X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Fsuper.c;h=a3c61a7480bebc602e5cf037a704d916e873f9b8;hb=38a93ad6b29474e281b466b29c971420e796ba4f;hp=16b8cbfc973a389ced7393b5a7614bec13156b4a;hpb=018de5aa899937a9dc3bc8cb9819cb218a59abf3;p=bcachefs-tools-debian

diff --git a/libbcachefs/super.c b/libbcachefs/super.c
index 16b8cbf..a3c61a7 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"
@@ -19,6 +22,7 @@
 #include "compress.h"
 #include "debug.h"
 #include "disk_groups.h"
+#include "ec.h"
 #include "error.h"
 #include "fs.h"
 #include "fs-io.h"
@@ -26,18 +30,18 @@
 #include "inode.h"
 #include "io.h"
 #include "journal.h"
-#include "journal_io.h"
 #include "journal_reclaim.h"
-#include "keylist.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 "super.h"
 #include "super-io.h"
 #include "sysfs.h"
-#include "tier.h"
 
 #include <linux/backing-dev.h>
 #include <linux/blkdev.h>
@@ -45,7 +49,6 @@
 #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>
@@ -145,43 +148,21 @@ 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);
 
 	rcu_read_lock();
-	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 {
-		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];
-
-		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;
-			}
-		}
-	}
+	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: */
@@ -201,53 +182,66 @@ int bch2_congested(void *data, int bdi_bits)
  * - allocator depends on the journal (when it rewrites prios and gens)
  */
 
-static void bch_fs_mark_clean(struct bch_fs *c)
-{
-	if (!bch2_journal_error(&c->journal) &&
-	    !test_bit(BCH_FS_ERROR, &c->flags) &&
-	    !test_bit(BCH_FS_EMERGENCY_RO, &c->flags)) {
-		mutex_lock(&c->sb_lock);
-		SET_BCH_SB_CLEAN(c->disk_sb.sb, true);
-		bch2_write_super(c);
-		mutex_unlock(&c->sb_lock);
-	}
-}
-
 static void __bch2_fs_read_only(struct bch_fs *c)
 {
 	struct bch_dev *ca;
-	unsigned i;
+	unsigned i, clean_passes = 0;
 
 	bch2_rebalance_stop(c);
-
-	for_each_member_device(ca, c, i)
-		bch2_copygc_stop(ca);
-
+	bch2_copygc_stop(c);
 	bch2_gc_thread_stop(c);
 
 	/*
 	 * Flush journal before stopping allocators, because flushing journal
 	 * blacklist entries involves allocating new btree nodes:
 	 */
-	bch2_journal_flush_pins(&c->journal, U64_MAX - 1);
-
-	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:
+	 * 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);
 
-	if (!test_bit(BCH_FS_EMERGENCY_RO, &c->flags))
-		bch2_btree_verify_flushed(c);
+	clear_bit(BCH_FS_ALLOCATOR_RUNNING, &c->flags);
+	clear_bit(BCH_FS_ALLOCATOR_STOPPING, &c->flags);
 
 	bch2_fs_journal_stop(&c->journal);
 
@@ -255,8 +249,7 @@ static void __bch2_fs_read_only(struct bch_fs *c)
 	 * 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);
+	bch2_btree_flush_all_writes(c);
 
 	/*
 	 * After stopping journal:
@@ -275,12 +268,12 @@ 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)) {
+		BUG_ON(c->journal.reclaim_thread);
 		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
@@ -292,6 +285,7 @@ void bch2_fs_read_only(struct bch_fs *c)
 	 */
 	percpu_ref_kill(&c->writes);
 
+	cancel_work_sync(&c->ec_stripe_delete_work);
 	cancel_delayed_work(&c->pd_controllers_update);
 
 	/*
@@ -311,13 +305,22 @@ void bch2_fs_read_only(struct bch_fs *c)
 
 	__bch2_fs_read_only(c);
 
-	bch_fs_mark_clean(c);
-
 	wait_event(bch_read_only_wait,
 		   test_bit(BCH_FS_WRITE_DISABLE_COMPLETE, &c->flags));
 
 	clear_bit(BCH_FS_WRITE_DISABLE_COMPLETE, &c->flags);
-	c->state = BCH_FS_RO;
+
+	if (!bch2_journal_error(&c->journal) &&
+	    !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);
+	}
+
+	clear_bit(BCH_FS_RW, &c->flags);
 }
 
 static void bch2_fs_read_only_work(struct work_struct *work)
@@ -325,9 +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)
@@ -339,89 +342,168 @@ 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_delayed_work(&c->pd_controllers_update, 5 * HZ);
+
+	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_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;
+
+	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);
+	ret = bch2_journal_reclaim_start(&c->journal);
+	if (ret) {
+		bch_err(c, "error starting journal reclaim: %i", ret);
+		return ret;
+	}
+
+	if (!early) {
+		ret = bch2_fs_read_write_late(c);
+		if (ret)
 			goto err;
-		}
+	}
 
-	err = "error starting rebalance thread";
-	if (bch2_rebalance_start(c))
-		goto err;
+	percpu_ref_reinit(&c->writes);
+	set_bit(BCH_FS_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)
 {
-#define BCH_TIME_STAT(name)				\
-	bch2_time_stats_exit(&c->name##_time);
-	BCH_TIME_STATS()
-#undef BCH_TIME_STAT
+	unsigned i;
+	int cpu;
+
+	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_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_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);
+	kfree(c->usage_scratch);
+	for (i = 0; i < ARRAY_SIZE(c->usage); i++)
+		free_percpu(c->usage[i]);
+	kfree(c->usage_base);
+
+	if (c->btree_iters_bufs)
+		for_each_possible_cpu(cpu)
+			kfree(per_cpu_ptr(c->btree_iters_bufs, cpu)->iter);
+
+	free_percpu(c->btree_iters_bufs);
+	free_percpu(c->pcpu);
+	mempool_exit(&c->large_bkey_pool);
 	mempool_exit(&c->btree_bounce_pool);
 	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(c->replicas.entries);
+	kfree(c->replicas_gc.entries);
 	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->copygc_wq)
 		destroy_workqueue(c->copygc_wq);
@@ -438,18 +520,23 @@ static void bch2_fs_release(struct kobject *kobj)
 {
 	struct bch_fs *c = container_of(kobj, struct bch_fs, kobj);
 
-	bch2_fs_free(c);
+	__bch2_fs_free(c);
 }
 
-void bch2_fs_stop(struct bch_fs *c)
+void __bch2_fs_stop(struct bch_fs *c)
 {
 	struct bch_dev *ca;
 	unsigned i;
 
-	mutex_lock(&c->state_lock);
-	BUG_ON(c->state == BCH_FS_STOPPING);
-	c->state = BCH_FS_STOPPING;
-	mutex_unlock(&c->state_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 &&
@@ -467,19 +554,6 @@ void bch2_fs_stop(struct bch_fs *c)
 	kobject_put(&c->opts_dir);
 	kobject_put(&c->internal);
 
-	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);
-
-	mutex_lock(&c->state_lock);
-	__bch2_fs_read_only(c);
-	mutex_unlock(&c->state_lock);
-
-	bch_fs_mark_clean(c);
-
 	/* btree prefetch might have kicked off reads in the background: */
 	bch2_btree_flush_all_reads(c);
 
@@ -490,13 +564,37 @@ void bch2_fs_stop(struct bch_fs *c)
 	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_free_super(&c->devs[i]->disk_sb);
+}
+
+void bch2_fs_free(struct bch_fs *c)
+{
+	unsigned i;
+
+	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)
+{
+	__bch2_fs_stop(c);
+	bch2_fs_free(c);
+}
+
 static const char *bch2_fs_online(struct bch_fs *c)
 {
 	struct bch_dev *ca;
@@ -525,7 +623,7 @@ static const char *bch2_fs_online(struct bch_fs *c)
 	    bch2_opts_create_sysfs_files(&c->opts_dir))
 		return "error creating sysfs objects";
 
-	mutex_lock(&c->state_lock);
+	down_write(&c->state_lock);
 
 	err = "error creating sysfs objects";
 	__for_each_member_device(ca, c, i, NULL)
@@ -535,7 +633,7 @@ static const char *bch2_fs_online(struct bch_fs *c)
 	list_add(&c->list, &bch_fs_list);
 	err = NULL;
 err:
-	mutex_unlock(&c->state_lock);
+	up_write(&c->state_lock);
 	return err;
 }
 
@@ -554,10 +652,18 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 
 	__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);
@@ -565,20 +671,19 @@ 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)				\
-	bch2_time_stats_init(&c->name##_time);
-	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_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);
 
@@ -586,23 +691,45 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 	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);
+
+	spin_lock_init(&c->ec_stripes_heap_lock);
+
 	seqcount_init(&c->gc_pos_lock);
 
+	seqcount_init(&c->usage_lock);
+
+	sema_init(&c->io_in_flight, 64);
+
 	c->copy_gc_enabled		= 1;
-	c->rebalance_enabled		= 1;
-	c->rebalance_percent		= 10;
+	c->rebalance.enabled		= 1;
 	c->promote_whole_extents	= true;
 
-	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->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)) {
@@ -621,40 +748,45 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_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 = sizeof(struct btree_node_iter_large) +
+	iter_size = sizeof(struct sort_iter) +
 		(btree_blocks(c) + 1) * 2 *
-		sizeof(struct btree_node_iter_set);
+		sizeof(struct sort_iter_set);
+
+	c->inode_shard_bits = ilog2(roundup_pow_of_two(num_possible_cpus()));
 
 	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)) ||
+				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)) ||
+	    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) ||
-	    !(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_iters_bufs = alloc_percpu(struct btree_iter_buf)) ||
+	    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_io_init(c) ||
 	    bch2_fs_encryption_init(c) ||
 	    bch2_fs_compress_init(c) ||
+	    bch2_fs_ec_init(c) ||
 	    bch2_fs_fsio_init(c))
 		goto err;
 
@@ -664,17 +796,13 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 		    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);
-
-	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);
+	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);
 
 	mutex_lock(&bch_fs_list_lock);
 	err = bch2_fs_online(c);
@@ -692,207 +820,109 @@ err:
 	goto out;
 }
 
-const char *bch2_fs_start(struct bch_fs *c)
+noinline_for_stack
+static void print_mount_opts(struct bch_fs *c)
+{
+	enum bch_opt_id i;
+	char buf[512];
+	struct printbuf p = PBUF(buf);
+	bool first = true;
+
+	strcpy(buf, "(null)");
+
+	if (c->opts.read_only) {
+		pr_buf(&p, "ro");
+		first = false;
+	}
+
+	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);
+
+		if (!(opt->mode & OPT_MOUNT))
+			continue;
+
+		if (v == bch2_opt_get_by_id(&bch2_opts_default, i))
+			continue;
+
+		if (!first)
+			pr_buf(&p, ",");
+		first = false;
+		bch2_opt_to_text(&p, c, opt, v, OPT_SHOW_MOUNT_STYLE);
+	}
+
+	bch_info(c, "mounted with opts: %s", buf);
+}
+
+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.sb)) {
-		ret = bch2_journal_read(c, &journal);
-		if (ret)
-			goto err;
-
-		j = &list_entry(journal.prev, struct journal_replay, list)->j;
-
-		c->bucket_clock[READ].hand = le16_to_cpu(j->read_clock);
-		c->bucket_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_fs_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_fs_journal_start(&c->journal);
-
-		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 (enabled_qtypes(c)) {
-			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);
-
-		ret = bch2_initial_gc(c, &journal);
-		if (ret)
-			goto err;
-
-		err = "unable to allocate journal buckets";
-		for_each_online_member(ca, c, i)
-			if (bch2_dev_journal_alloc(ca)) {
-				percpu_ref_put(&ca->io_ref);
-				goto err;
-			}
-
-		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_fs_journal_start(&c->journal);
-		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 (enabled_qtypes(c)) {
-			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:
+	 */
+	for_each_online_member(ca, c, i)
+		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.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.sb, true);
-	SET_BCH_SB_CLEAN(c->disk_sb.sb, false);
-
-	bch2_write_super(c);
-	mutex_unlock(&c->sb_lock);
-
-	set_bit(BCH_FS_STARTED, &c->flags);
-
-	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");
@@ -919,8 +949,8 @@ fsck_err:
 		break;
 	}
 
-	BUG_ON(!err);
-	set_bit(BCH_FS_ERROR, &c->flags);
+	if (ret >= 0)
+		ret = -EIO;
 	goto out;
 }
 
@@ -987,6 +1017,7 @@ 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]);
@@ -1080,11 +1111,6 @@ static struct bch_dev *__bch2_dev_alloc(struct bch_fs *c,
 
 	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);
 
 	bch2_time_stats_init(&ca->io_latency[READ]);
@@ -1093,10 +1119,14 @@ static struct bch_dev *__bch2_dev_alloc(struct bch_fs *c,
 	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) ||
+	    !(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) ||
@@ -1184,14 +1214,6 @@ static int __bch2_dev_attach_bdev(struct bch_dev *ca, struct bch_sb_handle *sb)
 		ca->disk_sb.bdev->bd_holder = ca;
 	memset(sb, 0, sizeof(*sb));
 
-	if (ca->fs)
-		mutex_lock(&ca->fs->sb_lock);
-
-	bch2_mark_dev_superblock(ca->fs, ca, BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE);
-
-	if (ca->fs)
-		mutex_unlock(&ca->fs->sb_lock);
-
 	percpu_ref_reinit(&ca->io_ref);
 
 	return 0;
@@ -1217,6 +1239,8 @@ static int bch2_dev_attach_bdev(struct bch_fs *c, struct bch_sb_handle *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);
@@ -1240,7 +1264,6 @@ 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;
 
@@ -1276,9 +1299,7 @@ bool bch2_dev_state_allowed(struct bch_fs *c, struct bch_dev *ca,
 		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(s, flags);
+		return bch2_have_enough_devs(c, new_online_devs, flags, false);
 	default:
 		BUG();
 	}
@@ -1286,14 +1307,18 @@ 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)
+		flags |= BCH_FORCE_IF_DEGRADED;
 
-	if (!c->opts.degraded) {
+	if (!c->opts.degraded &&
+	    !c->opts.very_degraded) {
 		mutex_lock(&c->sb_lock);
 		mi = bch2_sb_get_members(c->disk_sb.sb);
 
@@ -1313,14 +1338,16 @@ 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(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:
@@ -1328,6 +1355,8 @@ 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)
@@ -1342,9 +1371,6 @@ 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";
-
 	return NULL;
 }
 
@@ -1385,24 +1411,53 @@ 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: */
 
+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)) {
 		bch_err(ca, "Cannot remove without losing data");
@@ -1411,11 +1466,6 @@ int bch2_dev_remove(struct bch_fs *c, struct bch_dev *ca, int flags)
 
 	__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);
@@ -1428,23 +1478,7 @@ int bch2_dev_remove(struct bch_fs *c, struct bch_dev *ca, int flags)
 		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;
@@ -1454,12 +1488,34 @@ 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);
@@ -1482,12 +1538,15 @@ int bch2_dev_remove(struct bch_fs *c, struct bch_dev *ca, int flags)
 	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;
 }
 
@@ -1529,12 +1588,26 @@ int bch2_dev_add(struct bch_fs *c, const char *path)
 		return ret;
 	}
 
+	/*
+	 * 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...
+	 */
+	bch2_mark_dev_superblock(NULL, ca, 0);
+
 	err = "journal alloc failed";
 	ret = bch2_dev_journal_alloc(ca);
 	if (ret)
 		goto err;
 
-	mutex_lock(&c->state_lock);
+	down_write(&c->state_lock);
 	mutex_lock(&c->sb_lock);
 
 	err = "insufficient space in new superblock";
@@ -1578,7 +1651,7 @@ have_slot:
 	/* success: */
 
 	mi->members[dev_idx] = dev_mi;
-	mi->members[dev_idx].last_mount = cpu_to_le64(ktime_get_seconds());
+	mi->members[dev_idx].last_mount = cpu_to_le64(ktime_get_real_seconds());
 	c->disk_sb.sb->nr_devices	= nr_devices;
 
 	ca->disk_sb.sb->dev_idx	= dev_idx;
@@ -1587,18 +1660,25 @@ have_slot:
 	bch2_write_super(c);
 	mutex_unlock(&c->sb_lock);
 
+	bch2_dev_usage_journal_reserve(c);
+
+	err = "error marking superblock";
+	ret = bch2_trans_mark_dev_sb(c, NULL, ca);
+	if (ret)
+		goto err_late;
+
 	if (ca->mi.state == BCH_MEMBER_STATE_RW) {
 		err = __bch2_dev_read_write(c, ca);
 		if (err)
 			goto err_late;
 	}
 
-	mutex_unlock(&c->state_lock);
+	up_write(&c->state_lock);
 	return 0;
 
 err_unlock:
 	mutex_unlock(&c->sb_lock);
-	mutex_unlock(&c->state_lock);
+	up_write(&c->state_lock);
 err:
 	if (ca)
 		bch2_dev_free(ca);
@@ -1606,6 +1686,7 @@ err:
 	bch_err(c, "Unable to add device: %s", err);
 	return ret;
 err_late:
+	up_write(&c->state_lock);
 	bch_err(c, "Error going rw after adding device: %s", err);
 	return -EINVAL;
 }
@@ -1615,16 +1696,17 @@ 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;
 	}
 
@@ -1640,16 +1722,31 @@ int bch2_dev_online(struct bch_fs *c, const char *path)
 	}
 
 	ca = bch_dev_locked(c, dev_idx);
+
+	if (bch2_trans_mark_dev_sb(c, NULL, 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;
@@ -1657,23 +1754,23 @@ 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)) {
 		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;
 }
 
@@ -1682,7 +1779,7 @@ 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");
@@ -1713,14 +1810,13 @@ 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 block_device *bdev = lookup_bdev(path);
 	struct bch_dev *ca;
 	unsigned i;
@@ -1745,6 +1841,7 @@ 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;
@@ -1780,10 +1877,24 @@ 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;
@@ -1792,22 +1903,22 @@ 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_attach_bdev(c, &sb[i])) {
-			mutex_unlock(&c->state_lock);
+			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;
 	}
 out:
 	kfree(sb);
@@ -1834,6 +1945,7 @@ 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)
@@ -1866,8 +1978,9 @@ static const char *__bch2_fs_open_incremental(struct bch_sb_handle *sb,
 	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;
 	}
 
@@ -1908,6 +2021,7 @@ 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);
 }
@@ -1915,9 +2029,9 @@ 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())