
+		     !percpu_ref_tryget(&c->writes))) {
+		if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
+			return -EROFS;
+
+		btree_trans_unlock(trans);
+
+		ret = bch2_fs_read_write_early(c);
+		if (ret)
+			return ret;
+
+		percpu_ref_get(&c->writes);
+
+		if (!btree_trans_relock(trans)) {
+			ret = -EINTR;
+			goto err;
+		}
+	}
 retry:
 	ret = bch2_trans_journal_preres_get(trans);
 	if (ret)

fs/bcachefs/fs.c

@@ -1616,7 +1616,7 @@ static struct bch_fs *__bch2_open_as_blockdevs(const char *dev_name, char * cons
 
 	mutex_lock(&c->state_lock);
 
-	if (!bch2_fs_running(c)) {
+	if (!test_bit(BCH_FS_STARTED, &c->flags)) {
 		mutex_unlock(&c->state_lock);
 		closure_put(&c->cl);
 		pr_err("err mounting %s: incomplete filesystem", dev_name);
@@ -1672,8 +1672,6 @@ static int bch2_remount(struct super_block *sb, int *flags, char *data)
 		return ret;
 
 	if (opts.read_only != c->opts.read_only) {
-		const char *err = NULL;
-
 		mutex_lock(&c->state_lock);
 
 		if (opts.read_only) {
@@ -1681,9 +1679,9 @@ static int bch2_remount(struct super_block *sb, int *flags, char *data)
 
 			sb->s_flags |= SB_RDONLY;
 		} else {
-			err = bch2_fs_read_write(c);
-			if (err) {
-				bch_err(c, "error going rw: %s", err);
+			ret = bch2_fs_read_write(c);
+			if (ret) {
+				bch_err(c, "error going rw: %i", ret);
 				return -EINVAL;
 			}
 

fs/bcachefs/fsck.c

@@ -174,7 +174,8 @@ static int hash_redo_key(const struct bch_hash_desc desc,
 	bch2_hash_set(trans, desc, &h->info, k_iter->pos.inode,
 		      tmp, BCH_HASH_SET_MUST_CREATE);
 	ret = bch2_trans_commit(trans, NULL, NULL,
-				BTREE_INSERT_NOFAIL);
+				BTREE_INSERT_NOFAIL|
+				BTREE_INSERT_LAZY_RW);
 err:
 	kfree(tmp);
 	return ret;
@@ -204,7 +205,8 @@ static int fsck_hash_delete_at(const struct bch_hash_desc desc,
 	ret   = bch2_hash_delete_at(&trans, desc, info, iter) ?:
 		bch2_trans_commit(&trans, NULL, NULL,
 				  BTREE_INSERT_ATOMIC|
-				  BTREE_INSERT_NOFAIL);
+				  BTREE_INSERT_NOFAIL|
+				  BTREE_INSERT_LAZY_RW);
 err:
 	if (ret == -EINTR)
 		goto retry;
@@ -365,7 +367,9 @@ static int check_dirent_hash(struct btree_trans *trans, struct hash_check *h,
 		     buf, strlen(buf), d->v.d_name, len)) {
 		bch2_trans_update(trans, BTREE_INSERT_ENTRY(iter, &d->k_i));
 
-		ret = bch2_trans_commit(trans, NULL, NULL, BTREE_INSERT_NOFAIL);
+		ret = bch2_trans_commit(trans, NULL, NULL,
+					BTREE_INSERT_NOFAIL|
+					BTREE_INSERT_LAZY_RW);
 		if (ret)
 			goto err;
 
@@ -630,7 +634,8 @@ static int check_dirents(struct bch_fs *c)
 				BTREE_INSERT_ENTRY(iter, &n->k_i));
 
 			ret = bch2_trans_commit(&trans, NULL, NULL,
-						BTREE_INSERT_NOFAIL);
+						BTREE_INSERT_NOFAIL|
+						BTREE_INSERT_LAZY_RW);
 			kfree(n);
 			if (ret)
 				goto err;
@@ -1268,7 +1273,8 @@ static int check_inode(struct btree_trans *trans,
 		bch2_trans_update(trans, BTREE_INSERT_ENTRY(iter, &p.inode.k_i));
 
 		ret = bch2_trans_commit(trans, NULL, NULL,
-					BTREE_INSERT_NOFAIL);
+					BTREE_INSERT_NOFAIL|
+					BTREE_INSERT_LAZY_RW);
 		if (ret && ret != -EINTR)
 			bch_err(c, "error in fs gc: error %i "
 				"updating inode", ret);

fs/bcachefs/journal.c

@@ -1027,8 +1027,6 @@ void bch2_fs_journal_start(struct journal *j)
 	 * only have to go down with the next journal entry we write:
 	 */
 	bch2_journal_seq_blacklist_write(j);
-
-	queue_delayed_work(c->journal_reclaim_wq, &j->reclaim_work, 0);
 }
 
 /* init/exit: */

fs/bcachefs/journal_io.c

@@ -861,6 +861,7 @@ static int bch2_extent_replay_key(struct bch_fs *c, struct bkey_i *k)
 		ret = bch2_trans_commit(&trans, &disk_res, NULL,
 					BTREE_INSERT_ATOMIC|
 					BTREE_INSERT_NOFAIL|
+					BTREE_INSERT_LAZY_RW|
 					BTREE_INSERT_JOURNAL_REPLAY);
 	} while ((!ret || ret == -EINTR) &&
 		 bkey_cmp(k->k.p, iter->pos));
@@ -906,6 +907,7 @@ int bch2_journal_replay(struct bch_fs *c, struct list_head *list)
 				ret = bch2_btree_insert(c, entry->btree_id, k,
 						NULL, NULL,
 						BTREE_INSERT_NOFAIL|
+						BTREE_INSERT_LAZY_RW|
 						BTREE_INSERT_JOURNAL_REPLAY|
 						BTREE_INSERT_NOMARK);
 				break;

fs/bcachefs/recovery.c

@@ -119,8 +119,13 @@ static int verify_superblock_clean(struct bch_fs *c,
 	if (mustfix_fsck_err_on(j->seq != clean->journal_seq, c,
 			"superblock journal seq (%llu) doesn't match journal (%llu) after clean shutdown",
 			le64_to_cpu(clean->journal_seq),
-			le64_to_cpu(j->seq)))
-		bch2_fs_mark_clean(c, false);
+			le64_to_cpu(j->seq))) {
+		ret = bch2_fs_mark_dirty(c);
+		if (ret) {
+			bch_err(c, "error going rw");
+			return ret;
+		}
+	}
 
 	mustfix_fsck_err_on(j->read_clock != clean->read_clock, c,
 			"superblock read clock doesn't match journal after clean shutdown");
@@ -331,13 +336,6 @@ int bch2_fs_recovery(struct bch_fs *c)
 	if (c->opts.noreplay)
 		goto out;
 
-	/*
-	 * Mark dirty before journal replay, fsck:
-	 * XXX: after a clean shutdown, this could be done lazily only when fsck
-	 * finds an error
-	 */
-	bch2_fs_mark_clean(c, false);
-
 	/*
 	 * bch2_fs_journal_start() can't happen sooner, or btree_gc_finish()
 	 * will give spurious errors about oldest_gen > bucket_gen -
@@ -345,11 +343,6 @@ int bch2_fs_recovery(struct bch_fs *c)
 	 */
 	bch2_fs_journal_start(&c->journal);
 
-	err = "error starting allocator";
-	ret = bch2_fs_allocator_start(c);
-	if (ret)
-		goto err;
-
 	bch_verbose(c, "starting journal replay:");
 	err = "journal replay failed";
 	ret = bch2_journal_replay(c, &journal);
@@ -436,8 +429,8 @@ int bch2_fs_initialize(struct bch_fs *c)
 	bch2_fs_journal_start(&c->journal);
 	bch2_journal_set_replay_done(&c->journal);
 
-	err = "error starting allocator";
-	ret = bch2_fs_allocator_start(c);
+	err = "error going read write";
+	ret = bch2_fs_read_write_early(c);
 	if (ret)
 		goto err;
 

fs/bcachefs/super-io.c

@@ -886,7 +886,7 @@ void bch2_sb_clean_renumber(struct bch_sb_field_clean *clean, int write)
 		bch2_bkey_renumber(BKEY_TYPE_BTREE, bkey_to_packed(entry->start), write);
 }
 
-static void bch2_fs_mark_dirty(struct bch_fs *c)
+int bch2_fs_mark_dirty(struct bch_fs *c)
 {
 	mutex_lock(&c->sb_lock);
 	if (BCH_SB_CLEAN(c->disk_sb.sb) ||
@@ -896,6 +896,8 @@ static void bch2_fs_mark_dirty(struct bch_fs *c)
 		bch2_write_super(c);
 	}
 	mutex_unlock(&c->sb_lock);
+
+	return 0;
 }
 
 struct jset_entry *
@@ -997,17 +999,12 @@ bch2_journal_super_entries_add_common(struct bch_fs *c,
 	return entry;
 }
 
-void bch2_fs_mark_clean(struct bch_fs *c, bool clean)
+void bch2_fs_mark_clean(struct bch_fs *c)
 {
 	struct bch_sb_field_clean *sb_clean;
 	struct jset_entry *entry;
 	unsigned u64s;
 
-	if (!clean) {
-		bch2_fs_mark_dirty(c);
-		return;
-	}
-
 	mutex_lock(&c->sb_lock);
 	if (BCH_SB_CLEAN(c->disk_sb.sb))
 		goto out;

fs/bcachefs/super-io.h

@@ -141,7 +141,8 @@ bch2_journal_super_entries_add_common(struct bch_fs *,
 
 void bch2_sb_clean_renumber(struct bch_sb_field_clean *, int);
 
-void bch2_fs_mark_clean(struct bch_fs *, bool);
+int bch2_fs_mark_dirty(struct bch_fs *);
+void bch2_fs_mark_clean(struct bch_fs *);
 
 void bch2_sb_field_to_text(struct printbuf *, struct bch_sb *,
 			   struct bch_sb_field *);

fs/bcachefs/super.c

@@ -258,8 +258,10 @@ static void bch2_writes_disabled(struct percpu_ref *writes)
 
 void bch2_fs_read_only(struct bch_fs *c)
 {
-	if (c->state == BCH_FS_RO)
+	if (!test_bit(BCH_FS_RW, &c->flags)) {
+		cancel_delayed_work_sync(&c->journal.reclaim_work);
 		return;
+	}
 
 	BUG_ON(test_bit(BCH_FS_WRITE_DISABLE_COMPLETE, &c->flags));
 
@@ -301,10 +303,9 @@ void bch2_fs_read_only(struct bch_fs *c)
 	    !test_bit(BCH_FS_ERROR, &c->flags) &&
 	    !test_bit(BCH_FS_EMERGENCY_RO, &c->flags) &&
 	    test_bit(BCH_FS_STARTED, &c->flags))
-		bch2_fs_mark_clean(c, true);
+		bch2_fs_mark_clean(c);
 
-	if (c->state != BCH_FS_STOPPING)
-		c->state = BCH_FS_RO;
+	clear_bit(BCH_FS_RW, &c->flags);
 }
 
 static void bch2_fs_read_only_work(struct work_struct *work)
@@ -333,55 +334,106 @@ bool bch2_fs_emergency_read_only(struct bch_fs *c)
 	return ret;
 }
 
-const char *bch2_fs_read_write(struct bch_fs *c)
+static int bch2_fs_read_write_late(struct bch_fs *c)
 {
 	struct bch_dev *ca;
-	const char *err = NULL;
 	unsigned i;
+	int ret;
 
-	if (c->state == BCH_FS_RW)
-		return NULL;
+	ret = bch2_gc_thread_start(c);
+	if (ret) {
+		bch_err(c, "error starting gc thread");
+		return ret;
+	}
+
+	for_each_rw_member(ca, c, i) {
+		ret = bch2_copygc_start(c, ca);
+		if (ret) {
+			bch_err(c, "error starting copygc threads");
+			percpu_ref_put(&ca->io_ref);
+			return ret;
+		}
+	}
+
+	ret = bch2_rebalance_start(c);
+	if (ret) {
+		bch_err(c, "error starting rebalance thread");
+		return ret;
+	}
+
+	schedule_delayed_work(&c->pd_controllers_update, 5 * HZ);
+
+	return 0;
+}
+
+static int __bch2_fs_read_write(struct bch_fs *c, bool early)
+{
+	struct bch_dev *ca;
+	unsigned i;
+	int ret;
+
+	if (test_bit(BCH_FS_RW, &c->flags))
+		return 0;
 
-	bch2_fs_mark_clean(c, false);
+	ret = bch2_fs_mark_dirty(c);
+	if (ret)
+		goto err;
 
 	for_each_rw_member(ca, c, i)
 		bch2_dev_allocator_add(c, ca);
 	bch2_recalc_capacity(c);
 
-	err = "error starting allocator thread";
-	for_each_rw_member(ca, c, i)
-		if (bch2_dev_allocator_start(ca)) {
-			percpu_ref_put(&ca->io_ref);
+	if (!test_bit(BCH_FS_ALLOCATOR_STARTED, &c->flags)) {
+		ret = bch2_fs_allocator_start(c);
+		if (ret) {
+			bch_err(c, "error initializing allocator");
 			goto err;
 		}
 
-	set_bit(BCH_FS_ALLOCATOR_RUNNING, &c->flags);
-
-	err = "error starting btree GC thread";
-	if (bch2_gc_thread_start(c))
-		goto err;
+		set_bit(BCH_FS_ALLOCATOR_STARTED, &c->flags);
+	}
 
-	err = "error starting copygc thread";
-	for_each_rw_member(ca, c, i)
-		if (bch2_copygc_start(c, ca)) {
+	for_each_rw_member(ca, c, i) {
+		ret = bch2_dev_allocator_start(ca);
+		if (ret) {
+			bch_err(c, "error starting allocator threads");
 			percpu_ref_put(&ca->io_ref);
 			goto err;
 		}
+	}
 
-	err = "error starting rebalance thread";
-	if (bch2_rebalance_start(c))
-		goto err;
+	set_bit(BCH_FS_ALLOCATOR_RUNNING, &c->flags);
 
-	schedule_delayed_work(&c->pd_controllers_update, 5 * HZ);
+	if (!early) {
+		ret = bch2_fs_read_write_late(c);
+		if (ret)
+			goto err;
+	}
 
-	if (c->state != BCH_FS_STARTING)
-		percpu_ref_reinit(&c->writes);
+	percpu_ref_reinit(&c->writes);
+	set_bit(BCH_FS_RW, &c->flags);
 
-	c->state = BCH_FS_RW;
-	return NULL;
+	queue_delayed_work(c->journal_reclaim_wq,
+			   &c->journal.reclaim_work, 0);
+	return 0;
 err:
 	__bch2_fs_read_only(c);
-	return err;
+	return ret;
+}
+
+int bch2_fs_read_write(struct bch_fs *c)
+{
+	return __bch2_fs_read_write(c, false);
+}
+
+int bch2_fs_read_write_early(struct bch_fs *c)
+{
+	lockdep_assert_held(&c->state_lock);
+
+	if (c->opts.read_only)
+		return -EROFS;
+
+	return __bch2_fs_read_write(c, true);
 }
 
 /* Filesystem startup/shutdown: */
@@ -638,7 +690,8 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 				WQ_FREEZABLE|WQ_MEM_RECLAIM|WQ_HIGHPRI, 1)) ||
 	    !(c->journal_reclaim_wq = alloc_workqueue("bcache_journal",
 				WQ_FREEZABLE|WQ_MEM_RECLAIM|WQ_HIGHPRI, 1)) ||
-	    percpu_ref_init(&c->writes, bch2_writes_disabled, 0, GFP_KERNEL) ||
+	    percpu_ref_init(&c->writes, bch2_writes_disabled,
+			    PERCPU_REF_INIT_DEAD, GFP_KERNEL) ||
 	    mempool_init_kmalloc_pool(&c->btree_reserve_pool, 1,
 				      sizeof(struct btree_reserve)) ||
 	    mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
@@ -712,7 +765,7 @@ const char *bch2_fs_start(struct bch_fs *c)
 
 	mutex_lock(&c->state_lock);
 
-	BUG_ON(c->state != BCH_FS_STARTING);
+	BUG_ON(test_bit(BCH_FS_STARTED, &c->flags));
 
 	mutex_lock(&c->sb_lock);
 
@@ -746,9 +799,12 @@ const char *bch2_fs_start(struct bch_fs *c)
 	if (c->opts.read_only) {
 		bch2_fs_read_only(c);
 	} else {
-		err = bch2_fs_read_write(c);
-		if (err)
+		if (!test_bit(BCH_FS_RW, &c->flags)
+		    ? bch2_fs_read_write(c)
+		    : bch2_fs_read_write_late(c)) {
+			err = "error going read write";
 			goto err;
+		}
 	}
 
 	set_bit(BCH_FS_STARTED, &c->flags);

fs/bcachefs/super.h

@@ -217,7 +217,9 @@ struct bch_dev *bch2_dev_lookup(struct bch_fs *, const char *);
 
 bool bch2_fs_emergency_read_only(struct bch_fs *);
 void bch2_fs_read_only(struct bch_fs *);
-const char *bch2_fs_read_write(struct bch_fs *);
+
+int bch2_fs_read_write(struct bch_fs *);
+int bch2_fs_read_write_early(struct bch_fs *);
 
 void bch2_fs_stop(struct bch_fs *);
 

fs/bcachefs/sysfs.c

@@ -289,7 +289,7 @@ static ssize_t bch2_compression_stats(struct bch_fs *c, char *buf)
 	    compressed_sectors_compressed = 0,
 	    compressed_sectors_uncompressed = 0;
 
-	if (!bch2_fs_running(c))
+	if (!test_bit(BCH_FS_STARTED, &c->flags))
 		return -EPERM;
 
 	for_each_btree_key(&iter, c, BTREE_ID_EXTENTS, POS_MIN, 0, k)
@@ -482,7 +482,7 @@ STORE(__bch2_fs)
 	BCH_DEBUG_PARAMS()
 #undef BCH_DEBUG_PARAM
 
-	if (!bch2_fs_running(c))
+	if (!test_bit(BCH_FS_STARTED, &c->flags))
 		return -EPERM;
 
 	/* Debugging: */