X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Fsuper.c;h=0eb6b7e7531b5a4196c68dbf23f11450e5ee7575;hb=bca8b084ad754afc54e628d9db7721b90d9480b7;hp=951053f79bfeceb2ac8edfcf9b7b33489d60a259;hpb=85ee972555948337bb1a58f0702a4da95db6758f;p=bcachefs-tools-debian

diff --git a/libbcachefs/super.c b/libbcachefs/super.c
index 951053f..0eb6b7e 100644
--- a/libbcachefs/super.c
+++ b/libbcachefs/super.c
@@ -7,10 +7,10 @@
  */
 
 #include "bcachefs.h"
-#include "alloc.h"
+#include "alloc_background.h"
+#include "alloc_foreground.h"
 #include "btree_cache.h"
 #include "btree_gc.h"
-#include "btree_update.h"
 #include "btree_update_interior.h"
 #include "btree_io.h"
 #include "chardev.h"
@@ -18,20 +18,26 @@
 #include "clock.h"
 #include "compress.h"
 #include "debug.h"
+#include "disk_groups.h"
+#include "ec.h"
 #include "error.h"
 #include "fs.h"
+#include "fs-io.h"
 #include "fsck.h"
 #include "inode.h"
 #include "io.h"
 #include "journal.h"
-#include "keylist.h"
+#include "journal_reclaim.h"
 #include "move.h"
 #include "migrate.h"
 #include "movinggc.h"
+#include "quota.h"
+#include "rebalance.h"
+#include "recovery.h"
+#include "replicas.h"
 #include "super.h"
 #include "super-io.h"
 #include "sysfs.h"
-#include "tier.h"
 
 #include <linux/backing-dev.h>
 #include <linux/blkdev.h>
@@ -87,7 +93,7 @@ static DECLARE_WAIT_QUEUE_HEAD(bch_read_only_wait);
 
 static void bch2_dev_free(struct bch_dev *);
 static int bch2_dev_alloc(struct bch_fs *, unsigned);
-static int bch2_dev_sysfs_online(struct bch_dev *);
+static int bch2_dev_sysfs_online(struct bch_fs *, struct bch_dev *);
 static void __bch2_dev_read_only(struct bch_fs *, struct bch_dev *);
 
 struct bch_fs *bch2_bdev_to_fs(struct block_device *bdev)
@@ -120,7 +126,7 @@ static struct bch_fs *__bch2_uuid_to_fs(uuid_le uuid)
 	lockdep_assert_held(&bch_fs_list_lock);
 
 	list_for_each_entry(c, &bch_fs_list, list)
-		if (!memcmp(&c->disk_sb->uuid, &uuid, sizeof(uuid_le)))
+		if (!memcmp(&c->disk_sb.sb->uuid, &uuid, sizeof(uuid_le)))
 			return c;
 
 	return NULL;
@@ -139,13 +145,15 @@ struct bch_fs *bch2_uuid_to_fs(uuid_le uuid)
 	return c;
 }
 
-int bch2_congested(struct bch_fs *c, int bdi_bits)
+int bch2_congested(void *data, int bdi_bits)
 {
+	struct bch_fs *c = data;
 	struct backing_dev_info *bdi;
 	struct bch_dev *ca;
 	unsigned i;
 	int ret = 0;
 
+	rcu_read_lock();
 	if (bdi_bits & (1 << WB_sync_congested)) {
 		/* Reads - check all devices: */
 		for_each_readable_member(ca, c, i) {
@@ -157,12 +165,11 @@ int bch2_congested(struct bch_fs *c, int bdi_bits)
 			}
 		}
 	} else {
-		/* Writes prefer fastest tier: */
-		struct bch_tier *tier = READ_ONCE(c->fastest_tier);
-		struct bch_devs_mask *devs =
-			tier ? &tier->devs : &c->rw_devs[BCH_DATA_USER];
+		unsigned target = READ_ONCE(c->opts.foreground_target);
+		const struct bch_devs_mask *devs = target
+			? bch2_target_to_mask(c, target)
+			: &c->rw_devs[BCH_DATA_USER];
 
-		rcu_read_lock();
 		for_each_member_device_rcu(ca, c, i, devs) {
 			bdi = ca->disk_sb.bdev->bd_bdi;
 
@@ -171,27 +178,20 @@ int bch2_congested(struct bch_fs *c, int bdi_bits)
 				break;
 			}
 		}
-		rcu_read_unlock();
 	}
+	rcu_read_unlock();
 
 	return ret;
 }
 
-static int bch2_congested_fn(void *data, int bdi_bits)
-{
-	struct bch_fs *c = data;
-
-	return bch2_congested(c, bdi_bits);
-}
-
 /* Filesystem RO/RW: */
 
 /*
  * For startup/shutdown of RW stuff, the dependencies are:
  *
- * - foreground writes depend on copygc and tiering (to free up space)
+ * - foreground writes depend on copygc and rebalance (to free up space)
  *
- * - copygc and tiering depend on mark and sweep gc (they actually probably
+ * - copygc and rebalance depend on mark and sweep gc (they actually probably
  *   don't because they either reserve ahead of time or don't block if
  *   allocations fail, but allocations can require mark and sweep gc to run
  *   because of generation number wraparound)
@@ -206,10 +206,10 @@ static void __bch2_fs_read_only(struct bch_fs *c)
 	struct bch_dev *ca;
 	unsigned i;
 
-	bch2_tiering_stop(c);
+	bch2_rebalance_stop(c);
 
 	for_each_member_device(ca, c, i)
-		bch2_moving_gc_stop(ca);
+		bch2_copygc_stop(ca);
 
 	bch2_gc_thread_stop(c);
 
@@ -217,16 +217,37 @@ static void __bch2_fs_read_only(struct bch_fs *c)
 	 * Flush journal before stopping allocators, because flushing journal
 	 * blacklist entries involves allocating new btree nodes:
 	 */
-	bch2_journal_flush_pins(&c->journal, U64_MAX);
-
-	if (!bch2_journal_error(&c->journal))
-		bch2_btree_verify_flushed(c);
+	bch2_journal_flush_all_pins(&c->journal);
 
 	for_each_member_device(ca, c, i)
 		bch2_dev_allocator_stop(ca);
 
+	bch2_journal_flush_all_pins(&c->journal);
+
+	/*
+	 * We need to explicitly wait on btree interior updates to complete
+	 * before stopping the journal, flushing all journal pins isn't
+	 * sufficient, because in the BTREE_INTERIOR_UPDATING_ROOT case btree
+	 * interior updates have to drop their journal pin before they're
+	 * fully complete:
+	 */
+	closure_wait_event(&c->btree_interior_update_wait,
+			   !bch2_btree_interior_updates_nr_pending(c));
+
 	bch2_fs_journal_stop(&c->journal);
 
+	/*
+	 * the journal kicks off btree writes via reclaim - wait for in flight
+	 * writes after stopping journal:
+	 */
+	if (test_bit(BCH_FS_EMERGENCY_RO, &c->flags))
+		bch2_btree_flush_all_writes(c);
+	else
+		bch2_btree_verify_flushed(c);
+
+	/*
+	 * After stopping journal:
+	 */
 	for_each_member_device(ca, c, i)
 		bch2_dev_allocator_remove(c, ca);
 }
@@ -241,13 +262,10 @@ static void bch2_writes_disabled(struct percpu_ref *writes)
 
 void bch2_fs_read_only(struct bch_fs *c)
 {
-	mutex_lock(&c->state_lock);
-	if (c->state != BCH_FS_STARTING &&
-	    c->state != BCH_FS_RW)
-		goto out;
+	if (c->state == BCH_FS_RO)
+		return;
 
-	if (test_bit(BCH_FS_ERROR, &c->flags))
-		goto out;
+	BUG_ON(test_bit(BCH_FS_WRITE_DISABLE_COMPLETE, &c->flags));
 
 	/*
 	 * Block new foreground-end write operations from starting - any new
@@ -259,12 +277,8 @@ void bch2_fs_read_only(struct bch_fs *c)
 	 */
 	percpu_ref_kill(&c->writes);
 
-	del_timer(&c->foreground_write_wakeup);
 	cancel_delayed_work(&c->pd_controllers_update);
 
-	c->foreground_write_pd.rate.rate = UINT_MAX;
-	bch2_wake_delayed_writes((unsigned long) c);
-
 	/*
 	 * If we're not doing an emergency shutdown, we want to wait on
 	 * outstanding writes to complete so they don't see spurious errors due
@@ -288,16 +302,12 @@ void bch2_fs_read_only(struct bch_fs *c)
 	clear_bit(BCH_FS_WRITE_DISABLE_COMPLETE, &c->flags);
 
 	if (!bch2_journal_error(&c->journal) &&
-	    !test_bit(BCH_FS_ERROR, &c->flags)) {
-		mutex_lock(&c->sb_lock);
-		SET_BCH_SB_CLEAN(c->disk_sb, true);
-		bch2_write_super(c);
-		mutex_unlock(&c->sb_lock);
-	}
+	    !test_bit(BCH_FS_ERROR, &c->flags) &&
+	    !test_bit(BCH_FS_EMERGENCY_RO, &c->flags))
+		bch2_fs_mark_clean(c, true);
 
-	c->state = BCH_FS_RO;
-out:
-	mutex_unlock(&c->state_lock);
+	if (c->state != BCH_FS_STOPPING)
+		c->state = BCH_FS_RO;
 }
 
 static void bch2_fs_read_only_work(struct work_struct *work)
@@ -305,7 +315,9 @@ static void bch2_fs_read_only_work(struct work_struct *work)
 	struct bch_fs *c =
 		container_of(work, struct bch_fs, read_only_work);
 
+	mutex_lock(&c->state_lock);
 	bch2_fs_read_only(c);
+	mutex_unlock(&c->state_lock);
 }
 
 static void bch2_fs_read_only_async(struct bch_fs *c)
@@ -330,10 +342,10 @@ const char *bch2_fs_read_write(struct bch_fs *c)
 	const char *err = NULL;
 	unsigned i;
 
-	mutex_lock(&c->state_lock);
-	if (c->state != BCH_FS_STARTING &&
-	    c->state != BCH_FS_RO)
-		goto out;
+	if (c->state == BCH_FS_RW)
+		return NULL;
+
+	bch2_fs_mark_clean(c, false);
 
 	for_each_rw_member(ca, c, i)
 		bch2_dev_allocator_add(c, ca);
@@ -350,15 +362,15 @@ const char *bch2_fs_read_write(struct bch_fs *c)
 	if (bch2_gc_thread_start(c))
 		goto err;
 
-	err = "error starting moving GC thread";
+	err = "error starting copygc thread";
 	for_each_rw_member(ca, c, i)
-		if (bch2_moving_gc_start(ca)) {
+		if (bch2_copygc_start(c, ca)) {
 			percpu_ref_put(&ca->io_ref);
 			goto err;
 		}
 
-	err = "error starting tiering thread";
-	if (bch2_tiering_start(c))
+	err = "error starting rebalance thread";
+	if (bch2_rebalance_start(c))
 		goto err;
 
 	schedule_delayed_work(&c->pd_controllers_update, 5 * HZ);
@@ -367,74 +379,67 @@ const char *bch2_fs_read_write(struct bch_fs *c)
 		percpu_ref_reinit(&c->writes);
 
 	c->state = BCH_FS_RW;
-	err = NULL;
-out:
-	mutex_unlock(&c->state_lock);
-	return err;
+	return NULL;
 err:
 	__bch2_fs_read_only(c);
-	goto out;
+	return err;
 }
 
 /* Filesystem startup/shutdown: */
 
 static void bch2_fs_free(struct bch_fs *c)
 {
+	unsigned i;
+
+	for (i = 0; i < BCH_TIME_STAT_NR; i++)
+		bch2_time_stats_exit(&c->times[i]);
+
+	bch2_fs_quota_exit(c);
+	bch2_fs_fsio_exit(c);
+	bch2_fs_ec_exit(c);
 	bch2_fs_encryption_exit(c);
-	bch2_fs_btree_exit(c);
+	bch2_fs_io_exit(c);
+	bch2_fs_btree_cache_exit(c);
 	bch2_fs_journal_exit(&c->journal);
 	bch2_io_clock_exit(&c->io_clock[WRITE]);
 	bch2_io_clock_exit(&c->io_clock[READ]);
 	bch2_fs_compress_exit(c);
-	if (c->bdi.bdi_list.next)
-		bdi_destroy(&c->bdi);
-	lg_lock_free(&c->usage_lock);
-	free_percpu(c->usage_percpu);
+	percpu_free_rwsem(&c->usage_lock);
+	free_percpu(c->usage[0]);
+	mempool_exit(&c->btree_iters_pool);
 	mempool_exit(&c->btree_bounce_pool);
-	mempool_exit(&c->bio_bounce_pages);
-	bioset_exit(&c->bio_write);
-	bioset_exit(&c->bio_read_split);
-	bioset_exit(&c->bio_read);
-	bioset_exit(&c->btree_read_bio);
+	bioset_exit(&c->btree_bio);
 	mempool_exit(&c->btree_interior_update_pool);
 	mempool_exit(&c->btree_reserve_pool);
 	mempool_exit(&c->fill_iter);
 	percpu_ref_exit(&c->writes);
+	kfree(rcu_dereference_protected(c->replicas, 1));
+	kfree(rcu_dereference_protected(c->disk_groups, 1));
 
 	if (c->copygc_wq)
 		destroy_workqueue(c->copygc_wq);
 	if (c->wq)
 		destroy_workqueue(c->wq);
 
-	free_pages((unsigned long) c->disk_sb, c->disk_sb_order);
-	kfree(c);
+	free_pages((unsigned long) c->disk_sb.sb,
+		   c->disk_sb.page_order);
+	kvpfree(c, sizeof(*c));
 	module_put(THIS_MODULE);
 }
 
-static void bch2_fs_exit(struct bch_fs *c)
+static void bch2_fs_release(struct kobject *kobj)
 {
-	unsigned i;
-
-	del_timer_sync(&c->foreground_write_wakeup);
-	cancel_delayed_work_sync(&c->pd_controllers_update);
-	cancel_work_sync(&c->read_only_work);
-
-	for (i = 0; i < c->sb.nr_devices; i++)
-		if (c->devs[i])
-			bch2_dev_free(c->devs[i]);
+	struct bch_fs *c = container_of(kobj, struct bch_fs, kobj);
 
-	closure_debug_destroy(&c->cl);
-	kobject_put(&c->kobj);
+	bch2_fs_free(c);
 }
 
-static void bch2_fs_offline(struct bch_fs *c)
+void bch2_fs_stop(struct bch_fs *c)
 {
 	struct bch_dev *ca;
 	unsigned i;
 
-	mutex_lock(&bch_fs_list_lock);
-	list_del(&c->list);
-	mutex_unlock(&bch_fs_list_lock);
+	bch_verbose(c, "shutting down");
 
 	for_each_member_device(ca, c, i)
 		if (ca->kobj.state_in_sysfs &&
@@ -452,28 +457,76 @@ static void bch2_fs_offline(struct bch_fs *c)
 	kobject_put(&c->opts_dir);
 	kobject_put(&c->internal);
 
-	__bch2_fs_read_only(c);
-}
+	mutex_lock(&bch_fs_list_lock);
+	list_del(&c->list);
+	mutex_unlock(&bch_fs_list_lock);
 
-static void bch2_fs_release(struct kobject *kobj)
-{
-	struct bch_fs *c = container_of(kobj, struct bch_fs, kobj);
+	closure_sync(&c->cl);
+	closure_debug_destroy(&c->cl);
 
-	bch2_fs_free(c);
+	mutex_lock(&c->state_lock);
+	bch2_fs_read_only(c);
+	mutex_unlock(&c->state_lock);
+
+	/* btree prefetch might have kicked off reads in the background: */
+	bch2_btree_flush_all_reads(c);
+
+	for_each_member_device(ca, c, i)
+		cancel_work_sync(&ca->io_error_work);
+
+	cancel_work_sync(&c->btree_write_error_work);
+	cancel_delayed_work_sync(&c->pd_controllers_update);
+	cancel_work_sync(&c->read_only_work);
+
+	for (i = 0; i < c->sb.nr_devices; i++)
+		if (c->devs[i])
+			bch2_dev_free(rcu_dereference_protected(c->devs[i], 1));
+
+	bch_verbose(c, "shutdown complete");
+
+	kobject_put(&c->kobj);
 }
 
-void bch2_fs_stop(struct bch_fs *c)
+static const char *bch2_fs_online(struct bch_fs *c)
 {
-	mutex_lock(&c->state_lock);
-	BUG_ON(c->state == BCH_FS_STOPPING);
-	c->state = BCH_FS_STOPPING;
-	mutex_unlock(&c->state_lock);
+	struct bch_dev *ca;
+	const char *err = NULL;
+	unsigned i;
+	int ret;
 
-	bch2_fs_offline(c);
+	lockdep_assert_held(&bch_fs_list_lock);
 
-	closure_sync(&c->cl);
+	if (!list_empty(&c->list))
+		return NULL;
+
+	if (__bch2_uuid_to_fs(c->sb.uuid))
+		return "filesystem UUID already open";
+
+	ret = bch2_fs_chardev_init(c);
+	if (ret)
+		return "error creating character device";
+
+	bch2_fs_debug_init(c);
+
+	if (kobject_add(&c->kobj, NULL, "%pU", c->sb.user_uuid.b) ||
+	    kobject_add(&c->internal, &c->kobj, "internal") ||
+	    kobject_add(&c->opts_dir, &c->kobj, "options") ||
+	    kobject_add(&c->time_stats, &c->kobj, "time_stats") ||
+	    bch2_opts_create_sysfs_files(&c->opts_dir))
+		return "error creating sysfs objects";
+
+	mutex_lock(&c->state_lock);
+
+	err = "error creating sysfs objects";
+	__for_each_member_device(ca, c, i, NULL)
+		if (bch2_dev_sysfs_online(c, ca))
+			goto err;
 
-	bch2_fs_exit(c);
+	list_add(&c->list, &bch_fs_list);
+	err = NULL;
+err:
+	mutex_unlock(&c->state_lock);
+	return err;
 }
 
 static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
@@ -481,44 +534,42 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 	struct bch_sb_field_members *mi;
 	struct bch_fs *c;
 	unsigned i, iter_size;
+	const char *err;
+
+	pr_verbose_init(opts, "");
 
-	c = kzalloc(sizeof(struct bch_fs), GFP_KERNEL);
+	c = kvpmalloc(sizeof(struct bch_fs), GFP_KERNEL|__GFP_ZERO);
 	if (!c)
-		return NULL;
+		goto out;
 
 	__module_get(THIS_MODULE);
 
 	c->minor		= -1;
+	c->disk_sb.fs_sb	= true;
 
 	mutex_init(&c->state_lock);
 	mutex_init(&c->sb_lock);
 	mutex_init(&c->replicas_gc_lock);
-	mutex_init(&c->btree_cache_lock);
-	mutex_init(&c->bucket_lock);
 	mutex_init(&c->btree_root_lock);
 	INIT_WORK(&c->read_only_work, bch2_fs_read_only_work);
 
 	init_rwsem(&c->gc_lock);
 
-#define BCH_TIME_STAT(name, frequency_units, duration_units)		\
-	spin_lock_init(&c->name##_time.lock);
-	BCH_TIME_STATS()
-#undef BCH_TIME_STAT
+	for (i = 0; i < BCH_TIME_STAT_NR; i++)
+		bch2_time_stats_init(&c->times[i]);
 
-	bch2_fs_allocator_init(c);
-	bch2_fs_tiering_init(c);
+	bch2_fs_allocator_background_init(c);
+	bch2_fs_allocator_foreground_init(c);
+	bch2_fs_rebalance_init(c);
+	bch2_fs_quota_init(c);
 
 	INIT_LIST_HEAD(&c->list);
-	INIT_LIST_HEAD(&c->btree_cache);
-	INIT_LIST_HEAD(&c->btree_cache_freeable);
-	INIT_LIST_HEAD(&c->btree_cache_freed);
 
 	INIT_LIST_HEAD(&c->btree_interior_update_list);
 	mutex_init(&c->btree_reserve_cache_lock);
 	mutex_init(&c->btree_interior_update_lock);
 
 	mutex_init(&c->bio_bounce_pages_lock);
-	mutex_init(&c->zlib_workspace_lock);
 
 	bio_list_init(&c->btree_write_error_list);
 	spin_lock_init(&c->btree_write_error_lock);
@@ -527,26 +578,23 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 	INIT_LIST_HEAD(&c->fsck_errors);
 	mutex_init(&c->fsck_error_lock);
 
-	seqcount_init(&c->gc_pos_lock);
-
-	c->prio_clock[READ].hand = 1;
-	c->prio_clock[READ].min_prio = 0;
-	c->prio_clock[WRITE].hand = 1;
-	c->prio_clock[WRITE].min_prio = 0;
+	INIT_LIST_HEAD(&c->ec_new_stripe_list);
+	mutex_init(&c->ec_new_stripe_lock);
+	mutex_init(&c->ec_stripes_lock);
+	spin_lock_init(&c->ec_stripes_heap_lock);
 
-	init_waitqueue_head(&c->writeback_wait);
-	c->writeback_pages_max = (256 << 10) / PAGE_SIZE;
+	seqcount_init(&c->gc_pos_lock);
 
-	c->copy_gc_enabled = 1;
-	c->tiering_enabled = 1;
-	c->tiering_percent = 10;
+	c->copy_gc_enabled		= 1;
+	c->rebalance.enabled		= 1;
+	c->promote_whole_extents	= true;
 
-	c->foreground_target_percent = 20;
+	c->journal.write_time	= &c->times[BCH_TIME_journal_write];
+	c->journal.delay_time	= &c->times[BCH_TIME_journal_delay];
+	c->journal.blocked_time	= &c->times[BCH_TIME_journal_blocked];
+	c->journal.flush_seq_time = &c->times[BCH_TIME_journal_flush_seq];
 
-	c->journal.write_time	= &c->journal_write_time;
-	c->journal.delay_time	= &c->journal_delay_time;
-	c->journal.blocked_time	= &c->journal_blocked_time;
-	c->journal.flush_seq_time = &c->journal_flush_seq_time;
+	bch2_fs_btree_cache_init_early(&c->btree_cache);
 
 	mutex_lock(&c->sb_lock);
 
@@ -555,22 +603,25 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 		goto err;
 	}
 
-	c->block_bits		= ilog2(c->sb.block_size);
-
 	mutex_unlock(&c->sb_lock);
 
 	scnprintf(c->name, sizeof(c->name), "%pU", &c->sb.user_uuid);
 
-	bch2_opts_apply(&c->opts, bch2_sb_opts(sb));
+	c->opts = bch2_opts_default;
+	bch2_opts_apply(&c->opts, bch2_opts_from_sb(sb));
 	bch2_opts_apply(&c->opts, opts);
 
+	c->block_bits		= ilog2(c->opts.block_size);
+	c->btree_foreground_merge_threshold = BTREE_FOREGROUND_MERGE_THRESHOLD(c);
+
 	c->opts.nochanges	|= c->opts.noreplay;
 	c->opts.read_only	|= c->opts.nochanges;
 
 	if (bch2_fs_init_fault("fs_alloc"))
 		goto err;
 
-	iter_size = (btree_blocks(c) + 1) * 2 *
+	iter_size = sizeof(struct btree_node_iter_large) +
+		(btree_blocks(c) + 1) * 2 *
 		sizeof(struct btree_node_iter_set);
 
 	if (!(c->wq = alloc_workqueue("bcachefs",
@@ -583,36 +634,30 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 	    mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
 				      sizeof(struct btree_update)) ||
 	    mempool_init_kmalloc_pool(&c->fill_iter, 1, iter_size) ||
-	    bioset_init(&c->btree_read_bio, 1,
-			offsetof(struct btree_read_bio, bio)) ||
-	    bioset_init(&c->bio_read, 1, offsetof(struct bch_read_bio, bio)) ||
-	    bioset_init(&c->bio_read_split, 1, offsetof(struct bch_read_bio, bio)) ||
-	    bioset_init(&c->bio_write, 1, offsetof(struct bch_write_bio, bio)) ||
-	    mempool_init_page_pool(&c->bio_bounce_pages,
-				   max_t(unsigned,
-					 c->sb.btree_node_size,
-					 c->sb.encoded_extent_max) /
-				   PAGE_SECTORS, 0) ||
-	    !(c->usage_percpu = alloc_percpu(struct bch_fs_usage)) ||
-	    lg_lock_init(&c->usage_lock) ||
-	    mempool_init_vp_pool(&c->btree_bounce_pool, 1, btree_bytes(c)) ||
-	    bdi_setup_and_register(&c->bdi, "bcachefs") ||
+	    bioset_init(&c->btree_bio, 1,
+			max(offsetof(struct btree_read_bio, bio),
+			    offsetof(struct btree_write_bio, wbio.bio)),
+			BIOSET_NEED_BVECS) ||
+	    !(c->usage[0] = alloc_percpu(struct bch_fs_usage)) ||
+	    percpu_init_rwsem(&c->usage_lock) ||
+	    mempool_init_kvpmalloc_pool(&c->btree_bounce_pool, 1,
+					btree_bytes(c)) ||
+	    mempool_init_kmalloc_pool(&c->btree_iters_pool, 1,
+			sizeof(struct btree_iter) * BTREE_ITER_MAX) ||
 	    bch2_io_clock_init(&c->io_clock[READ]) ||
 	    bch2_io_clock_init(&c->io_clock[WRITE]) ||
 	    bch2_fs_journal_init(&c->journal) ||
-	    bch2_fs_btree_init(c) ||
+	    bch2_fs_btree_cache_init(c) ||
+	    bch2_fs_io_init(c) ||
 	    bch2_fs_encryption_init(c) ||
 	    bch2_fs_compress_init(c) ||
-	    bch2_check_set_has_compressed_data(c, c->opts.compression))
+	    bch2_fs_ec_init(c) ||
+	    bch2_fs_fsio_init(c))
 		goto err;
 
-	c->bdi.ra_pages		= VM_MAX_READAHEAD * 1024 / PAGE_SIZE;
-	c->bdi.congested_fn	= bch2_congested_fn;
-	c->bdi.congested_data	= c;
-
-	mi = bch2_sb_get_members(c->disk_sb);
+	mi = bch2_sb_get_members(c->disk_sb.sb);
 	for (i = 0; i < c->sb.nr_devices; i++)
-		if (bch2_dev_exists(c->disk_sb, mi, i) &&
+		if (bch2_dev_exists(c->disk_sb.sb, mi, i) &&
 		    bch2_dev_alloc(c, i))
 			goto err;
 
@@ -627,225 +672,61 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 	kobject_init(&c->internal, &bch2_fs_internal_ktype);
 	kobject_init(&c->opts_dir, &bch2_fs_opts_dir_ktype);
 	kobject_init(&c->time_stats, &bch2_fs_time_stats_ktype);
-	return c;
-err:
-	bch2_fs_free(c);
-	return NULL;
-}
-
-static const char *__bch2_fs_online(struct bch_fs *c)
-{
-	struct bch_dev *ca;
-	const char *err = NULL;
-	unsigned i;
-	int ret;
-
-	lockdep_assert_held(&bch_fs_list_lock);
-
-	if (!list_empty(&c->list))
-		return NULL;
-
-	if (__bch2_uuid_to_fs(c->sb.uuid))
-		return "filesystem UUID already open";
-
-	ret = bch2_fs_chardev_init(c);
-	if (ret)
-		return "error creating character device";
-
-	bch2_fs_debug_init(c);
-
-	if (kobject_add(&c->kobj, NULL, "%pU", c->sb.user_uuid.b) ||
-	    kobject_add(&c->internal, &c->kobj, "internal") ||
-	    kobject_add(&c->opts_dir, &c->kobj, "options") ||
-	    kobject_add(&c->time_stats, &c->kobj, "time_stats"))
-		return "error creating sysfs objects";
-
-	mutex_lock(&c->state_lock);
-
-	err = "error creating sysfs objects";
-	__for_each_member_device(ca, c, i, NULL)
-		if (bch2_dev_sysfs_online(ca))
-			goto err;
-
-	list_add(&c->list, &bch_fs_list);
-	err = NULL;
-err:
-	mutex_unlock(&c->state_lock);
-	return err;
-}
-
-static const char *bch2_fs_online(struct bch_fs *c)
-{
-	const char *err;
 
 	mutex_lock(&bch_fs_list_lock);
-	err = __bch2_fs_online(c);
+	err = bch2_fs_online(c);
 	mutex_unlock(&bch_fs_list_lock);
-
-	return err;
+	if (err) {
+		bch_err(c, "bch2_fs_online() error: %s", err);
+		goto err;
+	}
+out:
+	pr_verbose_init(opts, "ret %i", c ? 0 : -ENOMEM);
+	return c;
+err:
+	bch2_fs_free(c);
+	c = NULL;
+	goto out;
 }
 
-static const char *__bch2_fs_start(struct bch_fs *c)
+const char *bch2_fs_start(struct bch_fs *c)
 {
 	const char *err = "cannot allocate memory";
 	struct bch_sb_field_members *mi;
 	struct bch_dev *ca;
-	LIST_HEAD(journal);
-	struct jset *j;
-	struct closure cl;
-	time64_t now;
+	time64_t now = ktime_get_real_seconds();
 	unsigned i;
 	int ret = -EINVAL;
 
-	closure_init_stack(&cl);
+	mutex_lock(&c->state_lock);
 
 	BUG_ON(c->state != BCH_FS_STARTING);
 
 	mutex_lock(&c->sb_lock);
+
 	for_each_online_member(ca, c, i)
 		bch2_sb_from_fs(c, ca);
+
+	mi = bch2_sb_get_members(c->disk_sb.sb);
+	for_each_online_member(ca, c, i)
+		mi->members[ca->dev_idx].last_mount = cpu_to_le64(now);
+
 	mutex_unlock(&c->sb_lock);
 
 	for_each_rw_member(ca, c, i)
 		bch2_dev_allocator_add(c, ca);
 	bch2_recalc_capacity(c);
 
-	if (BCH_SB_INITIALIZED(c->disk_sb)) {
-		ret = bch2_journal_read(c, &journal);
-		if (ret)
-			goto err;
-
-		j = &list_entry(journal.prev, struct journal_replay, list)->j;
-
-		c->prio_clock[READ].hand = le16_to_cpu(j->read_clock);
-		c->prio_clock[WRITE].hand = le16_to_cpu(j->write_clock);
-
-		for (i = 0; i < BTREE_ID_NR; i++) {
-			unsigned level;
-			struct bkey_i *k;
-
-			err = "missing btree root";
-			k = bch2_journal_find_btree_root(c, j, i, &level);
-			if (!k && i < BTREE_ID_ALLOC)
-				goto err;
-
-			if (!k)
-				continue;
-
-			err = "error reading btree root";
-			if (bch2_btree_root_read(c, i, k, level))
-				goto err;
-		}
-
-		err = "error reading allocation information";
-		ret = bch2_alloc_read(c, &journal);
-		if (ret)
-			goto err;
-
-		bch_verbose(c, "starting mark and sweep:");
-		err = "error in recovery";
-		ret = bch2_initial_gc(c, &journal);
-		if (ret)
-			goto err;
-		bch_verbose(c, "mark and sweep done");
-
-		if (c->opts.noreplay)
-			goto recovery_done;
-
-		err = "cannot allocate new btree root";
-		for (i = 0; i < BTREE_ID_NR; i++)
-			if (!c->btree_roots[i].b &&
-			    bch2_btree_root_alloc(c, i, &cl))
-				goto err;
-
-		closure_sync(&cl);
-
-		/*
-		 * bch2_journal_start() can't happen sooner, or btree_gc_finish()
-		 * will give spurious errors about oldest_gen > bucket_gen -
-		 * this is a hack but oh well.
-		 */
-		bch2_journal_start(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);
-				goto err;
-			}
-
-		bch_verbose(c, "starting journal replay:");
-		err = "journal replay failed";
-		ret = bch2_journal_replay(c, &journal);
-		if (ret)
-			goto err;
-		bch_verbose(c, "journal replay done");
-
-		if (c->opts.norecovery)
-			goto recovery_done;
-
-		bch_verbose(c, "starting fsck:");
-		err = "error in fsck";
-		ret = bch2_fsck(c, !c->opts.nofsck);
-		if (ret)
-			goto err;
-		bch_verbose(c, "fsck done");
-	} else {
-		struct bch_inode_unpacked inode;
-		struct bkey_inode_buf packed_inode;
-
-		bch_notice(c, "initializing new filesystem");
-
-		ret = bch2_initial_gc(c, &journal);
-		if (ret)
-			goto err;
-
-		err = "unable to allocate journal buckets";
-		for_each_rw_member(ca, c, i)
-			if (bch2_dev_journal_alloc(ca)) {
-				percpu_ref_put(&ca->io_ref);
-				goto err;
-			}
-
-		err = "cannot allocate new btree root";
-		for (i = 0; i < BTREE_ID_NR; i++)
-			if (bch2_btree_root_alloc(c, i, &cl))
-				goto err;
-
-		/*
-		 * journal_res_get() will crash if called before this has
-		 * set up the journal.pin FIFO and journal.cur pointer:
-		 */
-		bch2_journal_start(c);
-		bch2_journal_set_replay_done(&c->journal);
-
-		err = "error starting allocator thread";
-		for_each_rw_member(ca, c, i)
-			if (bch2_dev_allocator_start(ca)) {
-				percpu_ref_put(&ca->io_ref);
-				goto err;
-			}
-
-		/* Wait for new btree roots to be written: */
-		closure_sync(&cl);
-
-		bch2_inode_init(c, &inode, 0, 0,
-			       S_IFDIR|S_IRWXU|S_IRUGO|S_IXUGO, 0);
-		inode.inum = BCACHEFS_ROOT_INO;
-
-		bch2_inode_pack(&packed_inode, &inode);
+	ret = BCH_SB_INITIALIZED(c->disk_sb.sb)
+		? bch2_fs_recovery(c)
+		: bch2_fs_initialize(c);
+	if (ret)
+		goto err;
 
-		err = "error creating root directory";
-		if (bch2_btree_insert(c, BTREE_ID_INODES,
-				     &packed_inode.inode.k_i,
-				     NULL, NULL, NULL, 0))
-			goto err;
+	ret = bch2_opts_check_may_set(c);
+	if (ret)
+		goto err;
 
-		err = "error writing first journal entry";
-		if (bch2_journal_meta(&c->journal))
-			goto err;
-	}
-recovery_done:
 	err = "dynamic fault";
 	if (bch2_fs_init_fault("fs_start"))
 		goto err;
@@ -858,26 +739,13 @@ recovery_done:
 			goto err;
 	}
 
-	mutex_lock(&c->sb_lock);
-	mi = bch2_sb_get_members(c->disk_sb);
-	now = ktime_get_seconds();
-
-	for_each_member_device(ca, c, i)
-		mi->members[ca->dev_idx].last_mount = cpu_to_le64(now);
-
-	SET_BCH_SB_INITIALIZED(c->disk_sb, true);
-	SET_BCH_SB_CLEAN(c->disk_sb, false);
-
-	bch2_write_super(c);
-	mutex_unlock(&c->sb_lock);
+	set_bit(BCH_FS_STARTED, &c->flags);
 
 	err = NULL;
 out:
-	bch2_journal_entries_free(&journal);
+	mutex_unlock(&c->state_lock);
 	return err;
 err:
-	closure_sync(&cl);
-
 	switch (ret) {
 	case BCH_FSCK_ERRORS_NOT_FIXED:
 		bch_err(c, "filesystem contains errors: please report this to the developers");
@@ -909,11 +777,6 @@ err:
 	goto out;
 }
 
-const char *bch2_fs_start(struct bch_fs *c)
-{
-	return __bch2_fs_start(c) ?: bch2_fs_online(c);
-}
-
 static const char *bch2_dev_may_add(struct bch_sb *sb, struct bch_fs *c)
 {
 	struct bch_sb_field_members *sb_mi;
@@ -922,11 +785,11 @@ static const char *bch2_dev_may_add(struct bch_sb *sb, struct bch_fs *c)
 	if (!sb_mi)
 		return "Invalid superblock: member info area missing";
 
-	if (le16_to_cpu(sb->block_size) != c->sb.block_size)
+	if (le16_to_cpu(sb->block_size) != c->opts.block_size)
 		return "mismatched block size";
 
 	if (le16_to_cpu(sb_mi->members[sb->dev_idx].bucket_size) <
-	    BCH_SB_BTREE_NODE_SIZE(c->disk_sb))
+	    BCH_SB_BTREE_NODE_SIZE(c->disk_sb.sb))
 		return "new cache bucket size is too small";
 
 	return NULL;
@@ -961,8 +824,6 @@ static void bch2_dev_release(struct kobject *kobj)
 
 static void bch2_dev_free(struct bch_dev *ca)
 {
-	unsigned i;
-
 	cancel_work_sync(&ca->io_error_work);
 
 	if (ca->kobj.state_in_sysfs &&
@@ -978,40 +839,29 @@ static void bch2_dev_free(struct bch_dev *ca)
 
 	free_percpu(ca->io_done);
 	bioset_exit(&ca->replica_set);
-	free_percpu(ca->usage_percpu);
-	kvpfree(ca->bucket_dirty, BITS_TO_LONGS(ca->mi.nbuckets) * sizeof(unsigned long));
-	kvpfree(ca->buckets,	 ca->mi.nbuckets * sizeof(struct bucket));
-	kvpfree(ca->oldest_gens, ca->mi.nbuckets * sizeof(u8));
-	free_heap(&ca->copygc_heap);
-	free_heap(&ca->alloc_heap);
-	free_fifo(&ca->free_inc);
+	bch2_dev_buckets_free(ca);
 
-	for (i = 0; i < RESERVE_NR; i++)
-		free_fifo(&ca->free[i]);
+	bch2_time_stats_exit(&ca->io_latency[WRITE]);
+	bch2_time_stats_exit(&ca->io_latency[READ]);
 
 	percpu_ref_exit(&ca->io_ref);
 	percpu_ref_exit(&ca->ref);
 	kobject_put(&ca->kobj);
 }
 
-static void bch2_dev_io_ref_release(struct percpu_ref *ref)
-{
-	struct bch_dev *ca = container_of(ref, struct bch_dev, io_ref);
-
-	complete(&ca->offline_complete);
-}
-
-static void __bch2_dev_offline(struct bch_dev *ca)
+static void __bch2_dev_offline(struct bch_fs *c, struct bch_dev *ca)
 {
-	struct bch_fs *c = ca->fs;
 
 	lockdep_assert_held(&c->state_lock);
 
+	if (percpu_ref_is_zero(&ca->io_ref))
+		return;
+
 	__bch2_dev_read_only(c, ca);
 
-	reinit_completion(&ca->offline_complete);
+	reinit_completion(&ca->io_ref_completion);
 	percpu_ref_kill(&ca->io_ref);
-	wait_for_completion(&ca->offline_complete);
+	wait_for_completion(&ca->io_ref_completion);
 
 	if (ca->kobj.state_in_sysfs) {
 		struct kobject *block =
@@ -1025,32 +875,22 @@ static void __bch2_dev_offline(struct bch_dev *ca)
 	bch2_dev_journal_exit(ca);
 }
 
-static void bch2_dev_ref_release(struct percpu_ref *ref)
+static void bch2_dev_ref_complete(struct percpu_ref *ref)
 {
 	struct bch_dev *ca = container_of(ref, struct bch_dev, ref);
 
-	complete(&ca->stop_complete);
+	complete(&ca->ref_completion);
 }
 
-static void bch2_dev_stop(struct bch_dev *ca)
+static void bch2_dev_io_ref_complete(struct percpu_ref *ref)
 {
-	struct bch_fs *c = ca->fs;
-
-	lockdep_assert_held(&c->state_lock);
-
-	BUG_ON(rcu_access_pointer(c->devs[ca->dev_idx]) != ca);
-	rcu_assign_pointer(c->devs[ca->dev_idx], NULL);
-
-	synchronize_rcu();
+	struct bch_dev *ca = container_of(ref, struct bch_dev, io_ref);
 
-	reinit_completion(&ca->stop_complete);
-	percpu_ref_kill(&ca->ref);
-	wait_for_completion(&ca->stop_complete);
+	complete(&ca->io_ref_completion);
 }
 
-static int bch2_dev_sysfs_online(struct bch_dev *ca)
+static int bch2_dev_sysfs_online(struct bch_fs *c, struct bch_dev *ca)
 {
-	struct bch_fs *c = ca->fs;
 	int ret;
 
 	if (!c->kobj.state_in_sysfs)
@@ -1078,160 +918,164 @@ static int bch2_dev_sysfs_online(struct bch_dev *ca)
 	return 0;
 }
 
-static int bch2_dev_alloc(struct bch_fs *c, unsigned dev_idx)
+static struct bch_dev *__bch2_dev_alloc(struct bch_fs *c,
+					struct bch_member *member)
 {
-	struct bch_member *member;
-	size_t reserve_none, movinggc_reserve, free_inc_reserve, total_reserve;
-	size_t heap_size;
-	unsigned i, btree_node_reserve_buckets;
 	struct bch_dev *ca;
 
-	if (bch2_fs_init_fault("dev_alloc"))
-		return -ENOMEM;
-
 	ca = kzalloc(sizeof(*ca), GFP_KERNEL);
 	if (!ca)
-		return -ENOMEM;
+		return NULL;
 
 	kobject_init(&ca->kobj, &bch2_dev_ktype);
-	init_completion(&ca->stop_complete);
-	init_completion(&ca->offline_complete);
+	init_completion(&ca->ref_completion);
+	init_completion(&ca->io_ref_completion);
 
-	ca->dev_idx = dev_idx;
-	__set_bit(ca->dev_idx, ca->self.d);
+	init_rwsem(&ca->bucket_lock);
 
-	ca->copygc_write_point.type = BCH_DATA_USER;
+	writepoint_init(&ca->copygc_write_point, BCH_DATA_USER);
 
 	spin_lock_init(&ca->freelist_lock);
-	bch2_dev_moving_gc_init(ca);
+	bch2_dev_copygc_init(ca);
 
 	INIT_WORK(&ca->io_error_work, bch2_io_error_work);
 
+	bch2_time_stats_init(&ca->io_latency[READ]);
+	bch2_time_stats_init(&ca->io_latency[WRITE]);
+
+	ca->mi = bch2_mi_to_cpu(member);
+	ca->uuid = member->uuid;
+
+	if (opt_defined(c->opts, discard))
+		ca->mi.discard = opt_get(c->opts, discard);
+
+	if (percpu_ref_init(&ca->ref, bch2_dev_ref_complete,
+			    0, GFP_KERNEL) ||
+	    percpu_ref_init(&ca->io_ref, bch2_dev_io_ref_complete,
+			    PERCPU_REF_INIT_DEAD, GFP_KERNEL) ||
+	    bch2_dev_buckets_alloc(c, ca) ||
+	    bioset_init(&ca->replica_set, 4,
+			offsetof(struct bch_write_bio, bio), 0) ||
+	    !(ca->io_done	= alloc_percpu(*ca->io_done)))
+		goto err;
+
+	return ca;
+err:
+	bch2_dev_free(ca);
+	return NULL;
+}
+
+static void bch2_dev_attach(struct bch_fs *c, struct bch_dev *ca,
+			    unsigned dev_idx)
+{
+	ca->dev_idx = dev_idx;
+	__set_bit(ca->dev_idx, ca->self.d);
+	scnprintf(ca->name, sizeof(ca->name), "dev-%u", dev_idx);
+
+	ca->fs = c;
+	rcu_assign_pointer(c->devs[ca->dev_idx], ca);
+
+	if (bch2_dev_sysfs_online(c, ca))
+		pr_warn("error creating sysfs objects");
+}
+
+static int bch2_dev_alloc(struct bch_fs *c, unsigned dev_idx)
+{
+	struct bch_member *member =
+		bch2_sb_get_members(c->disk_sb.sb)->members + dev_idx;
+	struct bch_dev *ca = NULL;
+	int ret = 0;
+
+	pr_verbose_init(c->opts, "");
+
 	if (bch2_fs_init_fault("dev_alloc"))
 		goto err;
 
-	member = bch2_sb_get_members(c->disk_sb)->members + dev_idx;
+	ca = __bch2_dev_alloc(c, member);
+	if (!ca)
+		goto err;
+
+	bch2_dev_attach(c, ca, dev_idx);
+out:
+	pr_verbose_init(c->opts, "ret %i", ret);
+	return ret;
+err:
+	if (ca)
+		bch2_dev_free(ca);
+	ret = -ENOMEM;
+	goto out;
+}
 
-	ca->mi = bch2_mi_to_cpu(member);
-	ca->uuid = member->uuid;
-	scnprintf(ca->name, sizeof(ca->name), "dev-%u", dev_idx);
+static int __bch2_dev_attach_bdev(struct bch_dev *ca, struct bch_sb_handle *sb)
+{
+	unsigned ret;
 
-	/* XXX: tune these */
-	movinggc_reserve = max_t(size_t, 16, ca->mi.nbuckets >> 7);
-	reserve_none = max_t(size_t, 4, ca->mi.nbuckets >> 9);
-	/*
-	 * free_inc must be smaller than the copygc reserve: if it was bigger,
-	 * one copygc iteration might not make enough buckets available to fill
-	 * up free_inc and allow the allocator to make forward progress
-	 */
-	free_inc_reserve = movinggc_reserve / 2;
-	heap_size = movinggc_reserve * 8;
+	if (bch2_dev_is_online(ca)) {
+		bch_err(ca, "already have device online in slot %u",
+			sb->sb->dev_idx);
+		return -EINVAL;
+	}
 
-	btree_node_reserve_buckets =
-		DIV_ROUND_UP(BTREE_NODE_RESERVE,
-			     ca->mi.bucket_size / c->sb.btree_node_size);
+	if (get_capacity(sb->bdev->bd_disk) <
+	    ca->mi.bucket_size * ca->mi.nbuckets) {
+		bch_err(ca, "cannot online: device too small");
+		return -EINVAL;
+	}
 
-	if (percpu_ref_init(&ca->ref, bch2_dev_ref_release,
-			    0, GFP_KERNEL) ||
-	    percpu_ref_init(&ca->io_ref, bch2_dev_io_ref_release,
-			    PERCPU_REF_INIT_DEAD, GFP_KERNEL) ||
-	    !init_fifo(&ca->free[RESERVE_BTREE], btree_node_reserve_buckets,
-		       GFP_KERNEL) ||
-	    !init_fifo(&ca->free[RESERVE_MOVINGGC],
-		       movinggc_reserve, GFP_KERNEL) ||
-	    !init_fifo(&ca->free[RESERVE_NONE], reserve_none, GFP_KERNEL) ||
-	    !init_fifo(&ca->free_inc,	free_inc_reserve, GFP_KERNEL) ||
-	    !init_heap(&ca->alloc_heap,	free_inc_reserve, GFP_KERNEL) ||
-	    !init_heap(&ca->copygc_heap,heap_size, GFP_KERNEL) ||
-	    !(ca->oldest_gens	= kvpmalloc(ca->mi.nbuckets *
-					    sizeof(u8),
-					    GFP_KERNEL|__GFP_ZERO)) ||
-	    !(ca->buckets	= kvpmalloc(ca->mi.nbuckets *
-					    sizeof(struct bucket),
-					    GFP_KERNEL|__GFP_ZERO)) ||
-	    !(ca->bucket_dirty	= kvpmalloc(BITS_TO_LONGS(ca->mi.nbuckets) *
-					    sizeof(unsigned long),
-					    GFP_KERNEL|__GFP_ZERO)) ||
-	    !(ca->usage_percpu = alloc_percpu(struct bch_dev_usage)) ||
-	    bioset_init(&ca->replica_set, 4,
-			offsetof(struct bch_write_bio, bio)) ||
-	    !(ca->io_done	= alloc_percpu(*ca->io_done)))
-		goto err;
+	BUG_ON(!percpu_ref_is_zero(&ca->io_ref));
 
-	total_reserve = ca->free_inc.size;
-	for (i = 0; i < RESERVE_NR; i++)
-		total_reserve += ca->free[i].size;
+	if (get_capacity(sb->bdev->bd_disk) <
+	    ca->mi.bucket_size * ca->mi.nbuckets) {
+		bch_err(ca, "device too small");
+		return -EINVAL;
+	}
 
-	ca->copygc_write_point.group = &ca->self;
+	ret = bch2_dev_journal_init(ca, sb->sb);
+	if (ret)
+		return ret;
 
-	ca->fs = c;
-	rcu_assign_pointer(c->devs[ca->dev_idx], ca);
+	/* Commit: */
+	ca->disk_sb = *sb;
+	if (sb->mode & FMODE_EXCL)
+		ca->disk_sb.bdev->bd_holder = ca;
+	memset(sb, 0, sizeof(*sb));
 
-	if (bch2_dev_sysfs_online(ca))
-		pr_warn("error creating sysfs objects");
+	percpu_ref_reinit(&ca->io_ref);
 
 	return 0;
-err:
-	bch2_dev_free(ca);
-	return -ENOMEM;
 }
 
-static int __bch2_dev_online(struct bch_fs *c, struct bcache_superblock *sb)
+static int bch2_dev_attach_bdev(struct bch_fs *c, struct bch_sb_handle *sb)
 {
 	struct bch_dev *ca;
 	int ret;
 
-	lockdep_assert_held(&c->sb_lock);
+	lockdep_assert_held(&c->state_lock);
 
 	if (le64_to_cpu(sb->sb->seq) >
-	    le64_to_cpu(c->disk_sb->seq))
+	    le64_to_cpu(c->disk_sb.sb->seq))
 		bch2_sb_to_fs(c, sb->sb);
 
 	BUG_ON(sb->sb->dev_idx >= c->sb.nr_devices ||
 	       !c->devs[sb->sb->dev_idx]);
 
-	ca = c->devs[sb->sb->dev_idx];
-	if (ca->disk_sb.bdev) {
-		bch_err(c, "already have device online in slot %u",
-			sb->sb->dev_idx);
-		return -EINVAL;
-	}
+	ca = bch_dev_locked(c, sb->sb->dev_idx);
 
-	ret = bch2_dev_journal_init(ca, sb->sb);
+	ret = __bch2_dev_attach_bdev(ca, sb);
 	if (ret)
 		return ret;
 
-	/*
-	 * Increase journal write timeout if flushes to this device are
-	 * expensive:
-	 */
-	if (!blk_queue_nonrot(bdev_get_queue(sb->bdev)) &&
-	    journal_flushes_device(ca))
-		c->journal.write_delay_ms =
-			max(c->journal.write_delay_ms, 1000U);
+	mutex_lock(&c->sb_lock);
+	bch2_mark_dev_superblock(ca->fs, ca, 0);
+	mutex_unlock(&c->sb_lock);
 
-	/* Commit: */
-	ca->disk_sb = *sb;
-	if (sb->mode & FMODE_EXCL)
-		ca->disk_sb.bdev->bd_holder = ca;
-	memset(sb, 0, sizeof(*sb));
+	bch2_dev_sysfs_online(c, ca);
 
 	if (c->sb.nr_devices == 1)
 		bdevname(ca->disk_sb.bdev, c->name);
 	bdevname(ca->disk_sb.bdev, ca->name);
 
-	if (bch2_dev_sysfs_online(ca))
-		pr_warn("error creating sysfs objects");
-
-	lg_local_lock(&c->usage_lock);
-	if (!gc_will_visit(c, gc_phase(GC_PHASE_SB_METADATA)))
-		bch2_mark_dev_metadata(c, ca);
-	lg_local_unlock(&c->usage_lock);
-
-	if (ca->mi.state == BCH_MEMBER_STATE_RW)
-		bch2_dev_allocator_add(c, ca);
-
-	percpu_ref_reinit(&ca->io_ref);
+	rebalance_wakeup(c);
 	return 0;
 }
 
@@ -1265,7 +1109,8 @@ bool bch2_dev_state_allowed(struct bch_fs *c, struct bch_dev *ca,
 
 		/* do we have enough devices to write to?  */
 		for_each_member_device(ca2, c, i)
-			nr_rw += ca2->mi.state == BCH_MEMBER_STATE_RW;
+			if (ca2 != ca)
+				nr_rw += ca2->mi.state == BCH_MEMBER_STATE_RW;
 
 		required = max(!(flags & BCH_FORCE_IF_METADATA_DEGRADED)
 			       ? c->opts.metadata_replicas
@@ -1274,7 +1119,7 @@ bool bch2_dev_state_allowed(struct bch_fs *c, struct bch_dev *ca,
 			       ? c->opts.data_replicas
 			       : c->opts.data_replicas_required);
 
-		return nr_rw - 1 <= required;
+		return nr_rw >= required;
 	case BCH_MEMBER_STATE_FAILED:
 	case BCH_MEMBER_STATE_SPARE:
 		if (ca->mi.state != BCH_MEMBER_STATE_RW &&
@@ -1287,7 +1132,7 @@ bool bch2_dev_state_allowed(struct bch_fs *c, struct bch_dev *ca,
 
 		s = __bch2_replicas_status(c, new_online_devs);
 
-		return bch2_have_enough_devs(c, s, flags);
+		return bch2_have_enough_devs(s, flags);
 	default:
 		BUG();
 	}
@@ -1297,41 +1142,46 @@ static bool bch2_fs_may_start(struct bch_fs *c)
 {
 	struct replicas_status s;
 	struct bch_sb_field_members *mi;
+	struct bch_dev *ca;
 	unsigned i, flags = c->opts.degraded
 		? BCH_FORCE_IF_DEGRADED
 		: 0;
 
 	if (!c->opts.degraded) {
 		mutex_lock(&c->sb_lock);
-		mi = bch2_sb_get_members(c->disk_sb);
+		mi = bch2_sb_get_members(c->disk_sb.sb);
 
-		for (i = 0; i < c->disk_sb->nr_devices; i++)
-			if (bch2_dev_exists(c->disk_sb, mi, i) &&
-			    !bch2_dev_is_online(c->devs[i]) &&
-			    (c->devs[i]->mi.state == BCH_MEMBER_STATE_RW ||
-			     c->devs[i]->mi.state == BCH_MEMBER_STATE_RO)) {
+		for (i = 0; i < c->disk_sb.sb->nr_devices; i++) {
+			if (!bch2_dev_exists(c->disk_sb.sb, mi, i))
+				continue;
+
+			ca = bch_dev_locked(c, i);
+
+			if (!bch2_dev_is_online(ca) &&
+			    (ca->mi.state == BCH_MEMBER_STATE_RW ||
+			     ca->mi.state == BCH_MEMBER_STATE_RO)) {
 				mutex_unlock(&c->sb_lock);
 				return false;
 			}
+		}
 		mutex_unlock(&c->sb_lock);
 	}
 
 	s = bch2_replicas_status(c);
 
-	return bch2_have_enough_devs(c, s, flags);
+	return bch2_have_enough_devs(s, flags);
 }
 
 static void __bch2_dev_read_only(struct bch_fs *c, struct bch_dev *ca)
 {
-	bch2_moving_gc_stop(ca);
+	bch2_copygc_stop(ca);
 
 	/*
-	 * This stops new data writes (e.g. to existing open data
-	 * buckets) and then waits for all existing writes to
-	 * complete.
+	 * The allocator thread itself allocates btree nodes, so stop it first:
 	 */
 	bch2_dev_allocator_stop(ca);
 	bch2_dev_allocator_remove(c, ca);
+	bch2_dev_journal_stop(&c->journal, ca);
 }
 
 static const char *__bch2_dev_read_write(struct bch_fs *c, struct bch_dev *ca)
@@ -1346,11 +1196,8 @@ static const char *__bch2_dev_read_write(struct bch_fs *c, struct bch_dev *ca)
 	if (bch2_dev_allocator_start(ca))
 		return "error starting allocator thread";
 
-	if (bch2_moving_gc_start(ca))
-		return "error starting moving GC thread";
-
-	if (bch2_tiering_start(c))
-		return "error starting tiering thread";
+	if (bch2_copygc_start(c, ca))
+		return "error starting copygc thread";
 
 	return NULL;
 }
@@ -1359,6 +1206,7 @@ int __bch2_dev_set_state(struct bch_fs *c, struct bch_dev *ca,
 			 enum bch_member_state new_state, int flags)
 {
 	struct bch_sb_field_members *mi;
+	int ret = 0;
 
 	if (ca->mi.state == new_state)
 		return 0;
@@ -1366,22 +1214,24 @@ int __bch2_dev_set_state(struct bch_fs *c, struct bch_dev *ca,
 	if (!bch2_dev_state_allowed(c, ca, new_state, flags))
 		return -EINVAL;
 
-	if (new_state == BCH_MEMBER_STATE_RW) {
-		if (__bch2_dev_read_write(c, ca))
-			return -ENOMEM;
-	} else {
+	if (new_state != BCH_MEMBER_STATE_RW)
 		__bch2_dev_read_only(c, ca);
-	}
 
 	bch_notice(ca, "%s", bch2_dev_state[new_state]);
 
 	mutex_lock(&c->sb_lock);
-	mi = bch2_sb_get_members(c->disk_sb);
+	mi = bch2_sb_get_members(c->disk_sb.sb);
 	SET_BCH_MEMBER_STATE(&mi->members[ca->dev_idx], new_state);
 	bch2_write_super(c);
 	mutex_unlock(&c->sb_lock);
 
-	return 0;
+	if (new_state == BCH_MEMBER_STATE_RW &&
+	    __bch2_dev_read_write(c, ca))
+		ret = -ENOMEM;
+
+	rebalance_wakeup(c);
+
+	return ret;
 }
 
 int bch2_dev_set_state(struct bch_fs *c, struct bch_dev *ca,
@@ -1408,37 +1258,70 @@ int bch2_dev_remove(struct bch_fs *c, struct bch_dev *ca, int flags)
 
 	percpu_ref_put(&ca->ref); /* XXX */
 
-	if (ca->mi.state == BCH_MEMBER_STATE_RW) {
-		bch_err(ca, "Cannot remove RW device");
-		goto err;
-	}
-
 	if (!bch2_dev_state_allowed(c, ca, BCH_MEMBER_STATE_FAILED, flags)) {
 		bch_err(ca, "Cannot remove without losing data");
 		goto err;
 	}
 
+	__bch2_dev_read_only(c, ca);
+
 	/*
 	 * XXX: verify that dev_idx is really not in use anymore, anywhere
 	 *
 	 * flag_data_bad() does not check btree pointers
 	 */
-	ret = bch2_flag_data_bad(ca);
+	ret = bch2_dev_data_drop(c, ca->dev_idx, flags);
+	if (ret) {
+		bch_err(ca, "Remove failed: error %i dropping data", ret);
+		goto err;
+	}
+
+	ret = bch2_journal_flush_device_pins(&c->journal, ca->dev_idx);
 	if (ret) {
-		bch_err(ca, "Remove failed");
+		bch_err(ca, "Remove failed: error %i flushing journal", ret);
 		goto err;
 	}
 
 	data = bch2_dev_has_data(c, ca);
 	if (data) {
-		bch_err(ca, "Remove failed, still has data (%x)", data);
+		char data_has_str[100];
+
+		bch2_string_opt_to_text(&PBUF(data_has_str),
+					bch2_data_types, data);
+		bch_err(ca, "Remove failed, still has data (%s)", data_has_str);
+		ret = -EBUSY;
+		goto err;
+	}
+
+	ret = bch2_btree_delete_range(c, BTREE_ID_ALLOC,
+				      POS(ca->dev_idx, 0),
+				      POS(ca->dev_idx + 1, 0),
+				      NULL);
+	if (ret) {
+		bch_err(ca, "Remove failed, error deleting alloc info");
 		goto err;
 	}
 
-	bch2_journal_meta(&c->journal);
+	/*
+	 * must flush all existing journal entries, they might have
+	 * (overwritten) keys that point to the device we're removing:
+	 */
+	bch2_journal_flush_all_pins(&c->journal);
+	ret = bch2_journal_error(&c->journal);
+	if (ret) {
+		bch_err(ca, "Remove failed, journal error");
+		goto err;
+	}
+
+	__bch2_dev_offline(c, ca);
+
+	mutex_lock(&c->sb_lock);
+	rcu_assign_pointer(c->devs[ca->dev_idx], NULL);
+	mutex_unlock(&c->sb_lock);
+
+	percpu_ref_kill(&ca->ref);
+	wait_for_completion(&ca->ref_completion);
 
-	__bch2_dev_offline(ca);
-	bch2_dev_stop(ca);
 	bch2_dev_free(ca);
 
 	/*
@@ -1446,7 +1329,7 @@ int bch2_dev_remove(struct bch_fs *c, struct bch_dev *ca, int flags)
 	 * this device must be gone:
 	 */
 	mutex_lock(&c->sb_lock);
-	mi = bch2_sb_get_members(c->disk_sb);
+	mi = bch2_sb_get_members(c->disk_sb.sb);
 	memset(&mi->members[dev_idx].uuid, 0, sizeof(mi->members[dev_idx].uuid));
 
 	bch2_write_super(c);
@@ -1455,50 +1338,112 @@ int bch2_dev_remove(struct bch_fs *c, struct bch_dev *ca, int flags)
 	mutex_unlock(&c->state_lock);
 	return 0;
 err:
+	if (ca->mi.state == BCH_MEMBER_STATE_RW)
+		__bch2_dev_read_write(c, ca);
 	mutex_unlock(&c->state_lock);
 	return ret;
 }
 
+static void dev_usage_clear(struct bch_dev *ca)
+{
+	struct bucket_array *buckets;
+	int cpu;
+
+	for_each_possible_cpu(cpu) {
+		struct bch_dev_usage *p =
+			per_cpu_ptr(ca->usage[0], cpu);
+		memset(p, 0, sizeof(*p));
+	}
+
+	down_read(&ca->bucket_lock);
+	buckets = bucket_array(ca);
+
+	memset(buckets->b, 0, sizeof(buckets->b[0]) * buckets->nbuckets);
+	up_read(&ca->bucket_lock);
+}
+
 /* Add new device to running filesystem: */
 int bch2_dev_add(struct bch_fs *c, const char *path)
 {
-	struct bcache_superblock sb;
+	struct bch_opts opts = bch2_opts_empty();
+	struct bch_sb_handle sb;
 	const char *err;
 	struct bch_dev *ca = NULL;
-	struct bch_sb_field_members *mi, *dev_mi;
-	struct bch_member saved_mi;
+	struct bch_sb_field_members *mi;
+	struct bch_member dev_mi;
 	unsigned dev_idx, nr_devices, u64s;
-	int ret = -EINVAL;
+	int ret;
 
-	err = bch2_read_super(&sb, bch2_opts_empty(), path);
-	if (err)
-		return -EINVAL;
+	ret = bch2_read_super(path, &opts, &sb);
+	if (ret)
+		return ret;
 
 	err = bch2_sb_validate(&sb);
 	if (err)
 		return -EINVAL;
 
+	dev_mi = bch2_sb_get_members(sb.sb)->members[sb.sb->dev_idx];
+
 	err = bch2_dev_may_add(sb.sb, c);
 	if (err)
 		return -EINVAL;
 
-	mutex_lock(&c->state_lock);
-	mutex_lock(&c->sb_lock);
+	ca = __bch2_dev_alloc(c, &dev_mi);
+	if (!ca) {
+		bch2_free_super(&sb);
+		return -ENOMEM;
+	}
+
+	ret = __bch2_dev_attach_bdev(ca, &sb);
+	if (ret) {
+		bch2_dev_free(ca);
+		return ret;
+	}
 
 	/*
-	 * Preserve the old cache member information (esp. tier)
-	 * before we start bashing the disk stuff.
+	 * We want to allocate journal on the new device before adding the new
+	 * device to the filesystem because allocating after we attach requires
+	 * spinning up the allocator thread, and the allocator thread requires
+	 * doing btree writes, which if the existing devices are RO isn't going
+	 * to work
+	 *
+	 * So we have to mark where the superblocks are, but marking allocated
+	 * data normally updates the filesystem usage too, so we have to mark,
+	 * allocate the journal, reset all the marks, then remark after we
+	 * attach...
 	 */
-	dev_mi = bch2_sb_get_members(sb.sb);
-	saved_mi = dev_mi->members[sb.sb->dev_idx];
-	saved_mi.last_mount = cpu_to_le64(ktime_get_seconds());
+	bch2_mark_dev_superblock(ca->fs, ca, 0);
+
+	err = "journal alloc failed";
+	ret = bch2_dev_journal_alloc(ca);
+	if (ret)
+		goto err;
+
+	dev_usage_clear(ca);
+
+	mutex_lock(&c->state_lock);
+	mutex_lock(&c->sb_lock);
+
+	err = "insufficient space in new superblock";
+	ret = bch2_sb_from_fs(c, ca);
+	if (ret)
+		goto err_unlock;
+
+	mi = bch2_sb_get_members(ca->disk_sb.sb);
+
+	if (!bch2_sb_resize_members(&ca->disk_sb,
+				le32_to_cpu(mi->field.u64s) +
+				sizeof(dev_mi) / sizeof(u64))) {
+		ret = -ENOSPC;
+		goto err_unlock;
+	}
 
 	if (dynamic_fault("bcachefs:add:no_slot"))
 		goto no_slot;
 
-	mi = bch2_sb_get_members(c->disk_sb);
+	mi = bch2_sb_get_members(c->disk_sb.sb);
 	for (dev_idx = 0; dev_idx < BCH_SB_MEMBERS_MAX; dev_idx++)
-		if (!bch2_dev_exists(c->disk_sb, mi, dev_idx))
+		if (!bch2_dev_exists(c->disk_sb.sb, mi, dev_idx))
 			goto have_slot;
 no_slot:
 	err = "no slots available in superblock";
@@ -1509,101 +1454,97 @@ have_slot:
 	nr_devices = max_t(unsigned, dev_idx + 1, c->sb.nr_devices);
 	u64s = (sizeof(struct bch_sb_field_members) +
 		sizeof(struct bch_member) * nr_devices) / sizeof(u64);
+
 	err = "no space in superblock for member info";
+	ret = -ENOSPC;
 
-	mi = bch2_fs_sb_resize_members(c, u64s);
+	mi = bch2_sb_resize_members(&c->disk_sb, u64s);
 	if (!mi)
 		goto err_unlock;
 
-	dev_mi = bch2_sb_resize_members(&sb, u64s);
-	if (!dev_mi)
-		goto err_unlock;
-
-	memcpy(dev_mi, mi, u64s * sizeof(u64));
-	dev_mi->members[dev_idx] = saved_mi;
+	/* success: */
 
-	sb.sb->uuid		= c->disk_sb->uuid;
-	sb.sb->dev_idx		= dev_idx;
-	sb.sb->nr_devices	= nr_devices;
+	mi->members[dev_idx] = dev_mi;
+	mi->members[dev_idx].last_mount = cpu_to_le64(ktime_get_real_seconds());
+	c->disk_sb.sb->nr_devices	= nr_devices;
 
-	/* commit new member info */
-	memcpy(mi, dev_mi, u64s * sizeof(u64));
-	c->disk_sb->nr_devices	= nr_devices;
-	c->sb.nr_devices	= nr_devices;
+	ca->disk_sb.sb->dev_idx	= dev_idx;
+	bch2_dev_attach(c, ca, dev_idx);
 
-	if (bch2_dev_alloc(c, dev_idx)) {
-		err = "cannot allocate memory";
-		ret = -ENOMEM;
-		goto err_unlock;
-	}
-
-	if (__bch2_dev_online(c, &sb)) {
-		err = "bch2_dev_online() error";
-		ret = -ENOMEM;
-		goto err_unlock;
-	}
+	bch2_mark_dev_superblock(c, ca, 0);
 
 	bch2_write_super(c);
 	mutex_unlock(&c->sb_lock);
 
-	ca = c->devs[dev_idx];
 	if (ca->mi.state == BCH_MEMBER_STATE_RW) {
-		err = "journal alloc failed";
-		if (bch2_dev_journal_alloc(ca))
-			goto err;
-
 		err = __bch2_dev_read_write(c, ca);
 		if (err)
-			goto err;
+			goto err_late;
 	}
 
 	mutex_unlock(&c->state_lock);
 	return 0;
+
 err_unlock:
 	mutex_unlock(&c->sb_lock);
-err:
 	mutex_unlock(&c->state_lock);
+err:
+	if (ca)
+		bch2_dev_free(ca);
 	bch2_free_super(&sb);
-
 	bch_err(c, "Unable to add device: %s", err);
-	return ret ?: -EINVAL;
+	return ret;
+err_late:
+	bch_err(c, "Error going rw after adding device: %s", err);
+	return -EINVAL;
 }
 
 /* Hot add existing device to running filesystem: */
 int bch2_dev_online(struct bch_fs *c, const char *path)
 {
-	struct bcache_superblock sb = { 0 };
+	struct bch_opts opts = bch2_opts_empty();
+	struct bch_sb_handle sb = { NULL };
+	struct bch_sb_field_members *mi;
 	struct bch_dev *ca;
 	unsigned dev_idx;
 	const char *err;
+	int ret;
 
 	mutex_lock(&c->state_lock);
 
-	err = bch2_read_super(&sb, bch2_opts_empty(), path);
-	if (err)
-		goto err;
+	ret = bch2_read_super(path, &opts, &sb);
+	if (ret) {
+		mutex_unlock(&c->state_lock);
+		return ret;
+	}
 
 	dev_idx = sb.sb->dev_idx;
 
-	err = bch2_dev_in_fs(c->disk_sb, sb.sb);
+	err = bch2_dev_in_fs(c->disk_sb.sb, sb.sb);
 	if (err)
 		goto err;
 
-	mutex_lock(&c->sb_lock);
-	if (__bch2_dev_online(c, &sb)) {
-		err = "__bch2_dev_online() error";
-		mutex_unlock(&c->sb_lock);
+	if (bch2_dev_attach_bdev(c, &sb)) {
+		err = "bch2_dev_attach_bdev() error";
 		goto err;
 	}
-	mutex_unlock(&c->sb_lock);
 
-	ca = c->devs[dev_idx];
+	ca = bch_dev_locked(c, dev_idx);
 	if (ca->mi.state == BCH_MEMBER_STATE_RW) {
 		err = __bch2_dev_read_write(c, ca);
 		if (err)
 			goto err;
 	}
 
+	mutex_lock(&c->sb_lock);
+	mi = bch2_sb_get_members(c->disk_sb.sb);
+
+	mi->members[ca->dev_idx].last_mount =
+		cpu_to_le64(ktime_get_real_seconds());
+
+	bch2_write_super(c);
+	mutex_unlock(&c->sb_lock);
+
 	mutex_unlock(&c->state_lock);
 	return 0;
 err:
@@ -1617,84 +1558,120 @@ int bch2_dev_offline(struct bch_fs *c, struct bch_dev *ca, int flags)
 {
 	mutex_lock(&c->state_lock);
 
+	if (!bch2_dev_is_online(ca)) {
+		bch_err(ca, "Already offline");
+		mutex_unlock(&c->state_lock);
+		return 0;
+	}
+
 	if (!bch2_dev_state_allowed(c, ca, BCH_MEMBER_STATE_FAILED, flags)) {
 		bch_err(ca, "Cannot offline required disk");
 		mutex_unlock(&c->state_lock);
 		return -EINVAL;
 	}
 
-	__bch2_dev_read_only(c, ca);
-	__bch2_dev_offline(ca);
+	__bch2_dev_offline(c, ca);
 
 	mutex_unlock(&c->state_lock);
 	return 0;
 }
 
-int bch2_dev_evacuate(struct bch_fs *c, struct bch_dev *ca)
+int bch2_dev_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
 {
-	unsigned data;
-	int ret;
+	struct bch_member *mi;
+	int ret = 0;
 
 	mutex_lock(&c->state_lock);
 
-	if (ca->mi.state == BCH_MEMBER_STATE_RW) {
-		bch_err(ca, "Cannot migrate data off RW device");
-		mutex_unlock(&c->state_lock);
-		return -EINVAL;
+	if (nbuckets < ca->mi.nbuckets) {
+		bch_err(ca, "Cannot shrink yet");
+		ret = -EINVAL;
+		goto err;
 	}
 
-	mutex_unlock(&c->state_lock);
-
-	ret = bch2_move_data_off_device(ca);
-	if (ret) {
-		bch_err(ca, "Error migrating data: %i", ret);
-		return ret;
+	if (bch2_dev_is_online(ca) &&
+	    get_capacity(ca->disk_sb.bdev->bd_disk) <
+	    ca->mi.bucket_size * nbuckets) {
+		bch_err(ca, "New size larger than device");
+		ret = -EINVAL;
+		goto err;
 	}
 
-	ret = bch2_move_metadata_off_device(ca);
+	ret = bch2_dev_buckets_resize(c, ca, nbuckets);
 	if (ret) {
-		bch_err(ca, "Error migrating metadata: %i", ret);
-		return ret;
+		bch_err(ca, "Resize error: %i", ret);
+		goto err;
 	}
 
-	data = bch2_dev_has_data(c, ca);
-	if (data) {
-		bch_err(ca, "Migrate error: data still present (%x)", data);
-		return -EINVAL;
-	}
+	mutex_lock(&c->sb_lock);
+	mi = &bch2_sb_get_members(c->disk_sb.sb)->members[ca->dev_idx];
+	mi->nbuckets = cpu_to_le64(nbuckets);
 
-	return 0;
+	bch2_write_super(c);
+	mutex_unlock(&c->sb_lock);
+
+	bch2_recalc_capacity(c);
+err:
+	mutex_unlock(&c->state_lock);
+	return ret;
+}
+
+/* return with ref on ca->ref: */
+struct bch_dev *bch2_dev_lookup(struct bch_fs *c, const char *path)
+{
+
+	struct block_device *bdev = lookup_bdev(path);
+	struct bch_dev *ca;
+	unsigned i;
+
+	if (IS_ERR(bdev))
+		return ERR_CAST(bdev);
+
+	for_each_member_device(ca, c, i)
+		if (ca->disk_sb.bdev == bdev)
+			goto found;
+
+	ca = ERR_PTR(-ENOENT);
+found:
+	bdput(bdev);
+	return ca;
 }
 
 /* Filesystem open: */
 
-const char *bch2_fs_open(char * const *devices, unsigned nr_devices,
-			 struct bch_opts opts, struct bch_fs **ret)
+struct bch_fs *bch2_fs_open(char * const *devices, unsigned nr_devices,
+			    struct bch_opts opts)
 {
-	const char *err;
+	struct bch_sb_handle *sb = NULL;
 	struct bch_fs *c = NULL;
-	struct bcache_superblock *sb;
 	unsigned i, best_sb = 0;
+	const char *err;
+	int ret = -ENOMEM;
+
+	pr_verbose_init(opts, "");
 
-	if (!nr_devices)
-		return "need at least one device";
+	if (!nr_devices) {
+		c = ERR_PTR(-EINVAL);
+		goto out2;
+	}
 
-	if (!try_module_get(THIS_MODULE))
-		return "module unloading";
+	if (!try_module_get(THIS_MODULE)) {
+		c = ERR_PTR(-ENODEV);
+		goto out2;
+	}
 
-	err = "cannot allocate memory";
 	sb = kcalloc(nr_devices, sizeof(*sb), GFP_KERNEL);
 	if (!sb)
 		goto err;
 
 	for (i = 0; i < nr_devices; i++) {
-		err = bch2_read_super(&sb[i], opts, devices[i]);
-		if (err)
+		ret = bch2_read_super(devices[i], &opts, &sb[i]);
+		if (ret)
 			goto err;
 
 		err = bch2_sb_validate(&sb[i]);
 		if (err)
-			goto err;
+			goto err_print;
 	}
 
 	for (i = 1; i < nr_devices; i++)
@@ -1705,59 +1682,52 @@ const char *bch2_fs_open(char * const *devices, unsigned nr_devices,
 	for (i = 0; i < nr_devices; i++) {
 		err = bch2_dev_in_fs(sb[best_sb].sb, sb[i].sb);
 		if (err)
-			goto err;
+			goto err_print;
 	}
 
-	err = "cannot allocate memory";
+	ret = -ENOMEM;
 	c = bch2_fs_alloc(sb[best_sb].sb, opts);
 	if (!c)
 		goto err;
 
 	err = "bch2_dev_online() error";
-	mutex_lock(&c->sb_lock);
+	mutex_lock(&c->state_lock);
 	for (i = 0; i < nr_devices; i++)
-		if (__bch2_dev_online(c, &sb[i])) {
-			mutex_unlock(&c->sb_lock);
-			goto err;
+		if (bch2_dev_attach_bdev(c, &sb[i])) {
+			mutex_unlock(&c->state_lock);
+			goto err_print;
 		}
-	mutex_unlock(&c->sb_lock);
+	mutex_unlock(&c->state_lock);
 
 	err = "insufficient devices";
 	if (!bch2_fs_may_start(c))
-		goto err;
+		goto err_print;
 
 	if (!c->opts.nostart) {
-		err = __bch2_fs_start(c);
+		err = bch2_fs_start(c);
 		if (err)
-			goto err;
+			goto err_print;
 	}
-
-	err = bch2_fs_online(c);
-	if (err)
-		goto err;
-
-	if (ret)
-		*ret = c;
-	else
-		closure_put(&c->cl);
-
-	err = NULL;
 out:
 	kfree(sb);
 	module_put(THIS_MODULE);
-	if (err)
-		c = NULL;
-	return err;
+out2:
+	pr_verbose_init(opts, "ret %i", PTR_ERR_OR_ZERO(c));
+	return c;
+err_print:
+	pr_err("bch_fs_open err opening %s: %s",
+	       devices[0], err);
+	ret = -EINVAL;
 err:
 	if (c)
 		bch2_fs_stop(c);
-
 	for (i = 0; i < nr_devices; i++)
 		bch2_free_super(&sb[i]);
+	c = ERR_PTR(ret);
 	goto out;
 }
 
-static const char *__bch2_fs_open_incremental(struct bcache_superblock *sb,
+static const char *__bch2_fs_open_incremental(struct bch_sb_handle *sb,
 					      struct bch_opts opts)
 {
 	const char *err;
@@ -1773,7 +1743,7 @@ static const char *__bch2_fs_open_incremental(struct bcache_superblock *sb,
 	if (c) {
 		closure_get(&c->cl);
 
-		err = bch2_dev_in_fs(c->disk_sb, sb->sb);
+		err = bch2_dev_in_fs(c->disk_sb.sb, sb->sb);
 		if (err)
 			goto err;
 	} else {
@@ -1788,22 +1758,18 @@ static const char *__bch2_fs_open_incremental(struct bcache_superblock *sb,
 	err = "bch2_dev_online() error";
 
 	mutex_lock(&c->sb_lock);
-	if (__bch2_dev_online(c, sb)) {
+	if (bch2_dev_attach_bdev(c, sb)) {
 		mutex_unlock(&c->sb_lock);
 		goto err;
 	}
 	mutex_unlock(&c->sb_lock);
 
 	if (!c->opts.nostart && bch2_fs_may_start(c)) {
-		err = __bch2_fs_start(c);
+		err = bch2_fs_start(c);
 		if (err)
 			goto err;
 	}
 
-	err = __bch2_fs_online(c);
-	if (err)
-		goto err;
-
 	closure_put(&c->cl);
 	mutex_unlock(&bch_fs_list_lock);
 
@@ -1821,13 +1787,12 @@ err:
 
 const char *bch2_fs_open_incremental(const char *path)
 {
-	struct bcache_superblock sb;
+	struct bch_sb_handle sb;
 	struct bch_opts opts = bch2_opts_empty();
 	const char *err;
 
-	err = bch2_read_super(&sb, opts, path);
-	if (err)
-		return err;
+	if (bch2_read_super(path, &opts, &sb))
+		return "error reading superblock";
 
 	err = __bch2_fs_open_incremental(&sb, opts);
 	bch2_free_super(&sb);