X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Fsuper.c;h=1feb7dee2e0c1ca744683cb9158f8596fdcb4f32;hb=e61b61c03bf1f1eedc5e2dbd6887f77e45144a31;hp=29ffba65cfc55f4d99ba36f6ab492e9d5622ec36;hpb=4de98a2712764bceb9e0f67b1ac2f2c7862feb77;p=bcachefs-tools-debian

diff --git a/libbcachefs/super.c b/libbcachefs/super.c
index 29ffba6..1feb7de 100644
--- a/libbcachefs/super.c
+++ b/libbcachefs/super.c
@@ -1,3 +1,4 @@
+// SPDX-License-Identifier: GPL-2.0
 /*
  * bcachefs setup/teardown code, and some metadata io - read a superblock and
  * figure out what to do with it.
@@ -7,10 +8,12 @@
  */
 
 #include "bcachefs.h"
-#include "alloc.h"
+#include "alloc_background.h"
+#include "alloc_foreground.h"
+#include "bkey_sort.h"
 #include "btree_cache.h"
 #include "btree_gc.h"
-#include "btree_update.h"
+#include "btree_key_cache.h"
 #include "btree_update_interior.h"
 #include "btree_io.h"
 #include "chardev.h"
@@ -18,6 +21,8 @@
 #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"
@@ -25,15 +30,19 @@
 #include "inode.h"
 #include "io.h"
 #include "journal.h"
-#include "keylist.h"
+#include "journal_reclaim.h"
+#include "journal_seq_blacklist.h"
 #include "move.h"
 #include "migrate.h"
 #include "movinggc.h"
 #include "quota.h"
+#include "rebalance.h"
+#include "recovery.h"
+#include "replicas.h"
+#include "subvolume.h"
 #include "super.h"
 #include "super-io.h"
 #include "sysfs.h"
-#include "tier.h"
 
 #include <linux/backing-dev.h>
 #include <linux/blkdev.h>
@@ -41,7 +50,6 @@
 #include <linux/device.h>
 #include <linux/genhd.h>
 #include <linux/idr.h>
-#include <linux/kthread.h>
 #include <linux/module.h>
 #include <linux/percpu.h>
 #include <linux/random.h>
@@ -92,7 +100,7 @@ static int bch2_dev_alloc(struct bch_fs *, unsigned);
 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)
+struct bch_fs *bch2_dev_to_fs(dev_t dev)
 {
 	struct bch_fs *c;
 	struct bch_dev *ca;
@@ -103,7 +111,7 @@ struct bch_fs *bch2_bdev_to_fs(struct block_device *bdev)
 
 	list_for_each_entry(c, &bch_fs_list, list)
 		for_each_member_device_rcu(ca, c, i, NULL)
-			if (ca->disk_sb.bdev == bdev) {
+			if (ca->disk_sb.bdev && ca->disk_sb.bdev->bd_dev == dev) {
 				closure_get(&c->cl);
 				goto found;
 			}
@@ -122,7 +130,7 @@ 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;
@@ -141,43 +149,21 @@ struct bch_fs *bch2_uuid_to_fs(uuid_le uuid)
 	return c;
 }
 
-int bch2_congested(void *data, int bdi_bits)
+static void bch2_dev_usage_journal_reserve(struct bch_fs *c)
 {
-	struct bch_fs *c = data;
-	struct backing_dev_info *bdi;
 	struct bch_dev *ca;
-	unsigned i;
-	int ret = 0;
+	unsigned i, nr = 0, u64s =
+		((sizeof(struct jset_entry_dev_usage) +
+		  sizeof(struct jset_entry_dev_usage_type) * BCH_DATA_NR)) /
+		sizeof(u64);
 
-	if (bdi_bits & (1 << WB_sync_congested)) {
-		/* Reads - check all devices: */
-		for_each_readable_member(ca, c, i) {
-			bdi = ca->disk_sb.bdev->bd_bdi;
-
-			if (bdi_congested(bdi, bdi_bits)) {
-				ret = 1;
-				break;
-			}
-		}
-	} 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];
-
-		rcu_read_lock();
-		for_each_member_device_rcu(ca, c, i, devs) {
-			bdi = ca->disk_sb.bdev->bd_bdi;
-
-			if (bdi_congested(bdi, bdi_bits)) {
-				ret = 1;
-				break;
-			}
-		}
-		rcu_read_unlock();
-	}
+	rcu_read_lock();
+	for_each_member_device_rcu(ca, c, i, NULL)
+		nr++;
+	rcu_read_unlock();
 
-	return ret;
+	bch2_journal_entry_res_resize(&c->journal,
+			&c->dev_usage_journal_res, u64s * nr);
 }
 
 /* Filesystem RO/RW: */
@@ -185,9 +171,9 @@ int bch2_congested(void *data, int bdi_bits)
 /*
  * 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)
@@ -200,13 +186,10 @@ int bch2_congested(void *data, int bdi_bits)
 static void __bch2_fs_read_only(struct bch_fs *c)
 {
 	struct bch_dev *ca;
-	unsigned i;
-
-	bch2_tiering_stop(c);
-
-	for_each_member_device(ca, c, i)
-		bch2_copygc_stop(ca);
+	unsigned i, clean_passes = 0;
 
+	bch2_rebalance_stop(c);
+	bch2_copygc_stop(c);
 	bch2_gc_thread_stop(c);
 
 	/*
@@ -215,15 +198,63 @@ static void __bch2_fs_read_only(struct bch_fs *c)
 	 */
 	bch2_journal_flush_all_pins(&c->journal);
 
+	/*
+	 * If the allocator threads didn't all start up, the btree updates to
+	 * write out alloc info aren't going to work:
+	 */
+	if (!test_bit(BCH_FS_ALLOCATOR_RUNNING, &c->flags))
+		goto nowrote_alloc;
+
+	bch_verbose(c, "flushing journal and stopping allocators");
+
+	bch2_journal_flush_all_pins(&c->journal);
+	set_bit(BCH_FS_ALLOCATOR_STOPPING, &c->flags);
+
+	do {
+		clean_passes++;
+
+		if (bch2_journal_flush_all_pins(&c->journal))
+			clean_passes = 0;
+
+		/*
+		 * In flight interior btree updates will generate more journal
+		 * updates and btree updates (alloc btree):
+		 */
+		if (bch2_btree_interior_updates_nr_pending(c)) {
+			closure_wait_event(&c->btree_interior_update_wait,
+					   !bch2_btree_interior_updates_nr_pending(c));
+			clean_passes = 0;
+		}
+		flush_work(&c->btree_interior_update_work);
+
+		if (bch2_journal_flush_all_pins(&c->journal))
+			clean_passes = 0;
+	} while (clean_passes < 2);
+	bch_verbose(c, "flushing journal and stopping allocators complete");
+
+	set_bit(BCH_FS_ALLOC_CLEAN, &c->flags);
+nowrote_alloc:
+	closure_wait_event(&c->btree_interior_update_wait,
+			   !bch2_btree_interior_updates_nr_pending(c));
+	flush_work(&c->btree_interior_update_work);
+
 	for_each_member_device(ca, c, i)
 		bch2_dev_allocator_stop(ca);
 
+	clear_bit(BCH_FS_ALLOCATOR_RUNNING, &c->flags);
+	clear_bit(BCH_FS_ALLOCATOR_STOPPING, &c->flags);
+
 	bch2_fs_journal_stop(&c->journal);
 
-	if (!bch2_journal_error(&c->journal) &&
-	    !test_bit(BCH_FS_ERROR, &c->flags))
-		bch2_btree_verify_flushed(c);
+	/*
+	 * the journal kicks off btree writes via reclaim - wait for in flight
+	 * writes after stopping journal:
+	 */
+	bch2_btree_flush_all_writes(c);
 
+	/*
+	 * After stopping journal:
+	 */
 	for_each_member_device(ca, c, i)
 		bch2_dev_allocator_remove(c, ca);
 }
@@ -238,12 +269,12 @@ static void bch2_writes_disabled(struct percpu_ref *writes)
 
 void bch2_fs_read_only(struct bch_fs *c)
 {
-	if (c->state != BCH_FS_STARTING &&
-	    c->state != BCH_FS_RW)
+	if (!test_bit(BCH_FS_RW, &c->flags)) {
+		bch2_journal_reclaim_stop(&c->journal);
 		return;
+	}
 
-	if (test_bit(BCH_FS_ERROR, &c->flags))
-		return;
+	BUG_ON(test_bit(BCH_FS_WRITE_DISABLE_COMPLETE, &c->flags));
 
 	/*
 	 * Block new foreground-end write operations from starting - any new
@@ -255,7 +286,7 @@ void bch2_fs_read_only(struct bch_fs *c)
 	 */
 	percpu_ref_kill(&c->writes);
 
-	cancel_delayed_work(&c->pd_controllers_update);
+	cancel_work_sync(&c->ec_stripe_delete_work);
 
 	/*
 	 * If we're not doing an emergency shutdown, we want to wait on
@@ -280,14 +311,16 @@ 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) &&
+	    test_bit(BCH_FS_STARTED, &c->flags) &&
+	    test_bit(BCH_FS_ALLOC_CLEAN, &c->flags) &&
+	    !c->opts.norecovery) {
+		bch_verbose(c, "marking filesystem clean");
+		bch2_fs_mark_clean(c);
 	}
 
-	c->state = BCH_FS_RO;
+	clear_bit(BCH_FS_RW, &c->flags);
 }
 
 static void bch2_fs_read_only_work(struct work_struct *work)
@@ -295,9 +328,9 @@ 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);
+	down_write(&c->state_lock);
 	bch2_fs_read_only(c);
-	mutex_unlock(&c->state_lock);
+	up_write(&c->state_lock);
 }
 
 static void bch2_fs_read_only_async(struct bch_fs *c)
@@ -309,126 +342,209 @@ bool bch2_fs_emergency_read_only(struct bch_fs *c)
 {
 	bool ret = !test_and_set_bit(BCH_FS_EMERGENCY_RO, &c->flags);
 
-	bch2_fs_read_only_async(c);
 	bch2_journal_halt(&c->journal);
+	bch2_fs_read_only_async(c);
 
 	wake_up(&bch_read_only_wait);
 	return ret;
 }
 
-const char *bch2_fs_read_write(struct bch_fs *c)
+static int bch2_fs_read_write_late(struct bch_fs *c)
+{
+	int ret;
+
+	ret = bch2_gc_thread_start(c);
+	if (ret) {
+		bch_err(c, "error starting gc thread");
+		return ret;
+	}
+
+	ret = bch2_copygc_start(c);
+	if (ret) {
+		bch_err(c, "error starting copygc thread");
+		return ret;
+	}
+
+	ret = bch2_rebalance_start(c);
+	if (ret) {
+		bch_err(c, "error starting rebalance thread");
+		return ret;
+	}
+
+	schedule_work(&c->ec_stripe_delete_work);
+
+	return 0;
+}
+
+static int __bch2_fs_read_write(struct bch_fs *c, bool early)
 {
 	struct bch_dev *ca;
-	const char *err = NULL;
 	unsigned i;
+	int ret;
 
-	if (c->state != BCH_FS_STARTING &&
-	    c->state != BCH_FS_RO)
-		return NULL;
+	if (test_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags)) {
+		bch_err(c, "cannot go rw, unfixed btree errors");
+		return -EROFS;
+	}
+
+	if (test_bit(BCH_FS_RW, &c->flags))
+		return 0;
+
+	/*
+	 * nochanges is used for fsck -n mode - we have to allow going rw
+	 * during recovery for that to work:
+	 */
+	if (c->opts.norecovery ||
+	    (c->opts.nochanges &&
+	     (!early || c->opts.read_only)))
+		return -EROFS;
+
+	bch_info(c, "going read-write");
+
+	ret = bch2_fs_mark_dirty(c);
+	if (ret)
+		goto err;
+
+	/*
+	 * We need to write out a journal entry before we start doing btree
+	 * updates, to ensure that on unclean shutdown new journal blacklist
+	 * entries are created:
+	 */
+	bch2_journal_meta(&c->journal);
+
+	clear_bit(BCH_FS_ALLOC_CLEAN, &c->flags);
 
 	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)) {
+	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 btree GC thread";
-	if (bch2_gc_thread_start(c))
-		goto err;
+	set_bit(BCH_FS_ALLOCATOR_RUNNING, &c->flags);
 
-	err = "error starting copygc thread";
 	for_each_rw_member(ca, c, i)
-		if (bch2_copygc_start(c, ca)) {
-			percpu_ref_put(&ca->io_ref);
+		bch2_wake_allocator(ca);
+
+	if (!early) {
+		ret = bch2_fs_read_write_late(c);
+		if (ret)
 			goto err;
-		}
+	}
 
-	err = "error starting tiering thread";
-	if (bch2_tiering_start(c))
-		goto err;
+	percpu_ref_reinit(&c->writes);
+	set_bit(BCH_FS_RW, &c->flags);
+	set_bit(BCH_FS_WAS_RW, &c->flags);
+	return 0;
+err:
+	__bch2_fs_read_only(c);
+	return ret;
+}
 
-	schedule_delayed_work(&c->pd_controllers_update, 5 * HZ);
+int bch2_fs_read_write(struct bch_fs *c)
+{
+	return __bch2_fs_read_write(c, false);
+}
 
-	if (c->state != BCH_FS_STARTING)
-		percpu_ref_reinit(&c->writes);
+int bch2_fs_read_write_early(struct bch_fs *c)
+{
+	lockdep_assert_held(&c->state_lock);
 
-	c->state = BCH_FS_RW;
-	return NULL;
-err:
-	__bch2_fs_read_only(c);
-	return err;
+	return __bch2_fs_read_write(c, true);
 }
 
 /* Filesystem startup/shutdown: */
 
-static void bch2_fs_free(struct bch_fs *c)
+static void __bch2_fs_free(struct bch_fs *c)
 {
+	unsigned i;
+	int cpu;
+
+	for (i = 0; i < BCH_TIME_STAT_NR; i++)
+		bch2_time_stats_exit(&c->times[i]);
+
+	bch2_fs_snapshots_exit(c);
 	bch2_fs_quota_exit(c);
 	bch2_fs_fsio_exit(c);
+	bch2_fs_ec_exit(c);
 	bch2_fs_encryption_exit(c);
+	bch2_fs_io_exit(c);
+	bch2_fs_btree_interior_update_exit(c);
+	bch2_fs_btree_iter_exit(c);
+	bch2_fs_btree_key_cache_exit(&c->btree_key_cache);
 	bch2_fs_btree_cache_exit(c);
+	bch2_fs_replicas_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);
-	lg_lock_free(&c->usage_lock);
-	free_percpu(c->usage_percpu);
+	bch2_journal_keys_free(&c->journal_keys);
+	bch2_journal_entries_free(&c->journal_entries);
+	percpu_free_rwsem(&c->mark_lock);
+
+	if (c->btree_paths_bufs)
+		for_each_possible_cpu(cpu)
+			kfree(per_cpu_ptr(c->btree_paths_bufs, cpu)->path);
+
+	free_percpu(c->online_reserved);
+	free_percpu(c->btree_paths_bufs);
+	free_percpu(c->pcpu);
+	mempool_exit(&c->large_bkey_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_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));
+	kfree(c->journal_seq_blacklist_table);
+	kfree(c->unused_inode_hints);
+	free_heap(&c->copygc_heap);
 
+	if (c->io_complete_wq )
+		destroy_workqueue(c->io_complete_wq );
 	if (c->copygc_wq)
 		destroy_workqueue(c->copygc_wq);
-	if (c->wq)
-		destroy_workqueue(c->wq);
+	if (c->btree_io_complete_wq)
+		destroy_workqueue(c->btree_io_complete_wq);
+	if (c->btree_update_wq)
+		destroy_workqueue(c->btree_update_wq);
 
-	free_pages((unsigned long) c->disk_sb, c->disk_sb_order);
+	bch2_free_super(&c->disk_sb);
 	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;
-
-	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));
+	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");
+
+	set_bit(BCH_FS_STOPPING, &c->flags);
+
+	cancel_work_sync(&c->journal_seq_blacklist_gc_work);
+
+	down_write(&c->state_lock);
+	bch2_fs_read_only(c);
+	up_write(&c->state_lock);
 
 	for_each_member_device(ca, c, i)
 		if (ca->kobj.state_in_sysfs &&
 		    ca->disk_sb.bdev)
-			sysfs_remove_link(&part_to_dev(ca->disk_sb.bdev->bd_part)->kobj,
-					  "bcachefs");
+			sysfs_remove_link(bdev_kobj(ca->disk_sb.bdev), "bcachefs");
 
 	if (c->kobj.state_in_sysfs)
 		kobject_del(&c->kobj);
@@ -440,30 +556,87 @@ static void bch2_fs_offline(struct bch_fs *c)
 	kobject_put(&c->opts_dir);
 	kobject_put(&c->internal);
 
-	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->read_only_work);
+
+	for (i = 0; i < c->sb.nr_devices; i++)
+		if (c->devs[i])
+			bch2_free_super(&c->devs[i]->disk_sb);
 }
 
-static void bch2_fs_release(struct kobject *kobj)
+void bch2_fs_free(struct bch_fs *c)
 {
-	struct bch_fs *c = container_of(kobj, struct bch_fs, kobj);
+	unsigned i;
 
-	bch2_fs_free(c);
+	mutex_lock(&bch_fs_list_lock);
+	list_del(&c->list);
+	mutex_unlock(&bch_fs_list_lock);
+
+	closure_sync(&c->cl);
+	closure_debug_destroy(&c->cl);
+
+	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)
 {
-	mutex_lock(&c->state_lock);
-	BUG_ON(c->state == BCH_FS_STOPPING);
-	c->state = BCH_FS_STOPPING;
-	mutex_unlock(&c->state_lock);
+	__bch2_fs_stop(c);
+	bch2_fs_free(c);
+}
+
+static const char *bch2_fs_online(struct bch_fs *c)
+{
+	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";
+
+	down_write(&c->state_lock);
+
+	err = "error creating sysfs objects";
+	for_each_member_device(ca, c, i)
+		if (bch2_dev_sysfs_online(c, ca)) {
+			percpu_ref_put(&ca->ref);
+			goto err;
+		}
 
-	bch2_fs_exit(c);
+	list_add(&c->list, &bch_fs_list);
+	err = NULL;
+err:
+	up_write(&c->state_lock);
+	return err;
 }
 
 static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
@@ -471,16 +644,28 @@ 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 = kvpmalloc(sizeof(struct bch_fs), GFP_KERNEL|__GFP_ZERO);
 	if (!c)
-		return NULL;
+		goto out;
 
 	__module_get(THIS_MODULE);
 
+	closure_init(&c->cl, NULL);
+
+	c->kobj.kset = bcachefs_kset;
+	kobject_init(&c->kobj, &bch2_fs_ktype);
+	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);
+
 	c->minor		= -1;
+	c->disk_sb.fs_sb	= true;
 
-	mutex_init(&c->state_lock);
+	init_rwsem(&c->state_lock);
 	mutex_init(&c->sb_lock);
 	mutex_init(&c->replicas_gc_lock);
 	mutex_init(&c->btree_root_lock);
@@ -488,47 +673,67 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 
 	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_copygc_init(c);
+	bch2_fs_btree_key_cache_init_early(&c->btree_key_cache);
+	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_interior_update_list);
-	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);
-	mutex_init(&c->zlib_workspace_lock);
+	mutex_init(&c->snapshot_table_lock);
 
-	bio_list_init(&c->btree_write_error_list);
 	spin_lock_init(&c->btree_write_error_lock);
-	INIT_WORK(&c->btree_write_error_work, bch2_btree_write_error_work);
+
+	INIT_WORK(&c->journal_seq_blacklist_gc_work,
+		  bch2_blacklist_entries_gc);
+
+	INIT_LIST_HEAD(&c->journal_entries);
+	INIT_LIST_HEAD(&c->journal_iters);
 
 	INIT_LIST_HEAD(&c->fsck_errors);
 	mutex_init(&c->fsck_error_lock);
 
+	INIT_LIST_HEAD(&c->ec_stripe_head_list);
+	mutex_init(&c->ec_stripe_head_lock);
+
+	INIT_LIST_HEAD(&c->ec_stripe_new_list);
+	mutex_init(&c->ec_stripe_new_lock);
+
+	INIT_LIST_HEAD(&c->data_progress_list);
+	mutex_init(&c->data_progress_lock);
+
+	spin_lock_init(&c->ec_stripes_heap_lock);
+
 	seqcount_init(&c->gc_pos_lock);
 
-	init_waitqueue_head(&c->writeback_wait);
-	c->writeback_pages_max = (256 << 10) / PAGE_SIZE;
+	seqcount_init(&c->usage_lock);
 
-	c->copy_gc_enabled = 1;
-	c->tiering_enabled = 1;
-	c->tiering_percent = 10;
+	sema_init(&c->io_in_flight, 64);
 
-	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;
+	c->copy_gc_enabled		= 1;
+	c->rebalance.enabled		= 1;
+	c->promote_whole_extents	= true;
+
+	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_blocked_journal];
+	c->journal.flush_seq_time = &c->times[BCH_TIME_journal_flush_seq];
 
 	bch2_fs_btree_cache_init_early(&c->btree_cache);
 
+	mutex_init(&c->sectors_available_lock);
+
+	if (percpu_init_rwsem(&c->mark_lock))
+		goto err;
+
 	mutex_lock(&c->sb_lock);
 
 	if (bch2_sb_to_fs(c, sb)) {
@@ -545,332 +750,196 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 	bch2_opts_apply(&c->opts, opts);
 
 	c->block_bits		= ilog2(c->opts.block_size);
-
-	c->opts.nochanges	|= c->opts.noreplay;
-	c->opts.read_only	|= c->opts.nochanges;
+	c->btree_foreground_merge_threshold = BTREE_FOREGROUND_MERGE_THRESHOLD(c);
 
 	if (bch2_fs_init_fault("fs_alloc"))
 		goto err;
 
-	iter_size = (btree_blocks(c) + 1) * 2 *
-		sizeof(struct btree_node_iter_set);
-
-	if (!(c->wq = alloc_workqueue("bcachefs",
-				WQ_FREEZABLE|WQ_MEM_RECLAIM|WQ_HIGHPRI, 1)) ||
-	    !(c->copygc_wq = alloc_workqueue("bcache_copygc",
-				WQ_FREEZABLE|WQ_MEM_RECLAIM|WQ_HIGHPRI, 1)) ||
-	    percpu_ref_init(&c->writes, bch2_writes_disabled, 0, 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,
-				      sizeof(struct btree_update)) ||
+	iter_size = sizeof(struct sort_iter) +
+		(btree_blocks(c) + 1) * 2 *
+		sizeof(struct sort_iter_set);
+
+	c->inode_shard_bits = ilog2(roundup_pow_of_two(num_possible_cpus()));
+
+	if (!(c->btree_update_wq = alloc_workqueue("bcachefs",
+				WQ_FREEZABLE|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE, 1)) ||
+	    !(c->btree_io_complete_wq = alloc_workqueue("bcachefs_btree_io",
+				WQ_FREEZABLE|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE, 1)) ||
+	    !(c->copygc_wq = alloc_workqueue("bcachefs_copygc",
+				WQ_FREEZABLE|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE, 1)) ||
+	    !(c->io_complete_wq = alloc_workqueue("bcachefs_io",
+				WQ_FREEZABLE|WQ_HIGHPRI|WQ_MEM_RECLAIM, 1)) ||
+	    percpu_ref_init(&c->writes, bch2_writes_disabled,
+			    PERCPU_REF_INIT_DEAD, GFP_KERNEL) ||
 	    mempool_init_kmalloc_pool(&c->fill_iter, 1, iter_size) ||
 	    bioset_init(&c->btree_bio, 1,
 			max(offsetof(struct btree_read_bio, bio),
 			    offsetof(struct btree_write_bio, wbio.bio)),
 			BIOSET_NEED_BVECS) ||
-	    bioset_init(&c->bio_read, 1, offsetof(struct bch_read_bio, bio),
-			BIOSET_NEED_BVECS) ||
-	    bioset_init(&c->bio_read_split, 1, offsetof(struct bch_read_bio, bio),
-			BIOSET_NEED_BVECS) ||
-	    bioset_init(&c->bio_write, 1, offsetof(struct bch_write_bio, bio),
-			BIOSET_NEED_BVECS) ||
-	    mempool_init_page_pool(&c->bio_bounce_pages,
-				   max_t(unsigned,
-					 c->opts.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)) ||
+	    !(c->pcpu = alloc_percpu(struct bch_fs_pcpu)) ||
+	    !(c->btree_paths_bufs = alloc_percpu(struct btree_path_buf)) ||
+	    !(c->online_reserved = alloc_percpu(u64)) ||
+	    mempool_init_kvpmalloc_pool(&c->btree_bounce_pool, 1,
+					btree_bytes(c)) ||
+	    mempool_init_kmalloc_pool(&c->large_bkey_pool, 1, 2048) ||
+	    !(c->unused_inode_hints = kcalloc(1U << c->inode_shard_bits,
+					      sizeof(u64), GFP_KERNEL)) ||
 	    bch2_io_clock_init(&c->io_clock[READ]) ||
 	    bch2_io_clock_init(&c->io_clock[WRITE]) ||
 	    bch2_fs_journal_init(&c->journal) ||
+	    bch2_fs_replicas_init(c) ||
 	    bch2_fs_btree_cache_init(c) ||
+	    bch2_fs_btree_key_cache_init(&c->btree_key_cache) ||
+	    bch2_fs_btree_iter_init(c) ||
+	    bch2_fs_btree_interior_update_init(c) ||
+	    bch2_fs_subvolumes_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;
 
-	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;
 
-	/*
-	 * Now that all allocations have succeeded, init various refcounty
-	 * things that let us shutdown:
-	 */
-	closure_init(&c->cl, NULL);
+	bch2_journal_entry_res_resize(&c->journal,
+			&c->btree_root_journal_res,
+			BTREE_ID_NR * (JSET_KEYS_U64s + BKEY_BTREE_PTR_U64s_MAX));
+	bch2_dev_usage_journal_reserve(c);
+	bch2_journal_entry_res_resize(&c->journal,
+			&c->clock_journal_res,
+			(sizeof(struct jset_entry_clock) / sizeof(u64)) * 2);
 
-	c->kobj.kset = bcachefs_kset;
-	kobject_init(&c->kobj, &bch2_fs_ktype);
-	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);
+	mutex_lock(&bch_fs_list_lock);
+	err = bch2_fs_online(c);
+	mutex_unlock(&bch_fs_list_lock);
+	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);
-	return NULL;
+	c = NULL;
+	goto out;
 }
 
-static const char *__bch2_fs_online(struct bch_fs *c)
+noinline_for_stack
+static void print_mount_opts(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";
+	enum bch_opt_id i;
+	char buf[512];
+	struct printbuf p = PBUF(buf);
+	bool first = true;
 
-	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";
+	strcpy(buf, "(null)");
 
-	mutex_lock(&c->state_lock);
+	if (c->opts.read_only) {
+		pr_buf(&p, "ro");
+		first = false;
+	}
 
-	err = "error creating sysfs objects";
-	__for_each_member_device(ca, c, i, NULL)
-		if (bch2_dev_sysfs_online(c, ca))
-			goto err;
+	for (i = 0; i < bch2_opts_nr; i++) {
+		const struct bch_option *opt = &bch2_opt_table[i];
+		u64 v = bch2_opt_get_by_id(&c->opts, i);
 
-	list_add(&c->list, &bch_fs_list);
-	err = NULL;
-err:
-	mutex_unlock(&c->state_lock);
-	return err;
-}
+		if (!(opt->mode & OPT_MOUNT))
+			continue;
 
-static const char *bch2_fs_online(struct bch_fs *c)
-{
-	const char *err;
+		if (v == bch2_opt_get_by_id(&bch2_opts_default, i))
+			continue;
 
-	mutex_lock(&bch_fs_list_lock);
-	err = __bch2_fs_online(c);
-	mutex_unlock(&bch_fs_list_lock);
+		if (!first)
+			pr_buf(&p, ",");
+		first = false;
+		bch2_opt_to_text(&p, c, opt, v, OPT_SHOW_MOUNT_STYLE);
+	}
 
-	return err;
+	bch_info(c, "mounted with opts: %s", buf);
 }
 
-static const char *__bch2_fs_start(struct bch_fs *c)
+int 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;
-	time64_t now;
+	time64_t now = ktime_get_real_seconds();
 	unsigned i;
 	int ret = -EINVAL;
 
-	mutex_lock(&c->state_lock);
+	down_write(&c->state_lock);
 
-	BUG_ON(c->state != BCH_FS_STARTING);
+	BUG_ON(test_bit(BCH_FS_STARTED, &c->flags));
 
 	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;
-
-			k = bch2_journal_find_btree_root(c, j, i, &level);
-			if (!k)
-				continue;
-
-			err = "invalid btree root pointer";
-			if (IS_ERR(k))
-				goto err;
-
-			err = "error reading btree root";
-			if (bch2_btree_root_read(c, i, k, level)) {
-				if (i != BTREE_ID_ALLOC)
-					goto err;
-
-				mustfix_fsck_err(c, "error reading btree root");
-			}
-		}
-
-		for (i = 0; i < BTREE_ID_NR; i++)
-			if (!c->btree_roots[i].b)
-				bch2_btree_root_alloc(c, i);
-
-		err = "error reading allocation information";
-		ret = bch2_alloc_read(c, &journal);
-		if (ret)
-			goto err;
-
-		set_bit(BCH_FS_ALLOC_READ_DONE, &c->flags);
-
-		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;
-
-		/*
-		 * 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";
-		if (bch2_fs_allocator_start(c))
-			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");
-
-		if (c->opts.usrquota || c->opts.grpquota) {
-			bch_verbose(c, "reading quotas:");
-			ret = bch2_fs_quota_read(c);
-			if (ret)
-				goto err;
-			bch_verbose(c, "quotas done");
-		}
-	} else {
-		struct bch_inode_unpacked inode;
-		struct bkey_inode_buf packed_inode;
-
-		bch_notice(c, "initializing new filesystem");
-
-		set_bit(BCH_FS_ALLOC_READ_DONE, &c->flags);
-		set_bit(BCH_FS_BRAND_NEW_FS, &c->flags);
-
-		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(c, ca)) {
-				percpu_ref_put(&ca->io_ref);
-				goto err;
-			}
-
-		clear_bit(BCH_FS_BRAND_NEW_FS, &c->flags);
-
-		for (i = 0; i < BTREE_ID_NR; i++)
-			bch2_btree_root_alloc(c, i);
-
-		/*
-		 * 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";
-		if (bch2_fs_allocator_start(c))
-			goto err;
-
-		bch2_inode_init(c, &inode, 0, 0,
-			       S_IFDIR|S_IRWXU|S_IRUGO|S_IXUGO, 0, NULL);
-		inode.bi_inum = BCACHEFS_ROOT_INO;
-
-		bch2_inode_pack(&packed_inode, &inode);
-
-		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 = BCH_SB_INITIALIZED(c->disk_sb.sb)
+		? bch2_fs_recovery(c)
+		: bch2_fs_initialize(c);
+	if (ret)
+		goto err;
 
-		if (c->opts.usrquota || c->opts.grpquota) {
-			ret = bch2_fs_quota_read(c);
-			if (ret)
-				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";
+	ret = -EINVAL;
 	if (bch2_fs_init_fault("fs_start"))
 		goto err;
 
-	if (c->opts.read_only) {
+	set_bit(BCH_FS_STARTED, &c->flags);
+
+	/*
+	 * Allocator threads don't start filling copygc reserve until after we
+	 * set BCH_FS_STARTED - wake them now:
+	 *
+	 * XXX ugly hack:
+	 * Need to set ca->allocator_state here instead of relying on the
+	 * allocator threads to do it to avoid racing with the copygc threads
+	 * checking it and thinking they have no alloc reserve:
+	 */
+	for_each_online_member(ca, c, i) {
+		ca->allocator_state = ALLOCATOR_running;
+		bch2_wake_allocator(ca);
+	}
+
+	if (c->opts.read_only || c->opts.nochanges) {
 		bch2_fs_read_only(c);
 	} else {
-		err = bch2_fs_read_write(c);
-		if (err)
+		err = "error going read write";
+		ret = !test_bit(BCH_FS_RW, &c->flags)
+			? bch2_fs_read_write(c)
+			: bch2_fs_read_write_late(c);
+		if (ret)
 			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);
-
-	err = NULL;
+	print_mount_opts(c);
+	ret = 0;
 out:
-	mutex_unlock(&c->state_lock);
-	bch2_journal_entries_free(&journal);
-	return err;
+	up_write(&c->state_lock);
+	return ret;
 err:
-fsck_err:
 	switch (ret) {
 	case BCH_FSCK_ERRORS_NOT_FIXED:
 		bch_err(c, "filesystem contains errors: please report this to the developers");
@@ -897,16 +966,11 @@ fsck_err:
 		break;
 	}
 
-	BUG_ON(!err);
-	set_bit(BCH_FS_ERROR, &c->flags);
+	if (ret >= 0)
+		ret = -EIO;
 	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;
@@ -919,7 +983,7 @@ static const char *bch2_dev_may_add(struct bch_sb *sb, struct bch_fs *c)
 		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;
@@ -954,12 +1018,13 @@ static void bch2_dev_release(struct kobject *kobj)
 
 static void bch2_dev_free(struct bch_dev *ca)
 {
+	bch2_dev_allocator_stop(ca);
+
 	cancel_work_sync(&ca->io_error_work);
 
 	if (ca->kobj.state_in_sysfs &&
 	    ca->disk_sb.bdev)
-		sysfs_remove_link(&part_to_dev(ca->disk_sb.bdev->bd_part)->kobj,
-				  "bcachefs");
+		sysfs_remove_link(bdev_kobj(ca->disk_sb.bdev), "bcachefs");
 
 	if (ca->kobj.state_in_sysfs)
 		kobject_del(&ca->kobj);
@@ -970,6 +1035,10 @@ static void bch2_dev_free(struct bch_dev *ca)
 	free_percpu(ca->io_done);
 	bioset_exit(&ca->replica_set);
 	bch2_dev_buckets_free(ca);
+	free_page((unsigned long) ca->sb_read_scratch);
+
+	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);
@@ -991,10 +1060,7 @@ static void __bch2_dev_offline(struct bch_fs *c, struct bch_dev *ca)
 	wait_for_completion(&ca->io_ref_completion);
 
 	if (ca->kobj.state_in_sysfs) {
-		struct kobject *block =
-			&part_to_dev(ca->disk_sb.bdev->bd_part)->kobj;
-
-		sysfs_remove_link(block, "bcachefs");
+		sysfs_remove_link(bdev_kobj(ca->disk_sb.bdev), "bcachefs");
 		sysfs_remove_link(&ca->kobj, "block");
 	}
 
@@ -1031,12 +1097,12 @@ static int bch2_dev_sysfs_online(struct bch_fs *c, struct bch_dev *ca)
 	}
 
 	if (ca->disk_sb.bdev) {
-		struct kobject *block =
-			&part_to_dev(ca->disk_sb.bdev->bd_part)->kobj;
+		struct kobject *block = bdev_kobj(ca->disk_sb.bdev);
 
 		ret = sysfs_create_link(block, &ca->kobj, "bcachefs");
 		if (ret)
 			return ret;
+
 		ret = sysfs_create_link(&ca->kobj, block, "block");
 		if (ret)
 			return ret;
@@ -1045,80 +1111,101 @@ static int bch2_dev_sysfs_online(struct bch_fs *c, 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;
 	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->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);
 
-	writepoint_init(&ca->copygc_write_point, BCH_DATA_USER);
-
-	spin_lock_init(&ca->freelist_lock);
-	bch2_dev_copygc_init(ca);
-
 	INIT_WORK(&ca->io_error_work, bch2_io_error_work);
 
-	if (bch2_fs_init_fault("dev_alloc"))
-		goto err;
-
-	member = bch2_sb_get_members(c->disk_sb)->members + dev_idx;
+	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;
-	scnprintf(ca->name, sizeof(ca->name), "dev-%u", dev_idx);
+
+	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) ||
+	    !(ca->sb_read_scratch = (void *) __get_free_page(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");
-
-	return 0;
-err:
-	bch2_dev_free(ca);
-	return -ENOMEM;
 }
 
-static int __bch2_dev_online(struct bch_fs *c, struct bch_sb_handle *sb)
+static int bch2_dev_alloc(struct bch_fs *c, unsigned dev_idx)
 {
-	struct bch_dev *ca;
-	int ret;
+	struct bch_member *member =
+		bch2_sb_get_members(c->disk_sb.sb)->members + dev_idx;
+	struct bch_dev *ca = NULL;
+	int ret = 0;
 
-	lockdep_assert_held(&c->state_lock);
+	pr_verbose_init(c->opts, "");
 
-	if (le64_to_cpu(sb->sb->seq) >
-	    le64_to_cpu(c->disk_sb->seq))
-		bch2_sb_to_fs(c, sb->sb);
+	if (bch2_fs_init_fault("dev_alloc"))
+		goto err;
 
-	BUG_ON(sb->sb->dev_idx >= c->sb.nr_devices ||
-	       !c->devs[sb->sb->dev_idx]);
+	ca = __bch2_dev_alloc(c, member);
+	if (!ca)
+		goto err;
 
-	ca = bch_dev_locked(c, sb->sb->dev_idx);
+	ca->fs = c;
+
+	if (ca->mi.state == BCH_MEMBER_STATE_rw &&
+	    bch2_dev_allocator_start(ca)) {
+		bch2_dev_free(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;
+}
+
+static int __bch2_dev_attach_bdev(struct bch_dev *ca, struct bch_sb_handle *sb)
+{
+	unsigned ret;
 
 	if (bch2_dev_is_online(ca)) {
 		bch_err(ca, "already have device online in slot %u",
@@ -1134,37 +1221,54 @@ static int __bch2_dev_online(struct bch_fs *c, struct bch_sb_handle *sb)
 
 	BUG_ON(!percpu_ref_is_zero(&ca->io_ref));
 
+	if (get_capacity(sb->bdev->bd_disk) <
+	    ca->mi.bucket_size * ca->mi.nbuckets) {
+		bch_err(ca, "device too small");
+		return -EINVAL;
+	}
+
 	ret = bch2_dev_journal_init(ca, sb->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);
-
 	/* Commit: */
 	ca->disk_sb = *sb;
 	if (sb->mode & FMODE_EXCL)
 		ca->disk_sb.bdev->bd_holder = ca;
 	memset(sb, 0, sizeof(*sb));
 
+	percpu_ref_reinit(&ca->io_ref);
+
+	return 0;
+}
+
+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->state_lock);
+
+	if (le64_to_cpu(sb->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 = bch_dev_locked(c, sb->sb->dev_idx);
+
+	ret = __bch2_dev_attach_bdev(ca, sb);
+	if (ret)
+		return ret;
+
+	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);
 
-	mutex_lock(&c->sb_lock);
-	bch2_mark_dev_superblock(c, ca, BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE);
-	mutex_unlock(&c->sb_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;
 }
 
@@ -1183,22 +1287,22 @@ bool bch2_dev_state_allowed(struct bch_fs *c, struct bch_dev *ca,
 			    enum bch_member_state new_state, int flags)
 {
 	struct bch_devs_mask new_online_devs;
-	struct replicas_status s;
 	struct bch_dev *ca2;
 	int i, nr_rw = 0, required;
 
 	lockdep_assert_held(&c->state_lock);
 
 	switch (new_state) {
-	case BCH_MEMBER_STATE_RW:
+	case BCH_MEMBER_STATE_rw:
 		return true;
-	case BCH_MEMBER_STATE_RO:
-		if (ca->mi.state != BCH_MEMBER_STATE_RW)
+	case BCH_MEMBER_STATE_ro:
+		if (ca->mi.state != BCH_MEMBER_STATE_rw)
 			return true;
 
 		/* 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
@@ -1207,20 +1311,18 @@ 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;
-	case BCH_MEMBER_STATE_FAILED:
-	case BCH_MEMBER_STATE_SPARE:
-		if (ca->mi.state != BCH_MEMBER_STATE_RW &&
-		    ca->mi.state != BCH_MEMBER_STATE_RO)
+		return nr_rw >= required;
+	case BCH_MEMBER_STATE_failed:
+	case BCH_MEMBER_STATE_spare:
+		if (ca->mi.state != BCH_MEMBER_STATE_rw &&
+		    ca->mi.state != BCH_MEMBER_STATE_ro)
 			return true;
 
 		/* do we have enough devices to read from?  */
 		new_online_devs = bch2_online_devs(c);
 		__clear_bit(ca->dev_idx, new_online_devs.d);
 
-		s = __bch2_replicas_status(c, new_online_devs);
-
-		return bch2_have_enough_devs(c, s, flags);
+		return bch2_have_enough_devs(c, new_online_devs, flags, false);
 	default:
 		BUG();
 	}
@@ -1228,26 +1330,30 @@ bool bch2_dev_state_allowed(struct bch_fs *c, struct bch_dev *ca,
 
 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;
+	unsigned i, flags = 0;
+
+	if (c->opts.very_degraded)
+		flags |= BCH_FORCE_IF_DEGRADED|BCH_FORCE_IF_LOST;
 
-	if (!c->opts.degraded) {
+	if (c->opts.degraded)
+		flags |= BCH_FORCE_IF_DEGRADED;
+
+	if (!c->opts.degraded &&
+	    !c->opts.very_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))
+		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)) {
+			    (ca->mi.state == BCH_MEMBER_STATE_rw ||
+			     ca->mi.state == BCH_MEMBER_STATE_ro)) {
 				mutex_unlock(&c->sb_lock);
 				return false;
 			}
@@ -1255,14 +1361,16 @@ static bool bch2_fs_may_start(struct bch_fs *c)
 		mutex_unlock(&c->sb_lock);
 	}
 
-	s = bch2_replicas_status(c);
-
-	return bch2_have_enough_devs(c, s, flags);
+	return bch2_have_enough_devs(c, bch2_online_devs(c), flags, true);
 }
 
 static void __bch2_dev_read_only(struct bch_fs *c, struct bch_dev *ca)
 {
-	bch2_copygc_stop(ca);
+	/*
+	 * Device going read only means the copygc reserve get smaller, so we
+	 * don't want that happening while copygc is in progress:
+	 */
+	bch2_copygc_stop(c);
 
 	/*
 	 * The allocator thread itself allocates btree nodes, so stop it first:
@@ -1270,13 +1378,15 @@ static void __bch2_dev_read_only(struct bch_fs *c, struct bch_dev *ca)
 	bch2_dev_allocator_stop(ca);
 	bch2_dev_allocator_remove(c, ca);
 	bch2_dev_journal_stop(&c->journal, ca);
+
+	bch2_copygc_start(c);
 }
 
 static const char *__bch2_dev_read_write(struct bch_fs *c, struct bch_dev *ca)
 {
 	lockdep_assert_held(&c->state_lock);
 
-	BUG_ON(ca->mi.state != BCH_MEMBER_STATE_RW);
+	BUG_ON(ca->mi.state != BCH_MEMBER_STATE_rw);
 
 	bch2_dev_allocator_add(c, ca);
 	bch2_recalc_capacity(c);
@@ -1284,12 +1394,6 @@ 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_copygc_start(c, ca))
-		return "error starting copygc thread";
-
-	if (bch2_tiering_start(c))
-		return "error starting tiering thread";
-
 	return NULL;
 }
 
@@ -1297,6 +1401,7 @@ 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;
@@ -1304,22 +1409,24 @@ 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]);
+	bch_notice(ca, "%s", bch2_member_states[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,
@@ -1327,66 +1434,74 @@ int bch2_dev_set_state(struct bch_fs *c, struct bch_dev *ca,
 {
 	int ret;
 
-	mutex_lock(&c->state_lock);
+	down_write(&c->state_lock);
 	ret = __bch2_dev_set_state(c, ca, new_state, flags);
-	mutex_unlock(&c->state_lock);
+	up_write(&c->state_lock);
 
 	return ret;
 }
 
 /* Device add/removal: */
 
+static int bch2_dev_remove_alloc(struct bch_fs *c, struct bch_dev *ca)
+{
+	struct btree_trans trans;
+	size_t i;
+	int ret;
+
+	bch2_trans_init(&trans, c, 0, 0);
+
+	for (i = 0; i < ca->mi.nbuckets; i++) {
+		ret = bch2_btree_key_cache_flush(&trans,
+				BTREE_ID_alloc, POS(ca->dev_idx, i));
+		if (ret)
+			break;
+	}
+	bch2_trans_exit(&trans);
+
+	if (ret)
+		return ret;
+
+	return bch2_btree_delete_range(c, BTREE_ID_alloc,
+				       POS(ca->dev_idx, 0),
+				       POS(ca->dev_idx + 1, 0),
+				       NULL);
+}
+
 int bch2_dev_remove(struct bch_fs *c, struct bch_dev *ca, int flags)
 {
 	struct bch_sb_field_members *mi;
 	unsigned dev_idx = ca->dev_idx, data;
 	int ret = -EINVAL;
 
-	mutex_lock(&c->state_lock);
+	down_write(&c->state_lock);
 
-	percpu_ref_put(&ca->ref); /* XXX */
+	/*
+	 * We consume a reference to ca->ref, regardless of whether we succeed
+	 * or fail:
+	 */
+	percpu_ref_put(&ca->ref);
 
-	if (!bch2_dev_state_allowed(c, ca, BCH_MEMBER_STATE_FAILED, flags)) {
+	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_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(&c->journal, ca->dev_idx);
+	ret = bch2_journal_flush_device_pins(&c->journal, ca->dev_idx);
 	if (ret) {
 		bch_err(ca, "Remove failed: error %i flushing journal", ret);
 		goto err;
 	}
 
-	data = bch2_dev_has_data(c, ca);
-	if (data) {
-		char data_has_str[100];
-		bch2_scnprint_flag_list(data_has_str,
-					sizeof(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),
-				      ZERO_VERSION,
-				      NULL, NULL, NULL);
+	ret = bch2_dev_remove_alloc(c, ca);
 	if (ret) {
 		bch_err(ca, "Remove failed, error deleting alloc info");
 		goto err;
@@ -1396,12 +1511,34 @@ int bch2_dev_remove(struct bch_fs *c, struct bch_dev *ca, int flags)
 	 * must flush all existing journal entries, they might have
 	 * (overwritten) keys that point to the device we're removing:
 	 */
-	ret = bch2_journal_flush_all_pins(&c->journal);
+	bch2_journal_flush_all_pins(&c->journal);
+	/*
+	 * hack to ensure bch2_replicas_gc2() clears out entries to this device
+	 */
+	bch2_journal_meta(&c->journal);
+	ret = bch2_journal_error(&c->journal);
 	if (ret) {
 		bch_err(ca, "Remove failed, journal error");
 		goto err;
 	}
 
+	ret = bch2_replicas_gc2(c);
+	if (ret) {
+		bch_err(ca, "Remove failed: error %i from replicas gc", ret);
+		goto err;
+	}
+
+	data = bch2_dev_has_data(c, ca);
+	if (data) {
+		char data_has_str[100];
+
+		bch2_flags_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;
+	}
+
 	__bch2_dev_offline(c, ca);
 
 	mutex_lock(&c->sb_lock);
@@ -1418,18 +1555,21 @@ 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);
 
 	mutex_unlock(&c->sb_lock);
-	mutex_unlock(&c->state_lock);
+	up_write(&c->state_lock);
+
+	bch2_dev_usage_journal_reserve(c);
 	return 0;
 err:
-	if (ca->mi.state == BCH_MEMBER_STATE_RW)
+	if (ca->mi.state == BCH_MEMBER_STATE_rw &&
+	    !percpu_ref_is_zero(&ca->io_ref))
 		__bch2_dev_read_write(c, ca);
-	mutex_unlock(&c->state_lock);
+	up_write(&c->state_lock);
 	return ret;
 }
 
@@ -1440,8 +1580,8 @@ int bch2_dev_add(struct bch_fs *c, const char *path)
 	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;
 
@@ -1453,27 +1593,66 @@ int bch2_dev_add(struct bch_fs *c, const char *path)
 	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(NULL, ca, 0);
+
+	err = "journal alloc failed";
+	ret = bch2_dev_journal_alloc(ca);
+	if (ret)
+		goto err;
+
+	down_write(&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";
@@ -1484,64 +1663,55 @@ 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";
 
-	dev_mi = bch2_sb_resize_members(&sb, u64s);
-	if (!dev_mi)
-		goto err_unlock;
+	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;
 
-	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);
 
 	bch2_write_super(c);
 	mutex_unlock(&c->sb_lock);
 
-	if (bch2_dev_alloc(c, dev_idx)) {
-		err = "cannot allocate memory";
-		ret = -ENOMEM;
-		goto err;
-	}
+	bch2_dev_usage_journal_reserve(c);
 
-	if (__bch2_dev_online(c, &sb)) {
-		err = "bch2_dev_online() error";
-		ret = -ENOMEM;
-		goto err;
-	}
+	err = "error marking superblock";
+	ret = bch2_trans_mark_dev_sb(c, ca);
+	if (ret)
+		goto err_late;
 
-	ca = bch_dev_locked(c, dev_idx);
-	if (ca->mi.state == BCH_MEMBER_STATE_RW) {
+	if (ca->mi.state == BCH_MEMBER_STATE_rw) {
 		err = __bch2_dev_read_write(c, ca);
 		if (err)
-			goto err;
-
-		err = "journal alloc failed";
-		if (bch2_dev_journal_alloc(c, ca))
-			goto err;
+			goto err_late;
 	}
 
-	mutex_unlock(&c->state_lock);
+	up_write(&c->state_lock);
 	return 0;
+
 err_unlock:
 	mutex_unlock(&c->sb_lock);
+	up_write(&c->state_lock);
 err:
-	mutex_unlock(&c->state_lock);
+	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:
+	up_write(&c->state_lock);
+	bch_err(c, "Error going rw after adding device: %s", err);
+	return -EINVAL;
 }
 
 /* Hot add existing device to running filesystem: */
@@ -1549,41 +1719,57 @@ int bch2_dev_online(struct bch_fs *c, const char *path)
 {
 	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);
+	down_write(&c->state_lock);
 
 	ret = bch2_read_super(path, &opts, &sb);
 	if (ret) {
-		mutex_unlock(&c->state_lock);
+		up_write(&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;
 
-	if (__bch2_dev_online(c, &sb)) {
-		err = "__bch2_dev_online() error";
+	if (bch2_dev_attach_bdev(c, &sb)) {
+		err = "bch2_dev_attach_bdev() error";
 		goto err;
 	}
 
 	ca = bch_dev_locked(c, dev_idx);
-	if (ca->mi.state == BCH_MEMBER_STATE_RW) {
+
+	if (bch2_trans_mark_dev_sb(c, ca)) {
+		err = "bch2_trans_mark_dev_sb() error";
+		goto err;
+	}
+
+	if (ca->mi.state == BCH_MEMBER_STATE_rw) {
 		err = __bch2_dev_read_write(c, ca);
 		if (err)
 			goto err;
 	}
 
-	mutex_unlock(&c->state_lock);
+	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);
+
+	up_write(&c->state_lock);
 	return 0;
 err:
-	mutex_unlock(&c->state_lock);
+	up_write(&c->state_lock);
 	bch2_free_super(&sb);
 	bch_err(c, "error bringing %s online: %s", path, err);
 	return -EINVAL;
@@ -1591,73 +1777,32 @@ err:
 
 int bch2_dev_offline(struct bch_fs *c, struct bch_dev *ca, int flags)
 {
-	mutex_lock(&c->state_lock);
+	down_write(&c->state_lock);
 
 	if (!bch2_dev_is_online(ca)) {
 		bch_err(ca, "Already offline");
-		mutex_unlock(&c->state_lock);
+		up_write(&c->state_lock);
 		return 0;
 	}
 
-	if (!bch2_dev_state_allowed(c, ca, BCH_MEMBER_STATE_FAILED, flags)) {
+	if (!bch2_dev_state_allowed(c, ca, BCH_MEMBER_STATE_failed, flags)) {
 		bch_err(ca, "Cannot offline required disk");
-		mutex_unlock(&c->state_lock);
+		up_write(&c->state_lock);
 		return -EINVAL;
 	}
 
 	__bch2_dev_offline(c, ca);
 
-	mutex_unlock(&c->state_lock);
+	up_write(&c->state_lock);
 	return 0;
 }
 
-int bch2_dev_evacuate(struct bch_fs *c, struct bch_dev *ca)
-{
-	unsigned data;
-	int ret = 0;
-
-	mutex_lock(&c->state_lock);
-
-	if (ca->mi.state == BCH_MEMBER_STATE_RW &&
-	    bch2_dev_is_online(ca)) {
-		bch_err(ca, "Cannot migrate data off RW device");
-		ret = -EINVAL;
-		goto err;
-	}
-
-	ret = bch2_dev_data_migrate(c, ca, 0);
-	if (ret) {
-		bch_err(ca, "Error migrating data: %i", ret);
-		goto err;
-	}
-
-	ret = bch2_journal_flush_device(&c->journal, ca->dev_idx);
-	if (ret) {
-		bch_err(ca, "Migrate failed: error %i flushing journal", ret);
-		goto err;
-	}
-
-	data = bch2_dev_has_data(c, ca);
-	if (data) {
-		char buf[100];
-
-		bch2_scnprint_flag_list(buf, sizeof(buf),
-					bch2_data_types, data);
-		bch_err(ca, "Migrate failed, still has data (%s)", buf);
-		ret = -EINVAL;
-		goto err;
-	}
-err:
-	mutex_unlock(&c->state_lock);
-	return ret;
-}
-
 int bch2_dev_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
 {
 	struct bch_member *mi;
 	int ret = 0;
 
-	mutex_lock(&c->state_lock);
+	down_write(&c->state_lock);
 
 	if (nbuckets < ca->mi.nbuckets) {
 		bch_err(ca, "Cannot shrink yet");
@@ -1679,8 +1824,13 @@ int bch2_dev_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
 		goto err;
 	}
 
+	ret = bch2_trans_mark_dev_sb(c, ca);
+	if (ret) {
+		goto err;
+	}
+
 	mutex_lock(&c->sb_lock);
-	mi = &bch2_sb_get_members(c->disk_sb)->members[ca->dev_idx];
+	mi = &bch2_sb_get_members(c->disk_sb.sb)->members[ca->dev_idx];
 	mi->nbuckets = cpu_to_le64(nbuckets);
 
 	bch2_write_super(c);
@@ -1688,10 +1838,33 @@ int bch2_dev_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
 
 	bch2_recalc_capacity(c);
 err:
-	mutex_unlock(&c->state_lock);
+	up_write(&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 bch_dev *ca;
+	dev_t dev;
+	unsigned i;
+	int ret;
+
+	ret = lookup_bdev(path, &dev);
+	if (ret)
+		return ERR_PTR(ret);
+
+	rcu_read_lock();
+	for_each_member_device_rcu(ca, c, i, NULL)
+		if (ca->disk_sb.bdev->bd_dev == dev)
+			goto found;
+	ca = ERR_PTR(-ENOENT);
+found:
+	rcu_read_unlock();
+
+	return ca;
+}
+
 /* Filesystem open: */
 
 struct bch_fs *bch2_fs_open(char * const *devices, unsigned nr_devices,
@@ -1699,15 +1872,22 @@ struct bch_fs *bch2_fs_open(char * const *devices, unsigned nr_devices,
 {
 	struct bch_sb_handle *sb = NULL;
 	struct bch_fs *c = NULL;
+	struct bch_sb_field_members *mi;
 	unsigned i, best_sb = 0;
 	const char *err;
 	int ret = -ENOMEM;
 
-	if (!nr_devices)
-		return ERR_PTR(-EINVAL);
+	pr_verbose_init(opts, "");
+
+	if (!nr_devices) {
+		c = ERR_PTR(-EINVAL);
+		goto out2;
+	}
 
-	if (!try_module_get(THIS_MODULE))
-		return ERR_PTR(-ENODEV);
+	if (!try_module_get(THIS_MODULE)) {
+		c = ERR_PTR(-ENODEV);
+		goto out2;
+	}
 
 	sb = kcalloc(nr_devices, sizeof(*sb), GFP_KERNEL);
 	if (!sb)
@@ -1728,10 +1908,24 @@ struct bch_fs *bch2_fs_open(char * const *devices, unsigned nr_devices,
 		    le64_to_cpu(sb[best_sb].sb->seq))
 			best_sb = i;
 
-	for (i = 0; i < nr_devices; i++) {
+	mi = bch2_sb_get_members(sb[best_sb].sb);
+
+	i = 0;
+	while (i < nr_devices) {
+		if (i != best_sb &&
+		    !bch2_dev_exists(sb[best_sb].sb, mi, sb[i].sb->dev_idx)) {
+			char buf[BDEVNAME_SIZE];
+			pr_info("%s has been removed, skipping",
+				bdevname(sb[i].bdev, buf));
+			bch2_free_super(&sb[i]);
+			array_remove_item(sb, nr_devices, i);
+			continue;
+		}
+
 		err = bch2_dev_in_fs(sb[best_sb].sb, sb[i].sb);
 		if (err)
 			goto err_print;
+		i++;
 	}
 
 	ret = -ENOMEM;
@@ -1740,30 +1934,28 @@ struct bch_fs *bch2_fs_open(char * const *devices, unsigned nr_devices,
 		goto err;
 
 	err = "bch2_dev_online() error";
-	mutex_lock(&c->state_lock);
+	down_write(&c->state_lock);
 	for (i = 0; i < nr_devices; i++)
-		if (__bch2_dev_online(c, &sb[i])) {
-			mutex_unlock(&c->state_lock);
+		if (bch2_dev_attach_bdev(c, &sb[i])) {
+			up_write(&c->state_lock);
 			goto err_print;
 		}
-	mutex_unlock(&c->state_lock);
+	up_write(&c->state_lock);
 
 	err = "insufficient devices";
 	if (!bch2_fs_may_start(c))
 		goto err_print;
 
 	if (!c->opts.nostart) {
-		err = __bch2_fs_start(c);
-		if (err)
-			goto err_print;
+		ret = bch2_fs_start(c);
+		if (ret)
+			goto err;
 	}
-
-	err = bch2_fs_online(c);
-	if (err)
-		goto err_print;
-
+out:
 	kfree(sb);
 	module_put(THIS_MODULE);
+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",
@@ -1772,12 +1964,10 @@ err_print:
 err:
 	if (c)
 		bch2_fs_stop(c);
-
 	for (i = 0; i < nr_devices; i++)
 		bch2_free_super(&sb[i]);
-	kfree(sb);
-	module_put(THIS_MODULE);
-	return ERR_PTR(ret);
+	c = ERR_PTR(ret);
+	goto out;
 }
 
 static const char *__bch2_fs_open_incremental(struct bch_sb_handle *sb,
@@ -1786,6 +1976,7 @@ static const char *__bch2_fs_open_incremental(struct bch_sb_handle *sb,
 	const char *err;
 	struct bch_fs *c;
 	bool allocated_fs = false;
+	int ret;
 
 	err = bch2_sb_validate(sb);
 	if (err)
@@ -1796,7 +1987,7 @@ static const char *__bch2_fs_open_incremental(struct bch_sb_handle *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 {
@@ -1811,22 +2002,19 @@ static const char *__bch2_fs_open_incremental(struct bch_sb_handle *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);
-		if (err)
+		err = "error starting filesystem";
+		ret = bch2_fs_start(c);
+		if (ret)
 			goto err;
 	}
 
-	err = __bch2_fs_online(c);
-	if (err)
-		goto err;
-
 	closure_put(&c->cl);
 	mutex_unlock(&bch_fs_list_lock);
 
@@ -1864,6 +2052,7 @@ static void bcachefs_exit(void)
 	bch2_debug_exit();
 	bch2_vfs_exit();
 	bch2_chardev_exit();
+	bch2_btree_key_cache_exit();
 	if (bcachefs_kset)
 		kset_unregister(bcachefs_kset);
 }
@@ -1871,9 +2060,9 @@ static void bcachefs_exit(void)
 static int __init bcachefs_init(void)
 {
 	bch2_bkey_pack_test();
-	bch2_inode_pack_test();
 
 	if (!(bcachefs_kset = kset_create_and_add("bcachefs", NULL, fs_kobj)) ||
+	    bch2_btree_key_cache_init() ||
 	    bch2_chardev_init() ||
 	    bch2_vfs_init() ||
 	    bch2_debug_init())