xref: /linux/fs/bcachefs/super.c (revision db1ecca22edf27c5a3dd66af406c88b5b5ac7cc1)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * bcachefs setup/teardown code, and some metadata io - read a superblock and
4  * figure out what to do with it.
5  *
6  * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
7  * Copyright 2012 Google, Inc.
8  */
9 
10 #include "bcachefs.h"
11 #include "alloc_background.h"
12 #include "alloc_foreground.h"
13 #include "bkey_sort.h"
14 #include "btree_cache.h"
15 #include "btree_gc.h"
16 #include "btree_journal_iter.h"
17 #include "btree_key_cache.h"
18 #include "btree_update_interior.h"
19 #include "btree_io.h"
20 #include "btree_write_buffer.h"
21 #include "buckets_waiting_for_journal.h"
22 #include "chardev.h"
23 #include "checksum.h"
24 #include "clock.h"
25 #include "compress.h"
26 #include "counters.h"
27 #include "debug.h"
28 #include "disk_groups.h"
29 #include "ec.h"
30 #include "errcode.h"
31 #include "error.h"
32 #include "fs.h"
33 #include "fs-io.h"
34 #include "fs-io-buffered.h"
35 #include "fs-io-direct.h"
36 #include "fsck.h"
37 #include "inode.h"
38 #include "io_read.h"
39 #include "io_write.h"
40 #include "journal.h"
41 #include "journal_reclaim.h"
42 #include "journal_seq_blacklist.h"
43 #include "move.h"
44 #include "migrate.h"
45 #include "movinggc.h"
46 #include "nocow_locking.h"
47 #include "quota.h"
48 #include "rebalance.h"
49 #include "recovery.h"
50 #include "replicas.h"
51 #include "sb-clean.h"
52 #include "sb-errors.h"
53 #include "sb-members.h"
54 #include "snapshot.h"
55 #include "subvolume.h"
56 #include "super.h"
57 #include "super-io.h"
58 #include "sysfs.h"
59 #include "trace.h"
60 
61 #include <linux/backing-dev.h>
62 #include <linux/blkdev.h>
63 #include <linux/debugfs.h>
64 #include <linux/device.h>
65 #include <linux/idr.h>
66 #include <linux/module.h>
67 #include <linux/percpu.h>
68 #include <linux/random.h>
69 #include <linux/sysfs.h>
70 #include <crypto/hash.h>
71 
72 MODULE_LICENSE("GPL");
73 MODULE_AUTHOR("Kent Overstreet <kent.overstreet@gmail.com>");
74 MODULE_DESCRIPTION("bcachefs filesystem");
75 
76 #define KTYPE(type)							\
77 static const struct attribute_group type ## _group = {			\
78 	.attrs = type ## _files						\
79 };									\
80 									\
81 static const struct attribute_group *type ## _groups[] = {		\
82 	&type ## _group,						\
83 	NULL								\
84 };									\
85 									\
86 static const struct kobj_type type ## _ktype = {			\
87 	.release	= type ## _release,				\
88 	.sysfs_ops	= &type ## _sysfs_ops,				\
89 	.default_groups = type ## _groups				\
90 }
91 
92 static void bch2_fs_release(struct kobject *);
93 static void bch2_dev_release(struct kobject *);
94 static void bch2_fs_counters_release(struct kobject *k)
95 {
96 }
97 
98 static void bch2_fs_internal_release(struct kobject *k)
99 {
100 }
101 
102 static void bch2_fs_opts_dir_release(struct kobject *k)
103 {
104 }
105 
106 static void bch2_fs_time_stats_release(struct kobject *k)
107 {
108 }
109 
110 KTYPE(bch2_fs);
111 KTYPE(bch2_fs_counters);
112 KTYPE(bch2_fs_internal);
113 KTYPE(bch2_fs_opts_dir);
114 KTYPE(bch2_fs_time_stats);
115 KTYPE(bch2_dev);
116 
117 static struct kset *bcachefs_kset;
118 static LIST_HEAD(bch_fs_list);
119 static DEFINE_MUTEX(bch_fs_list_lock);
120 
121 DECLARE_WAIT_QUEUE_HEAD(bch2_read_only_wait);
122 
123 static void bch2_dev_free(struct bch_dev *);
124 static int bch2_dev_alloc(struct bch_fs *, unsigned);
125 static int bch2_dev_sysfs_online(struct bch_fs *, struct bch_dev *);
126 static void __bch2_dev_read_only(struct bch_fs *, struct bch_dev *);
127 
128 struct bch_fs *bch2_dev_to_fs(dev_t dev)
129 {
130 	struct bch_fs *c;
131 	struct bch_dev *ca;
132 	unsigned i;
133 
134 	mutex_lock(&bch_fs_list_lock);
135 	rcu_read_lock();
136 
137 	list_for_each_entry(c, &bch_fs_list, list)
138 		for_each_member_device_rcu(ca, c, i, NULL)
139 			if (ca->disk_sb.bdev && ca->disk_sb.bdev->bd_dev == dev) {
140 				closure_get(&c->cl);
141 				goto found;
142 			}
143 	c = NULL;
144 found:
145 	rcu_read_unlock();
146 	mutex_unlock(&bch_fs_list_lock);
147 
148 	return c;
149 }
150 
151 static struct bch_fs *__bch2_uuid_to_fs(__uuid_t uuid)
152 {
153 	struct bch_fs *c;
154 
155 	lockdep_assert_held(&bch_fs_list_lock);
156 
157 	list_for_each_entry(c, &bch_fs_list, list)
158 		if (!memcmp(&c->disk_sb.sb->uuid, &uuid, sizeof(uuid)))
159 			return c;
160 
161 	return NULL;
162 }
163 
164 struct bch_fs *bch2_uuid_to_fs(__uuid_t uuid)
165 {
166 	struct bch_fs *c;
167 
168 	mutex_lock(&bch_fs_list_lock);
169 	c = __bch2_uuid_to_fs(uuid);
170 	if (c)
171 		closure_get(&c->cl);
172 	mutex_unlock(&bch_fs_list_lock);
173 
174 	return c;
175 }
176 
177 static void bch2_dev_usage_journal_reserve(struct bch_fs *c)
178 {
179 	struct bch_dev *ca;
180 	unsigned i, nr = 0, u64s =
181 		((sizeof(struct jset_entry_dev_usage) +
182 		  sizeof(struct jset_entry_dev_usage_type) * BCH_DATA_NR)) /
183 		sizeof(u64);
184 
185 	rcu_read_lock();
186 	for_each_member_device_rcu(ca, c, i, NULL)
187 		nr++;
188 	rcu_read_unlock();
189 
190 	bch2_journal_entry_res_resize(&c->journal,
191 			&c->dev_usage_journal_res, u64s * nr);
192 }
193 
194 /* Filesystem RO/RW: */
195 
196 /*
197  * For startup/shutdown of RW stuff, the dependencies are:
198  *
199  * - foreground writes depend on copygc and rebalance (to free up space)
200  *
201  * - copygc and rebalance depend on mark and sweep gc (they actually probably
202  *   don't because they either reserve ahead of time or don't block if
203  *   allocations fail, but allocations can require mark and sweep gc to run
204  *   because of generation number wraparound)
205  *
206  * - all of the above depends on the allocator threads
207  *
208  * - allocator depends on the journal (when it rewrites prios and gens)
209  */
210 
211 static void __bch2_fs_read_only(struct bch_fs *c)
212 {
213 	struct bch_dev *ca;
214 	unsigned i, clean_passes = 0;
215 	u64 seq = 0;
216 
217 	bch2_fs_ec_stop(c);
218 	bch2_open_buckets_stop(c, NULL, true);
219 	bch2_rebalance_stop(c);
220 	bch2_copygc_stop(c);
221 	bch2_gc_thread_stop(c);
222 	bch2_fs_ec_flush(c);
223 
224 	bch_verbose(c, "flushing journal and stopping allocators, journal seq %llu",
225 		    journal_cur_seq(&c->journal));
226 
227 	do {
228 		clean_passes++;
229 
230 		if (bch2_btree_interior_updates_flush(c) ||
231 		    bch2_journal_flush_all_pins(&c->journal) ||
232 		    bch2_btree_flush_all_writes(c) ||
233 		    seq != atomic64_read(&c->journal.seq)) {
234 			seq = atomic64_read(&c->journal.seq);
235 			clean_passes = 0;
236 		}
237 	} while (clean_passes < 2);
238 
239 	bch_verbose(c, "flushing journal and stopping allocators complete, journal seq %llu",
240 		    journal_cur_seq(&c->journal));
241 
242 	if (test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags) &&
243 	    !test_bit(BCH_FS_EMERGENCY_RO, &c->flags))
244 		set_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags);
245 	bch2_fs_journal_stop(&c->journal);
246 
247 	/*
248 	 * After stopping journal:
249 	 */
250 	for_each_member_device(ca, c, i)
251 		bch2_dev_allocator_remove(c, ca);
252 }
253 
254 #ifndef BCH_WRITE_REF_DEBUG
255 static void bch2_writes_disabled(struct percpu_ref *writes)
256 {
257 	struct bch_fs *c = container_of(writes, struct bch_fs, writes);
258 
259 	set_bit(BCH_FS_WRITE_DISABLE_COMPLETE, &c->flags);
260 	wake_up(&bch2_read_only_wait);
261 }
262 #endif
263 
264 void bch2_fs_read_only(struct bch_fs *c)
265 {
266 	if (!test_bit(BCH_FS_RW, &c->flags)) {
267 		bch2_journal_reclaim_stop(&c->journal);
268 		return;
269 	}
270 
271 	BUG_ON(test_bit(BCH_FS_WRITE_DISABLE_COMPLETE, &c->flags));
272 
273 	/*
274 	 * Block new foreground-end write operations from starting - any new
275 	 * writes will return -EROFS:
276 	 */
277 	set_bit(BCH_FS_GOING_RO, &c->flags);
278 #ifndef BCH_WRITE_REF_DEBUG
279 	percpu_ref_kill(&c->writes);
280 #else
281 	for (unsigned i = 0; i < BCH_WRITE_REF_NR; i++)
282 		bch2_write_ref_put(c, i);
283 #endif
284 
285 	/*
286 	 * If we're not doing an emergency shutdown, we want to wait on
287 	 * outstanding writes to complete so they don't see spurious errors due
288 	 * to shutting down the allocator:
289 	 *
290 	 * If we are doing an emergency shutdown outstanding writes may
291 	 * hang until we shutdown the allocator so we don't want to wait
292 	 * on outstanding writes before shutting everything down - but
293 	 * we do need to wait on them before returning and signalling
294 	 * that going RO is complete:
295 	 */
296 	wait_event(bch2_read_only_wait,
297 		   test_bit(BCH_FS_WRITE_DISABLE_COMPLETE, &c->flags) ||
298 		   test_bit(BCH_FS_EMERGENCY_RO, &c->flags));
299 
300 	__bch2_fs_read_only(c);
301 
302 	wait_event(bch2_read_only_wait,
303 		   test_bit(BCH_FS_WRITE_DISABLE_COMPLETE, &c->flags));
304 
305 	clear_bit(BCH_FS_WRITE_DISABLE_COMPLETE, &c->flags);
306 	clear_bit(BCH_FS_GOING_RO, &c->flags);
307 
308 	if (!bch2_journal_error(&c->journal) &&
309 	    !test_bit(BCH_FS_ERROR, &c->flags) &&
310 	    !test_bit(BCH_FS_EMERGENCY_RO, &c->flags) &&
311 	    test_bit(BCH_FS_STARTED, &c->flags) &&
312 	    test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags) &&
313 	    !c->opts.norecovery) {
314 		BUG_ON(c->journal.last_empty_seq != journal_cur_seq(&c->journal));
315 		BUG_ON(atomic_read(&c->btree_cache.dirty));
316 		BUG_ON(atomic_long_read(&c->btree_key_cache.nr_dirty));
317 		BUG_ON(c->btree_write_buffer.state.nr);
318 
319 		bch_verbose(c, "marking filesystem clean");
320 		bch2_fs_mark_clean(c);
321 	}
322 
323 	clear_bit(BCH_FS_RW, &c->flags);
324 }
325 
326 static void bch2_fs_read_only_work(struct work_struct *work)
327 {
328 	struct bch_fs *c =
329 		container_of(work, struct bch_fs, read_only_work);
330 
331 	down_write(&c->state_lock);
332 	bch2_fs_read_only(c);
333 	up_write(&c->state_lock);
334 }
335 
336 static void bch2_fs_read_only_async(struct bch_fs *c)
337 {
338 	queue_work(system_long_wq, &c->read_only_work);
339 }
340 
341 bool bch2_fs_emergency_read_only(struct bch_fs *c)
342 {
343 	bool ret = !test_and_set_bit(BCH_FS_EMERGENCY_RO, &c->flags);
344 
345 	bch2_journal_halt(&c->journal);
346 	bch2_fs_read_only_async(c);
347 
348 	wake_up(&bch2_read_only_wait);
349 	return ret;
350 }
351 
352 static int bch2_fs_read_write_late(struct bch_fs *c)
353 {
354 	int ret;
355 
356 	/*
357 	 * Data move operations can't run until after check_snapshots has
358 	 * completed, and bch2_snapshot_is_ancestor() is available.
359 	 *
360 	 * Ideally we'd start copygc/rebalance earlier instead of waiting for
361 	 * all of recovery/fsck to complete:
362 	 */
363 	ret = bch2_copygc_start(c);
364 	if (ret) {
365 		bch_err(c, "error starting copygc thread");
366 		return ret;
367 	}
368 
369 	ret = bch2_rebalance_start(c);
370 	if (ret) {
371 		bch_err(c, "error starting rebalance thread");
372 		return ret;
373 	}
374 
375 	return 0;
376 }
377 
378 static int __bch2_fs_read_write(struct bch_fs *c, bool early)
379 {
380 	struct bch_dev *ca;
381 	unsigned i;
382 	int ret;
383 
384 	if (test_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags)) {
385 		bch_err(c, "cannot go rw, unfixed btree errors");
386 		return -BCH_ERR_erofs_unfixed_errors;
387 	}
388 
389 	if (test_bit(BCH_FS_RW, &c->flags))
390 		return 0;
391 
392 	if (c->opts.norecovery)
393 		return -BCH_ERR_erofs_norecovery;
394 
395 	/*
396 	 * nochanges is used for fsck -n mode - we have to allow going rw
397 	 * during recovery for that to work:
398 	 */
399 	if (c->opts.nochanges && (!early || c->opts.read_only))
400 		return -BCH_ERR_erofs_nochanges;
401 
402 	bch_info(c, "going read-write");
403 
404 	ret = bch2_sb_members_v2_init(c);
405 	if (ret)
406 		goto err;
407 
408 	ret = bch2_fs_mark_dirty(c);
409 	if (ret)
410 		goto err;
411 
412 	clear_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags);
413 
414 	/*
415 	 * First journal write must be a flush write: after a clean shutdown we
416 	 * don't read the journal, so the first journal write may end up
417 	 * overwriting whatever was there previously, and there must always be
418 	 * at least one non-flush write in the journal or recovery will fail:
419 	 */
420 	set_bit(JOURNAL_NEED_FLUSH_WRITE, &c->journal.flags);
421 
422 	for_each_rw_member(ca, c, i)
423 		bch2_dev_allocator_add(c, ca);
424 	bch2_recalc_capacity(c);
425 
426 	set_bit(BCH_FS_RW, &c->flags);
427 	set_bit(BCH_FS_WAS_RW, &c->flags);
428 
429 #ifndef BCH_WRITE_REF_DEBUG
430 	percpu_ref_reinit(&c->writes);
431 #else
432 	for (i = 0; i < BCH_WRITE_REF_NR; i++) {
433 		BUG_ON(atomic_long_read(&c->writes[i]));
434 		atomic_long_inc(&c->writes[i]);
435 	}
436 #endif
437 
438 	ret = bch2_gc_thread_start(c);
439 	if (ret) {
440 		bch_err(c, "error starting gc thread");
441 		return ret;
442 	}
443 
444 	ret = bch2_journal_reclaim_start(&c->journal);
445 	if (ret)
446 		goto err;
447 
448 	if (!early) {
449 		ret = bch2_fs_read_write_late(c);
450 		if (ret)
451 			goto err;
452 	}
453 
454 	bch2_do_discards(c);
455 	bch2_do_invalidates(c);
456 	bch2_do_stripe_deletes(c);
457 	bch2_do_pending_node_rewrites(c);
458 	return 0;
459 err:
460 	if (test_bit(BCH_FS_RW, &c->flags))
461 		bch2_fs_read_only(c);
462 	else
463 		__bch2_fs_read_only(c);
464 	return ret;
465 }
466 
467 int bch2_fs_read_write(struct bch_fs *c)
468 {
469 	return __bch2_fs_read_write(c, false);
470 }
471 
472 int bch2_fs_read_write_early(struct bch_fs *c)
473 {
474 	lockdep_assert_held(&c->state_lock);
475 
476 	return __bch2_fs_read_write(c, true);
477 }
478 
479 /* Filesystem startup/shutdown: */
480 
481 static void __bch2_fs_free(struct bch_fs *c)
482 {
483 	unsigned i;
484 
485 	for (i = 0; i < BCH_TIME_STAT_NR; i++)
486 		bch2_time_stats_exit(&c->times[i]);
487 
488 	bch2_free_pending_node_rewrites(c);
489 	bch2_fs_sb_errors_exit(c);
490 	bch2_fs_counters_exit(c);
491 	bch2_fs_snapshots_exit(c);
492 	bch2_fs_quota_exit(c);
493 	bch2_fs_fs_io_direct_exit(c);
494 	bch2_fs_fs_io_buffered_exit(c);
495 	bch2_fs_fsio_exit(c);
496 	bch2_fs_ec_exit(c);
497 	bch2_fs_encryption_exit(c);
498 	bch2_fs_nocow_locking_exit(c);
499 	bch2_fs_io_write_exit(c);
500 	bch2_fs_io_read_exit(c);
501 	bch2_fs_buckets_waiting_for_journal_exit(c);
502 	bch2_fs_btree_interior_update_exit(c);
503 	bch2_fs_btree_iter_exit(c);
504 	bch2_fs_btree_key_cache_exit(&c->btree_key_cache);
505 	bch2_fs_btree_cache_exit(c);
506 	bch2_fs_replicas_exit(c);
507 	bch2_fs_journal_exit(&c->journal);
508 	bch2_io_clock_exit(&c->io_clock[WRITE]);
509 	bch2_io_clock_exit(&c->io_clock[READ]);
510 	bch2_fs_compress_exit(c);
511 	bch2_journal_keys_put_initial(c);
512 	BUG_ON(atomic_read(&c->journal_keys.ref));
513 	bch2_fs_btree_write_buffer_exit(c);
514 	percpu_free_rwsem(&c->mark_lock);
515 	free_percpu(c->online_reserved);
516 
517 	darray_exit(&c->btree_roots_extra);
518 	free_percpu(c->pcpu);
519 	mempool_exit(&c->large_bkey_pool);
520 	mempool_exit(&c->btree_bounce_pool);
521 	bioset_exit(&c->btree_bio);
522 	mempool_exit(&c->fill_iter);
523 #ifndef BCH_WRITE_REF_DEBUG
524 	percpu_ref_exit(&c->writes);
525 #endif
526 	kfree(rcu_dereference_protected(c->disk_groups, 1));
527 	kfree(c->journal_seq_blacklist_table);
528 	kfree(c->unused_inode_hints);
529 
530 	if (c->write_ref_wq)
531 		destroy_workqueue(c->write_ref_wq);
532 	if (c->io_complete_wq)
533 		destroy_workqueue(c->io_complete_wq);
534 	if (c->copygc_wq)
535 		destroy_workqueue(c->copygc_wq);
536 	if (c->btree_io_complete_wq)
537 		destroy_workqueue(c->btree_io_complete_wq);
538 	if (c->btree_update_wq)
539 		destroy_workqueue(c->btree_update_wq);
540 
541 	bch2_free_super(&c->disk_sb);
542 	kvpfree(c, sizeof(*c));
543 	module_put(THIS_MODULE);
544 }
545 
546 static void bch2_fs_release(struct kobject *kobj)
547 {
548 	struct bch_fs *c = container_of(kobj, struct bch_fs, kobj);
549 
550 	__bch2_fs_free(c);
551 }
552 
553 void __bch2_fs_stop(struct bch_fs *c)
554 {
555 	struct bch_dev *ca;
556 	unsigned i;
557 
558 	bch_verbose(c, "shutting down");
559 
560 	set_bit(BCH_FS_STOPPING, &c->flags);
561 
562 	cancel_work_sync(&c->journal_seq_blacklist_gc_work);
563 
564 	down_write(&c->state_lock);
565 	bch2_fs_read_only(c);
566 	up_write(&c->state_lock);
567 
568 	for_each_member_device(ca, c, i)
569 		if (ca->kobj.state_in_sysfs &&
570 		    ca->disk_sb.bdev)
571 			sysfs_remove_link(bdev_kobj(ca->disk_sb.bdev), "bcachefs");
572 
573 	if (c->kobj.state_in_sysfs)
574 		kobject_del(&c->kobj);
575 
576 	bch2_fs_debug_exit(c);
577 	bch2_fs_chardev_exit(c);
578 
579 	kobject_put(&c->counters_kobj);
580 	kobject_put(&c->time_stats);
581 	kobject_put(&c->opts_dir);
582 	kobject_put(&c->internal);
583 
584 	/* btree prefetch might have kicked off reads in the background: */
585 	bch2_btree_flush_all_reads(c);
586 
587 	for_each_member_device(ca, c, i)
588 		cancel_work_sync(&ca->io_error_work);
589 
590 	cancel_work_sync(&c->read_only_work);
591 }
592 
593 void bch2_fs_free(struct bch_fs *c)
594 {
595 	unsigned i;
596 
597 	mutex_lock(&bch_fs_list_lock);
598 	list_del(&c->list);
599 	mutex_unlock(&bch_fs_list_lock);
600 
601 	closure_sync(&c->cl);
602 	closure_debug_destroy(&c->cl);
603 
604 	for (i = 0; i < c->sb.nr_devices; i++) {
605 		struct bch_dev *ca = rcu_dereference_protected(c->devs[i], true);
606 
607 		if (ca) {
608 			bch2_free_super(&ca->disk_sb);
609 			bch2_dev_free(ca);
610 		}
611 	}
612 
613 	bch_verbose(c, "shutdown complete");
614 
615 	kobject_put(&c->kobj);
616 }
617 
618 void bch2_fs_stop(struct bch_fs *c)
619 {
620 	__bch2_fs_stop(c);
621 	bch2_fs_free(c);
622 }
623 
624 static int bch2_fs_online(struct bch_fs *c)
625 {
626 	struct bch_dev *ca;
627 	unsigned i;
628 	int ret = 0;
629 
630 	lockdep_assert_held(&bch_fs_list_lock);
631 
632 	if (__bch2_uuid_to_fs(c->sb.uuid)) {
633 		bch_err(c, "filesystem UUID already open");
634 		return -EINVAL;
635 	}
636 
637 	ret = bch2_fs_chardev_init(c);
638 	if (ret) {
639 		bch_err(c, "error creating character device");
640 		return ret;
641 	}
642 
643 	bch2_fs_debug_init(c);
644 
645 	ret = kobject_add(&c->kobj, NULL, "%pU", c->sb.user_uuid.b) ?:
646 	    kobject_add(&c->internal, &c->kobj, "internal") ?:
647 	    kobject_add(&c->opts_dir, &c->kobj, "options") ?:
648 	    kobject_add(&c->time_stats, &c->kobj, "time_stats") ?:
649 	    kobject_add(&c->counters_kobj, &c->kobj, "counters") ?:
650 	    bch2_opts_create_sysfs_files(&c->opts_dir);
651 	if (ret) {
652 		bch_err(c, "error creating sysfs objects");
653 		return ret;
654 	}
655 
656 	down_write(&c->state_lock);
657 
658 	for_each_member_device(ca, c, i) {
659 		ret = bch2_dev_sysfs_online(c, ca);
660 		if (ret) {
661 			bch_err(c, "error creating sysfs objects");
662 			percpu_ref_put(&ca->ref);
663 			goto err;
664 		}
665 	}
666 
667 	BUG_ON(!list_empty(&c->list));
668 	list_add(&c->list, &bch_fs_list);
669 err:
670 	up_write(&c->state_lock);
671 	return ret;
672 }
673 
674 static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
675 {
676 	struct bch_fs *c;
677 	struct printbuf name = PRINTBUF;
678 	unsigned i, iter_size;
679 	int ret = 0;
680 
681 	c = kvpmalloc(sizeof(struct bch_fs), GFP_KERNEL|__GFP_ZERO);
682 	if (!c) {
683 		c = ERR_PTR(-BCH_ERR_ENOMEM_fs_alloc);
684 		goto out;
685 	}
686 
687 	__module_get(THIS_MODULE);
688 
689 	closure_init(&c->cl, NULL);
690 
691 	c->kobj.kset = bcachefs_kset;
692 	kobject_init(&c->kobj, &bch2_fs_ktype);
693 	kobject_init(&c->internal, &bch2_fs_internal_ktype);
694 	kobject_init(&c->opts_dir, &bch2_fs_opts_dir_ktype);
695 	kobject_init(&c->time_stats, &bch2_fs_time_stats_ktype);
696 	kobject_init(&c->counters_kobj, &bch2_fs_counters_ktype);
697 
698 	c->minor		= -1;
699 	c->disk_sb.fs_sb	= true;
700 
701 	init_rwsem(&c->state_lock);
702 	mutex_init(&c->sb_lock);
703 	mutex_init(&c->replicas_gc_lock);
704 	mutex_init(&c->btree_root_lock);
705 	INIT_WORK(&c->read_only_work, bch2_fs_read_only_work);
706 
707 	init_rwsem(&c->gc_lock);
708 	mutex_init(&c->gc_gens_lock);
709 	atomic_set(&c->journal_keys.ref, 1);
710 	c->journal_keys.initial_ref_held = true;
711 
712 	for (i = 0; i < BCH_TIME_STAT_NR; i++)
713 		bch2_time_stats_init(&c->times[i]);
714 
715 	bch2_fs_copygc_init(c);
716 	bch2_fs_btree_key_cache_init_early(&c->btree_key_cache);
717 	bch2_fs_btree_interior_update_init_early(c);
718 	bch2_fs_allocator_background_init(c);
719 	bch2_fs_allocator_foreground_init(c);
720 	bch2_fs_rebalance_init(c);
721 	bch2_fs_quota_init(c);
722 	bch2_fs_ec_init_early(c);
723 	bch2_fs_move_init(c);
724 	bch2_fs_sb_errors_init_early(c);
725 
726 	INIT_LIST_HEAD(&c->list);
727 
728 	mutex_init(&c->usage_scratch_lock);
729 
730 	mutex_init(&c->bio_bounce_pages_lock);
731 	mutex_init(&c->snapshot_table_lock);
732 	init_rwsem(&c->snapshot_create_lock);
733 
734 	spin_lock_init(&c->btree_write_error_lock);
735 
736 	INIT_WORK(&c->journal_seq_blacklist_gc_work,
737 		  bch2_blacklist_entries_gc);
738 
739 	INIT_LIST_HEAD(&c->journal_iters);
740 
741 	INIT_LIST_HEAD(&c->fsck_error_msgs);
742 	mutex_init(&c->fsck_error_msgs_lock);
743 
744 	seqcount_init(&c->gc_pos_lock);
745 
746 	seqcount_init(&c->usage_lock);
747 
748 	sema_init(&c->io_in_flight, 128);
749 
750 	INIT_LIST_HEAD(&c->vfs_inodes_list);
751 	mutex_init(&c->vfs_inodes_lock);
752 
753 	c->copy_gc_enabled		= 1;
754 	c->rebalance.enabled		= 1;
755 	c->promote_whole_extents	= true;
756 
757 	c->journal.flush_write_time	= &c->times[BCH_TIME_journal_flush_write];
758 	c->journal.noflush_write_time	= &c->times[BCH_TIME_journal_noflush_write];
759 	c->journal.blocked_time		= &c->times[BCH_TIME_blocked_journal];
760 	c->journal.flush_seq_time	= &c->times[BCH_TIME_journal_flush_seq];
761 
762 	bch2_fs_btree_cache_init_early(&c->btree_cache);
763 
764 	mutex_init(&c->sectors_available_lock);
765 
766 	ret = percpu_init_rwsem(&c->mark_lock);
767 	if (ret)
768 		goto err;
769 
770 	mutex_lock(&c->sb_lock);
771 	ret = bch2_sb_to_fs(c, sb);
772 	mutex_unlock(&c->sb_lock);
773 
774 	if (ret)
775 		goto err;
776 
777 	pr_uuid(&name, c->sb.user_uuid.b);
778 	strscpy(c->name, name.buf, sizeof(c->name));
779 	printbuf_exit(&name);
780 
781 	ret = name.allocation_failure ? -BCH_ERR_ENOMEM_fs_name_alloc : 0;
782 	if (ret)
783 		goto err;
784 
785 	/* Compat: */
786 	if (le16_to_cpu(sb->version) <= bcachefs_metadata_version_inode_v2 &&
787 	    !BCH_SB_JOURNAL_FLUSH_DELAY(sb))
788 		SET_BCH_SB_JOURNAL_FLUSH_DELAY(sb, 1000);
789 
790 	if (le16_to_cpu(sb->version) <= bcachefs_metadata_version_inode_v2 &&
791 	    !BCH_SB_JOURNAL_RECLAIM_DELAY(sb))
792 		SET_BCH_SB_JOURNAL_RECLAIM_DELAY(sb, 100);
793 
794 	c->opts = bch2_opts_default;
795 	ret = bch2_opts_from_sb(&c->opts, sb);
796 	if (ret)
797 		goto err;
798 
799 	bch2_opts_apply(&c->opts, opts);
800 
801 	c->btree_key_cache_btrees |= 1U << BTREE_ID_alloc;
802 	if (c->opts.inodes_use_key_cache)
803 		c->btree_key_cache_btrees |= 1U << BTREE_ID_inodes;
804 	c->btree_key_cache_btrees |= 1U << BTREE_ID_logged_ops;
805 
806 	c->block_bits		= ilog2(block_sectors(c));
807 	c->btree_foreground_merge_threshold = BTREE_FOREGROUND_MERGE_THRESHOLD(c);
808 
809 	if (bch2_fs_init_fault("fs_alloc")) {
810 		bch_err(c, "fs_alloc fault injected");
811 		ret = -EFAULT;
812 		goto err;
813 	}
814 
815 	iter_size = sizeof(struct sort_iter) +
816 		(btree_blocks(c) + 1) * 2 *
817 		sizeof(struct sort_iter_set);
818 
819 	c->inode_shard_bits = ilog2(roundup_pow_of_two(num_possible_cpus()));
820 
821 	if (!(c->btree_update_wq = alloc_workqueue("bcachefs",
822 				WQ_FREEZABLE|WQ_UNBOUND|WQ_MEM_RECLAIM, 512)) ||
823 	    !(c->btree_io_complete_wq = alloc_workqueue("bcachefs_btree_io",
824 				WQ_FREEZABLE|WQ_MEM_RECLAIM, 1)) ||
825 	    !(c->copygc_wq = alloc_workqueue("bcachefs_copygc",
826 				WQ_FREEZABLE|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE, 1)) ||
827 	    !(c->io_complete_wq = alloc_workqueue("bcachefs_io",
828 				WQ_FREEZABLE|WQ_HIGHPRI|WQ_MEM_RECLAIM, 1)) ||
829 	    !(c->write_ref_wq = alloc_workqueue("bcachefs_write_ref",
830 				WQ_FREEZABLE, 0)) ||
831 #ifndef BCH_WRITE_REF_DEBUG
832 	    percpu_ref_init(&c->writes, bch2_writes_disabled,
833 			    PERCPU_REF_INIT_DEAD, GFP_KERNEL) ||
834 #endif
835 	    mempool_init_kmalloc_pool(&c->fill_iter, 1, iter_size) ||
836 	    bioset_init(&c->btree_bio, 1,
837 			max(offsetof(struct btree_read_bio, bio),
838 			    offsetof(struct btree_write_bio, wbio.bio)),
839 			BIOSET_NEED_BVECS) ||
840 	    !(c->pcpu = alloc_percpu(struct bch_fs_pcpu)) ||
841 	    !(c->online_reserved = alloc_percpu(u64)) ||
842 	    mempool_init_kvpmalloc_pool(&c->btree_bounce_pool, 1,
843 					btree_bytes(c)) ||
844 	    mempool_init_kmalloc_pool(&c->large_bkey_pool, 1, 2048) ||
845 	    !(c->unused_inode_hints = kcalloc(1U << c->inode_shard_bits,
846 					      sizeof(u64), GFP_KERNEL))) {
847 		ret = -BCH_ERR_ENOMEM_fs_other_alloc;
848 		goto err;
849 	}
850 
851 	ret = bch2_fs_counters_init(c) ?:
852 	    bch2_fs_sb_errors_init(c) ?:
853 	    bch2_io_clock_init(&c->io_clock[READ]) ?:
854 	    bch2_io_clock_init(&c->io_clock[WRITE]) ?:
855 	    bch2_fs_journal_init(&c->journal) ?:
856 	    bch2_fs_replicas_init(c) ?:
857 	    bch2_fs_btree_cache_init(c) ?:
858 	    bch2_fs_btree_key_cache_init(&c->btree_key_cache) ?:
859 	    bch2_fs_btree_iter_init(c) ?:
860 	    bch2_fs_btree_interior_update_init(c) ?:
861 	    bch2_fs_buckets_waiting_for_journal_init(c) ?:
862 	    bch2_fs_btree_write_buffer_init(c) ?:
863 	    bch2_fs_subvolumes_init(c) ?:
864 	    bch2_fs_io_read_init(c) ?:
865 	    bch2_fs_io_write_init(c) ?:
866 	    bch2_fs_nocow_locking_init(c) ?:
867 	    bch2_fs_encryption_init(c) ?:
868 	    bch2_fs_compress_init(c) ?:
869 	    bch2_fs_ec_init(c) ?:
870 	    bch2_fs_fsio_init(c) ?:
871 	    bch2_fs_fs_io_buffered_init(c) ?:
872 	    bch2_fs_fs_io_direct_init(c);
873 	if (ret)
874 		goto err;
875 
876 	for (i = 0; i < c->sb.nr_devices; i++)
877 		if (bch2_dev_exists(c->disk_sb.sb, i) &&
878 		    bch2_dev_alloc(c, i)) {
879 			ret = -EEXIST;
880 			goto err;
881 		}
882 
883 	bch2_journal_entry_res_resize(&c->journal,
884 			&c->btree_root_journal_res,
885 			BTREE_ID_NR * (JSET_KEYS_U64s + BKEY_BTREE_PTR_U64s_MAX));
886 	bch2_dev_usage_journal_reserve(c);
887 	bch2_journal_entry_res_resize(&c->journal,
888 			&c->clock_journal_res,
889 			(sizeof(struct jset_entry_clock) / sizeof(u64)) * 2);
890 
891 	mutex_lock(&bch_fs_list_lock);
892 	ret = bch2_fs_online(c);
893 	mutex_unlock(&bch_fs_list_lock);
894 
895 	if (ret)
896 		goto err;
897 out:
898 	return c;
899 err:
900 	bch2_fs_free(c);
901 	c = ERR_PTR(ret);
902 	goto out;
903 }
904 
905 noinline_for_stack
906 static void print_mount_opts(struct bch_fs *c)
907 {
908 	enum bch_opt_id i;
909 	struct printbuf p = PRINTBUF;
910 	bool first = true;
911 
912 	prt_str(&p, "mounting version ");
913 	bch2_version_to_text(&p, c->sb.version);
914 
915 	if (c->opts.read_only) {
916 		prt_str(&p, " opts=");
917 		first = false;
918 		prt_printf(&p, "ro");
919 	}
920 
921 	for (i = 0; i < bch2_opts_nr; i++) {
922 		const struct bch_option *opt = &bch2_opt_table[i];
923 		u64 v = bch2_opt_get_by_id(&c->opts, i);
924 
925 		if (!(opt->flags & OPT_MOUNT))
926 			continue;
927 
928 		if (v == bch2_opt_get_by_id(&bch2_opts_default, i))
929 			continue;
930 
931 		prt_str(&p, first ? " opts=" : ",");
932 		first = false;
933 		bch2_opt_to_text(&p, c, c->disk_sb.sb, opt, v, OPT_SHOW_MOUNT_STYLE);
934 	}
935 
936 	bch_info(c, "%s", p.buf);
937 	printbuf_exit(&p);
938 }
939 
940 int bch2_fs_start(struct bch_fs *c)
941 {
942 	struct bch_dev *ca;
943 	time64_t now = ktime_get_real_seconds();
944 	unsigned i;
945 	int ret;
946 
947 	print_mount_opts(c);
948 
949 	down_write(&c->state_lock);
950 
951 	BUG_ON(test_bit(BCH_FS_STARTED, &c->flags));
952 
953 	mutex_lock(&c->sb_lock);
954 
955 	ret = bch2_sb_members_v2_init(c);
956 	if (ret) {
957 		mutex_unlock(&c->sb_lock);
958 		goto err;
959 	}
960 
961 	for_each_online_member(ca, c, i)
962 		bch2_members_v2_get_mut(c->disk_sb.sb, i)->last_mount = cpu_to_le64(now);
963 
964 	mutex_unlock(&c->sb_lock);
965 
966 	for_each_rw_member(ca, c, i)
967 		bch2_dev_allocator_add(c, ca);
968 	bch2_recalc_capacity(c);
969 
970 	ret = BCH_SB_INITIALIZED(c->disk_sb.sb)
971 		? bch2_fs_recovery(c)
972 		: bch2_fs_initialize(c);
973 	if (ret)
974 		goto err;
975 
976 	ret = bch2_opts_check_may_set(c);
977 	if (ret)
978 		goto err;
979 
980 	if (bch2_fs_init_fault("fs_start")) {
981 		bch_err(c, "fs_start fault injected");
982 		ret = -EINVAL;
983 		goto err;
984 	}
985 
986 	set_bit(BCH_FS_STARTED, &c->flags);
987 
988 	if (c->opts.read_only || c->opts.nochanges) {
989 		bch2_fs_read_only(c);
990 	} else {
991 		ret = !test_bit(BCH_FS_RW, &c->flags)
992 			? bch2_fs_read_write(c)
993 			: bch2_fs_read_write_late(c);
994 		if (ret)
995 			goto err;
996 	}
997 
998 	ret = 0;
999 out:
1000 	up_write(&c->state_lock);
1001 	return ret;
1002 err:
1003 	bch_err_msg(c, ret, "starting filesystem");
1004 	goto out;
1005 }
1006 
1007 static int bch2_dev_may_add(struct bch_sb *sb, struct bch_fs *c)
1008 {
1009 	struct bch_member m = bch2_sb_member_get(sb, sb->dev_idx);
1010 
1011 	if (le16_to_cpu(sb->block_size) != block_sectors(c))
1012 		return -BCH_ERR_mismatched_block_size;
1013 
1014 	if (le16_to_cpu(m.bucket_size) <
1015 	    BCH_SB_BTREE_NODE_SIZE(c->disk_sb.sb))
1016 		return -BCH_ERR_bucket_size_too_small;
1017 
1018 	return 0;
1019 }
1020 
1021 static int bch2_dev_in_fs(struct bch_sb *fs, struct bch_sb *sb)
1022 {
1023 	struct bch_sb *newest =
1024 		le64_to_cpu(fs->seq) > le64_to_cpu(sb->seq) ? fs : sb;
1025 
1026 	if (!uuid_equal(&fs->uuid, &sb->uuid))
1027 		return -BCH_ERR_device_not_a_member_of_filesystem;
1028 
1029 	if (!bch2_dev_exists(newest, sb->dev_idx))
1030 		return -BCH_ERR_device_has_been_removed;
1031 
1032 	if (fs->block_size != sb->block_size)
1033 		return -BCH_ERR_mismatched_block_size;
1034 
1035 	return 0;
1036 }
1037 
1038 /* Device startup/shutdown: */
1039 
1040 static void bch2_dev_release(struct kobject *kobj)
1041 {
1042 	struct bch_dev *ca = container_of(kobj, struct bch_dev, kobj);
1043 
1044 	kfree(ca);
1045 }
1046 
1047 static void bch2_dev_free(struct bch_dev *ca)
1048 {
1049 	cancel_work_sync(&ca->io_error_work);
1050 
1051 	if (ca->kobj.state_in_sysfs &&
1052 	    ca->disk_sb.bdev)
1053 		sysfs_remove_link(bdev_kobj(ca->disk_sb.bdev), "bcachefs");
1054 
1055 	if (ca->kobj.state_in_sysfs)
1056 		kobject_del(&ca->kobj);
1057 
1058 	bch2_free_super(&ca->disk_sb);
1059 	bch2_dev_journal_exit(ca);
1060 
1061 	free_percpu(ca->io_done);
1062 	bioset_exit(&ca->replica_set);
1063 	bch2_dev_buckets_free(ca);
1064 	free_page((unsigned long) ca->sb_read_scratch);
1065 
1066 	bch2_time_stats_exit(&ca->io_latency[WRITE]);
1067 	bch2_time_stats_exit(&ca->io_latency[READ]);
1068 
1069 	percpu_ref_exit(&ca->io_ref);
1070 	percpu_ref_exit(&ca->ref);
1071 	kobject_put(&ca->kobj);
1072 }
1073 
1074 static void __bch2_dev_offline(struct bch_fs *c, struct bch_dev *ca)
1075 {
1076 
1077 	lockdep_assert_held(&c->state_lock);
1078 
1079 	if (percpu_ref_is_zero(&ca->io_ref))
1080 		return;
1081 
1082 	__bch2_dev_read_only(c, ca);
1083 
1084 	reinit_completion(&ca->io_ref_completion);
1085 	percpu_ref_kill(&ca->io_ref);
1086 	wait_for_completion(&ca->io_ref_completion);
1087 
1088 	if (ca->kobj.state_in_sysfs) {
1089 		sysfs_remove_link(bdev_kobj(ca->disk_sb.bdev), "bcachefs");
1090 		sysfs_remove_link(&ca->kobj, "block");
1091 	}
1092 
1093 	bch2_free_super(&ca->disk_sb);
1094 	bch2_dev_journal_exit(ca);
1095 }
1096 
1097 static void bch2_dev_ref_complete(struct percpu_ref *ref)
1098 {
1099 	struct bch_dev *ca = container_of(ref, struct bch_dev, ref);
1100 
1101 	complete(&ca->ref_completion);
1102 }
1103 
1104 static void bch2_dev_io_ref_complete(struct percpu_ref *ref)
1105 {
1106 	struct bch_dev *ca = container_of(ref, struct bch_dev, io_ref);
1107 
1108 	complete(&ca->io_ref_completion);
1109 }
1110 
1111 static int bch2_dev_sysfs_online(struct bch_fs *c, struct bch_dev *ca)
1112 {
1113 	int ret;
1114 
1115 	if (!c->kobj.state_in_sysfs)
1116 		return 0;
1117 
1118 	if (!ca->kobj.state_in_sysfs) {
1119 		ret = kobject_add(&ca->kobj, &c->kobj,
1120 				  "dev-%u", ca->dev_idx);
1121 		if (ret)
1122 			return ret;
1123 	}
1124 
1125 	if (ca->disk_sb.bdev) {
1126 		struct kobject *block = bdev_kobj(ca->disk_sb.bdev);
1127 
1128 		ret = sysfs_create_link(block, &ca->kobj, "bcachefs");
1129 		if (ret)
1130 			return ret;
1131 
1132 		ret = sysfs_create_link(&ca->kobj, block, "block");
1133 		if (ret)
1134 			return ret;
1135 	}
1136 
1137 	return 0;
1138 }
1139 
1140 static struct bch_dev *__bch2_dev_alloc(struct bch_fs *c,
1141 					struct bch_member *member)
1142 {
1143 	struct bch_dev *ca;
1144 	unsigned i;
1145 
1146 	ca = kzalloc(sizeof(*ca), GFP_KERNEL);
1147 	if (!ca)
1148 		return NULL;
1149 
1150 	kobject_init(&ca->kobj, &bch2_dev_ktype);
1151 	init_completion(&ca->ref_completion);
1152 	init_completion(&ca->io_ref_completion);
1153 
1154 	init_rwsem(&ca->bucket_lock);
1155 
1156 	INIT_WORK(&ca->io_error_work, bch2_io_error_work);
1157 
1158 	bch2_time_stats_init(&ca->io_latency[READ]);
1159 	bch2_time_stats_init(&ca->io_latency[WRITE]);
1160 
1161 	ca->mi = bch2_mi_to_cpu(member);
1162 
1163 	for (i = 0; i < ARRAY_SIZE(member->errors); i++)
1164 		atomic64_set(&ca->errors[i], le64_to_cpu(member->errors[i]));
1165 
1166 	ca->uuid = member->uuid;
1167 
1168 	ca->nr_btree_reserve = DIV_ROUND_UP(BTREE_NODE_RESERVE,
1169 			     ca->mi.bucket_size / btree_sectors(c));
1170 
1171 	if (percpu_ref_init(&ca->ref, bch2_dev_ref_complete,
1172 			    0, GFP_KERNEL) ||
1173 	    percpu_ref_init(&ca->io_ref, bch2_dev_io_ref_complete,
1174 			    PERCPU_REF_INIT_DEAD, GFP_KERNEL) ||
1175 	    !(ca->sb_read_scratch = (void *) __get_free_page(GFP_KERNEL)) ||
1176 	    bch2_dev_buckets_alloc(c, ca) ||
1177 	    bioset_init(&ca->replica_set, 4,
1178 			offsetof(struct bch_write_bio, bio), 0) ||
1179 	    !(ca->io_done	= alloc_percpu(*ca->io_done)))
1180 		goto err;
1181 
1182 	return ca;
1183 err:
1184 	bch2_dev_free(ca);
1185 	return NULL;
1186 }
1187 
1188 static void bch2_dev_attach(struct bch_fs *c, struct bch_dev *ca,
1189 			    unsigned dev_idx)
1190 {
1191 	ca->dev_idx = dev_idx;
1192 	__set_bit(ca->dev_idx, ca->self.d);
1193 	scnprintf(ca->name, sizeof(ca->name), "dev-%u", dev_idx);
1194 
1195 	ca->fs = c;
1196 	rcu_assign_pointer(c->devs[ca->dev_idx], ca);
1197 
1198 	if (bch2_dev_sysfs_online(c, ca))
1199 		pr_warn("error creating sysfs objects");
1200 }
1201 
1202 static int bch2_dev_alloc(struct bch_fs *c, unsigned dev_idx)
1203 {
1204 	struct bch_member member = bch2_sb_member_get(c->disk_sb.sb, dev_idx);
1205 	struct bch_dev *ca = NULL;
1206 	int ret = 0;
1207 
1208 	if (bch2_fs_init_fault("dev_alloc"))
1209 		goto err;
1210 
1211 	ca = __bch2_dev_alloc(c, &member);
1212 	if (!ca)
1213 		goto err;
1214 
1215 	ca->fs = c;
1216 
1217 	bch2_dev_attach(c, ca, dev_idx);
1218 	return ret;
1219 err:
1220 	if (ca)
1221 		bch2_dev_free(ca);
1222 	return -BCH_ERR_ENOMEM_dev_alloc;
1223 }
1224 
1225 static int __bch2_dev_attach_bdev(struct bch_dev *ca, struct bch_sb_handle *sb)
1226 {
1227 	unsigned ret;
1228 
1229 	if (bch2_dev_is_online(ca)) {
1230 		bch_err(ca, "already have device online in slot %u",
1231 			sb->sb->dev_idx);
1232 		return -BCH_ERR_device_already_online;
1233 	}
1234 
1235 	if (get_capacity(sb->bdev->bd_disk) <
1236 	    ca->mi.bucket_size * ca->mi.nbuckets) {
1237 		bch_err(ca, "cannot online: device too small");
1238 		return -BCH_ERR_device_size_too_small;
1239 	}
1240 
1241 	BUG_ON(!percpu_ref_is_zero(&ca->io_ref));
1242 
1243 	ret = bch2_dev_journal_init(ca, sb->sb);
1244 	if (ret)
1245 		return ret;
1246 
1247 	/* Commit: */
1248 	ca->disk_sb = *sb;
1249 	memset(sb, 0, sizeof(*sb));
1250 
1251 	ca->dev = ca->disk_sb.bdev->bd_dev;
1252 
1253 	percpu_ref_reinit(&ca->io_ref);
1254 
1255 	return 0;
1256 }
1257 
1258 static int bch2_dev_attach_bdev(struct bch_fs *c, struct bch_sb_handle *sb)
1259 {
1260 	struct bch_dev *ca;
1261 	int ret;
1262 
1263 	lockdep_assert_held(&c->state_lock);
1264 
1265 	if (le64_to_cpu(sb->sb->seq) >
1266 	    le64_to_cpu(c->disk_sb.sb->seq))
1267 		bch2_sb_to_fs(c, sb->sb);
1268 
1269 	BUG_ON(sb->sb->dev_idx >= c->sb.nr_devices ||
1270 	       !c->devs[sb->sb->dev_idx]);
1271 
1272 	ca = bch_dev_locked(c, sb->sb->dev_idx);
1273 
1274 	ret = __bch2_dev_attach_bdev(ca, sb);
1275 	if (ret)
1276 		return ret;
1277 
1278 	bch2_dev_sysfs_online(c, ca);
1279 
1280 	if (c->sb.nr_devices == 1)
1281 		snprintf(c->name, sizeof(c->name), "%pg", ca->disk_sb.bdev);
1282 	snprintf(ca->name, sizeof(ca->name), "%pg", ca->disk_sb.bdev);
1283 
1284 	rebalance_wakeup(c);
1285 	return 0;
1286 }
1287 
1288 /* Device management: */
1289 
1290 /*
1291  * Note: this function is also used by the error paths - when a particular
1292  * device sees an error, we call it to determine whether we can just set the
1293  * device RO, or - if this function returns false - we'll set the whole
1294  * filesystem RO:
1295  *
1296  * XXX: maybe we should be more explicit about whether we're changing state
1297  * because we got an error or what have you?
1298  */
1299 bool bch2_dev_state_allowed(struct bch_fs *c, struct bch_dev *ca,
1300 			    enum bch_member_state new_state, int flags)
1301 {
1302 	struct bch_devs_mask new_online_devs;
1303 	struct bch_dev *ca2;
1304 	int i, nr_rw = 0, required;
1305 
1306 	lockdep_assert_held(&c->state_lock);
1307 
1308 	switch (new_state) {
1309 	case BCH_MEMBER_STATE_rw:
1310 		return true;
1311 	case BCH_MEMBER_STATE_ro:
1312 		if (ca->mi.state != BCH_MEMBER_STATE_rw)
1313 			return true;
1314 
1315 		/* do we have enough devices to write to?  */
1316 		for_each_member_device(ca2, c, i)
1317 			if (ca2 != ca)
1318 				nr_rw += ca2->mi.state == BCH_MEMBER_STATE_rw;
1319 
1320 		required = max(!(flags & BCH_FORCE_IF_METADATA_DEGRADED)
1321 			       ? c->opts.metadata_replicas
1322 			       : c->opts.metadata_replicas_required,
1323 			       !(flags & BCH_FORCE_IF_DATA_DEGRADED)
1324 			       ? c->opts.data_replicas
1325 			       : c->opts.data_replicas_required);
1326 
1327 		return nr_rw >= required;
1328 	case BCH_MEMBER_STATE_failed:
1329 	case BCH_MEMBER_STATE_spare:
1330 		if (ca->mi.state != BCH_MEMBER_STATE_rw &&
1331 		    ca->mi.state != BCH_MEMBER_STATE_ro)
1332 			return true;
1333 
1334 		/* do we have enough devices to read from?  */
1335 		new_online_devs = bch2_online_devs(c);
1336 		__clear_bit(ca->dev_idx, new_online_devs.d);
1337 
1338 		return bch2_have_enough_devs(c, new_online_devs, flags, false);
1339 	default:
1340 		BUG();
1341 	}
1342 }
1343 
1344 static bool bch2_fs_may_start(struct bch_fs *c)
1345 {
1346 	struct bch_dev *ca;
1347 	unsigned i, flags = 0;
1348 
1349 	if (c->opts.very_degraded)
1350 		flags |= BCH_FORCE_IF_DEGRADED|BCH_FORCE_IF_LOST;
1351 
1352 	if (c->opts.degraded)
1353 		flags |= BCH_FORCE_IF_DEGRADED;
1354 
1355 	if (!c->opts.degraded &&
1356 	    !c->opts.very_degraded) {
1357 		mutex_lock(&c->sb_lock);
1358 
1359 		for (i = 0; i < c->disk_sb.sb->nr_devices; i++) {
1360 			if (!bch2_dev_exists(c->disk_sb.sb, i))
1361 				continue;
1362 
1363 			ca = bch_dev_locked(c, i);
1364 
1365 			if (!bch2_dev_is_online(ca) &&
1366 			    (ca->mi.state == BCH_MEMBER_STATE_rw ||
1367 			     ca->mi.state == BCH_MEMBER_STATE_ro)) {
1368 				mutex_unlock(&c->sb_lock);
1369 				return false;
1370 			}
1371 		}
1372 		mutex_unlock(&c->sb_lock);
1373 	}
1374 
1375 	return bch2_have_enough_devs(c, bch2_online_devs(c), flags, true);
1376 }
1377 
1378 static void __bch2_dev_read_only(struct bch_fs *c, struct bch_dev *ca)
1379 {
1380 	/*
1381 	 * The allocator thread itself allocates btree nodes, so stop it first:
1382 	 */
1383 	bch2_dev_allocator_remove(c, ca);
1384 	bch2_dev_journal_stop(&c->journal, ca);
1385 }
1386 
1387 static void __bch2_dev_read_write(struct bch_fs *c, struct bch_dev *ca)
1388 {
1389 	lockdep_assert_held(&c->state_lock);
1390 
1391 	BUG_ON(ca->mi.state != BCH_MEMBER_STATE_rw);
1392 
1393 	bch2_dev_allocator_add(c, ca);
1394 	bch2_recalc_capacity(c);
1395 }
1396 
1397 int __bch2_dev_set_state(struct bch_fs *c, struct bch_dev *ca,
1398 			 enum bch_member_state new_state, int flags)
1399 {
1400 	struct bch_member *m;
1401 	int ret = 0;
1402 
1403 	if (ca->mi.state == new_state)
1404 		return 0;
1405 
1406 	if (!bch2_dev_state_allowed(c, ca, new_state, flags))
1407 		return -BCH_ERR_device_state_not_allowed;
1408 
1409 	if (new_state != BCH_MEMBER_STATE_rw)
1410 		__bch2_dev_read_only(c, ca);
1411 
1412 	bch_notice(ca, "%s", bch2_member_states[new_state]);
1413 
1414 	mutex_lock(&c->sb_lock);
1415 	m = bch2_members_v2_get_mut(c->disk_sb.sb, ca->dev_idx);
1416 	SET_BCH_MEMBER_STATE(m, new_state);
1417 	bch2_write_super(c);
1418 	mutex_unlock(&c->sb_lock);
1419 
1420 	if (new_state == BCH_MEMBER_STATE_rw)
1421 		__bch2_dev_read_write(c, ca);
1422 
1423 	rebalance_wakeup(c);
1424 
1425 	return ret;
1426 }
1427 
1428 int bch2_dev_set_state(struct bch_fs *c, struct bch_dev *ca,
1429 		       enum bch_member_state new_state, int flags)
1430 {
1431 	int ret;
1432 
1433 	down_write(&c->state_lock);
1434 	ret = __bch2_dev_set_state(c, ca, new_state, flags);
1435 	up_write(&c->state_lock);
1436 
1437 	return ret;
1438 }
1439 
1440 /* Device add/removal: */
1441 
1442 static int bch2_dev_remove_alloc(struct bch_fs *c, struct bch_dev *ca)
1443 {
1444 	struct bpos start	= POS(ca->dev_idx, 0);
1445 	struct bpos end		= POS(ca->dev_idx, U64_MAX);
1446 	int ret;
1447 
1448 	/*
1449 	 * We clear the LRU and need_discard btrees first so that we don't race
1450 	 * with bch2_do_invalidates() and bch2_do_discards()
1451 	 */
1452 	ret =   bch2_btree_delete_range(c, BTREE_ID_lru, start, end,
1453 					BTREE_TRIGGER_NORUN, NULL) ?:
1454 		bch2_btree_delete_range(c, BTREE_ID_need_discard, start, end,
1455 					BTREE_TRIGGER_NORUN, NULL) ?:
1456 		bch2_btree_delete_range(c, BTREE_ID_freespace, start, end,
1457 					BTREE_TRIGGER_NORUN, NULL) ?:
1458 		bch2_btree_delete_range(c, BTREE_ID_backpointers, start, end,
1459 					BTREE_TRIGGER_NORUN, NULL) ?:
1460 		bch2_btree_delete_range(c, BTREE_ID_alloc, start, end,
1461 					BTREE_TRIGGER_NORUN, NULL) ?:
1462 		bch2_btree_delete_range(c, BTREE_ID_bucket_gens, start, end,
1463 					BTREE_TRIGGER_NORUN, NULL);
1464 	if (ret)
1465 		bch_err_msg(c, ret, "removing dev alloc info");
1466 
1467 	return ret;
1468 }
1469 
1470 int bch2_dev_remove(struct bch_fs *c, struct bch_dev *ca, int flags)
1471 {
1472 	struct bch_member *m;
1473 	unsigned dev_idx = ca->dev_idx, data;
1474 	int ret;
1475 
1476 	down_write(&c->state_lock);
1477 
1478 	/*
1479 	 * We consume a reference to ca->ref, regardless of whether we succeed
1480 	 * or fail:
1481 	 */
1482 	percpu_ref_put(&ca->ref);
1483 
1484 	if (!bch2_dev_state_allowed(c, ca, BCH_MEMBER_STATE_failed, flags)) {
1485 		bch_err(ca, "Cannot remove without losing data");
1486 		ret = -BCH_ERR_device_state_not_allowed;
1487 		goto err;
1488 	}
1489 
1490 	__bch2_dev_read_only(c, ca);
1491 
1492 	ret = bch2_dev_data_drop(c, ca->dev_idx, flags);
1493 	if (ret) {
1494 		bch_err_msg(ca, ret, "dropping data");
1495 		goto err;
1496 	}
1497 
1498 	ret = bch2_dev_remove_alloc(c, ca);
1499 	if (ret) {
1500 		bch_err_msg(ca, ret, "deleting alloc info");
1501 		goto err;
1502 	}
1503 
1504 	ret = bch2_journal_flush_device_pins(&c->journal, ca->dev_idx);
1505 	if (ret) {
1506 		bch_err_msg(ca, ret, "flushing journal");
1507 		goto err;
1508 	}
1509 
1510 	ret = bch2_journal_flush(&c->journal);
1511 	if (ret) {
1512 		bch_err(ca, "journal error");
1513 		goto err;
1514 	}
1515 
1516 	ret = bch2_replicas_gc2(c);
1517 	if (ret) {
1518 		bch_err_msg(ca, ret, "in replicas_gc2()");
1519 		goto err;
1520 	}
1521 
1522 	data = bch2_dev_has_data(c, ca);
1523 	if (data) {
1524 		struct printbuf data_has = PRINTBUF;
1525 
1526 		prt_bitflags(&data_has, bch2_data_types, data);
1527 		bch_err(ca, "Remove failed, still has data (%s)", data_has.buf);
1528 		printbuf_exit(&data_has);
1529 		ret = -EBUSY;
1530 		goto err;
1531 	}
1532 
1533 	__bch2_dev_offline(c, ca);
1534 
1535 	mutex_lock(&c->sb_lock);
1536 	rcu_assign_pointer(c->devs[ca->dev_idx], NULL);
1537 	mutex_unlock(&c->sb_lock);
1538 
1539 	percpu_ref_kill(&ca->ref);
1540 	wait_for_completion(&ca->ref_completion);
1541 
1542 	bch2_dev_free(ca);
1543 
1544 	/*
1545 	 * At this point the device object has been removed in-core, but the
1546 	 * on-disk journal might still refer to the device index via sb device
1547 	 * usage entries. Recovery fails if it sees usage information for an
1548 	 * invalid device. Flush journal pins to push the back of the journal
1549 	 * past now invalid device index references before we update the
1550 	 * superblock, but after the device object has been removed so any
1551 	 * further journal writes elide usage info for the device.
1552 	 */
1553 	bch2_journal_flush_all_pins(&c->journal);
1554 
1555 	/*
1556 	 * Free this device's slot in the bch_member array - all pointers to
1557 	 * this device must be gone:
1558 	 */
1559 	mutex_lock(&c->sb_lock);
1560 	m = bch2_members_v2_get_mut(c->disk_sb.sb, dev_idx);
1561 	memset(&m->uuid, 0, sizeof(m->uuid));
1562 
1563 	bch2_write_super(c);
1564 
1565 	mutex_unlock(&c->sb_lock);
1566 	up_write(&c->state_lock);
1567 
1568 	bch2_dev_usage_journal_reserve(c);
1569 	return 0;
1570 err:
1571 	if (ca->mi.state == BCH_MEMBER_STATE_rw &&
1572 	    !percpu_ref_is_zero(&ca->io_ref))
1573 		__bch2_dev_read_write(c, ca);
1574 	up_write(&c->state_lock);
1575 	return ret;
1576 }
1577 
1578 /* Add new device to running filesystem: */
1579 int bch2_dev_add(struct bch_fs *c, const char *path)
1580 {
1581 	struct bch_opts opts = bch2_opts_empty();
1582 	struct bch_sb_handle sb;
1583 	struct bch_dev *ca = NULL;
1584 	struct bch_sb_field_members_v2 *mi;
1585 	struct bch_member dev_mi;
1586 	unsigned dev_idx, nr_devices, u64s;
1587 	struct printbuf errbuf = PRINTBUF;
1588 	struct printbuf label = PRINTBUF;
1589 	int ret;
1590 
1591 	ret = bch2_read_super(path, &opts, &sb);
1592 	if (ret) {
1593 		bch_err_msg(c, ret, "reading super");
1594 		goto err;
1595 	}
1596 
1597 	dev_mi = bch2_sb_member_get(sb.sb, sb.sb->dev_idx);
1598 
1599 	if (BCH_MEMBER_GROUP(&dev_mi)) {
1600 		bch2_disk_path_to_text_sb(&label, sb.sb, BCH_MEMBER_GROUP(&dev_mi) - 1);
1601 		if (label.allocation_failure) {
1602 			ret = -ENOMEM;
1603 			goto err;
1604 		}
1605 	}
1606 
1607 	ret = bch2_dev_may_add(sb.sb, c);
1608 	if (ret) {
1609 		bch_err_fn(c, ret);
1610 		goto err;
1611 	}
1612 
1613 	ca = __bch2_dev_alloc(c, &dev_mi);
1614 	if (!ca) {
1615 		ret = -ENOMEM;
1616 		goto err;
1617 	}
1618 
1619 	bch2_dev_usage_init(ca);
1620 
1621 	ret = __bch2_dev_attach_bdev(ca, &sb);
1622 	if (ret)
1623 		goto err;
1624 
1625 	ret = bch2_dev_journal_alloc(ca);
1626 	if (ret) {
1627 		bch_err_msg(c, ret, "allocating journal");
1628 		goto err;
1629 	}
1630 
1631 	down_write(&c->state_lock);
1632 	mutex_lock(&c->sb_lock);
1633 
1634 	ret = bch2_sb_from_fs(c, ca);
1635 	if (ret) {
1636 		bch_err_msg(c, ret, "setting up new superblock");
1637 		goto err_unlock;
1638 	}
1639 
1640 	if (dynamic_fault("bcachefs:add:no_slot"))
1641 		goto no_slot;
1642 
1643 	for (dev_idx = 0; dev_idx < BCH_SB_MEMBERS_MAX; dev_idx++)
1644 		if (!bch2_dev_exists(c->disk_sb.sb, dev_idx))
1645 			goto have_slot;
1646 no_slot:
1647 	ret = -BCH_ERR_ENOSPC_sb_members;
1648 	bch_err_msg(c, ret, "setting up new superblock");
1649 	goto err_unlock;
1650 
1651 have_slot:
1652 	nr_devices = max_t(unsigned, dev_idx + 1, c->sb.nr_devices);
1653 
1654 	mi = bch2_sb_field_get(c->disk_sb.sb, members_v2);
1655 	u64s = DIV_ROUND_UP(sizeof(struct bch_sb_field_members_v2) +
1656 			    le16_to_cpu(mi->member_bytes) * nr_devices, sizeof(u64));
1657 
1658 	mi = bch2_sb_field_resize(&c->disk_sb, members_v2, u64s);
1659 	if (!mi) {
1660 		ret = -BCH_ERR_ENOSPC_sb_members;
1661 		bch_err_msg(c, ret, "setting up new superblock");
1662 		goto err_unlock;
1663 	}
1664 	struct bch_member *m = bch2_members_v2_get_mut(c->disk_sb.sb, dev_idx);
1665 
1666 	/* success: */
1667 
1668 	*m = dev_mi;
1669 	m->last_mount = cpu_to_le64(ktime_get_real_seconds());
1670 	c->disk_sb.sb->nr_devices	= nr_devices;
1671 
1672 	ca->disk_sb.sb->dev_idx	= dev_idx;
1673 	bch2_dev_attach(c, ca, dev_idx);
1674 
1675 	if (BCH_MEMBER_GROUP(&dev_mi)) {
1676 		ret = __bch2_dev_group_set(c, ca, label.buf);
1677 		if (ret) {
1678 			bch_err_msg(c, ret, "creating new label");
1679 			goto err_unlock;
1680 		}
1681 	}
1682 
1683 	bch2_write_super(c);
1684 	mutex_unlock(&c->sb_lock);
1685 
1686 	bch2_dev_usage_journal_reserve(c);
1687 
1688 	ret = bch2_trans_mark_dev_sb(c, ca);
1689 	if (ret) {
1690 		bch_err_msg(ca, ret, "marking new superblock");
1691 		goto err_late;
1692 	}
1693 
1694 	ret = bch2_fs_freespace_init(c);
1695 	if (ret) {
1696 		bch_err_msg(ca, ret, "initializing free space");
1697 		goto err_late;
1698 	}
1699 
1700 	ca->new_fs_bucket_idx = 0;
1701 
1702 	if (ca->mi.state == BCH_MEMBER_STATE_rw)
1703 		__bch2_dev_read_write(c, ca);
1704 
1705 	up_write(&c->state_lock);
1706 	return 0;
1707 
1708 err_unlock:
1709 	mutex_unlock(&c->sb_lock);
1710 	up_write(&c->state_lock);
1711 err:
1712 	if (ca)
1713 		bch2_dev_free(ca);
1714 	bch2_free_super(&sb);
1715 	printbuf_exit(&label);
1716 	printbuf_exit(&errbuf);
1717 	return ret;
1718 err_late:
1719 	up_write(&c->state_lock);
1720 	ca = NULL;
1721 	goto err;
1722 }
1723 
1724 /* Hot add existing device to running filesystem: */
1725 int bch2_dev_online(struct bch_fs *c, const char *path)
1726 {
1727 	struct bch_opts opts = bch2_opts_empty();
1728 	struct bch_sb_handle sb = { NULL };
1729 	struct bch_dev *ca;
1730 	unsigned dev_idx;
1731 	int ret;
1732 
1733 	down_write(&c->state_lock);
1734 
1735 	ret = bch2_read_super(path, &opts, &sb);
1736 	if (ret) {
1737 		up_write(&c->state_lock);
1738 		return ret;
1739 	}
1740 
1741 	dev_idx = sb.sb->dev_idx;
1742 
1743 	ret = bch2_dev_in_fs(c->disk_sb.sb, sb.sb);
1744 	if (ret) {
1745 		bch_err_msg(c, ret, "bringing %s online", path);
1746 		goto err;
1747 	}
1748 
1749 	ret = bch2_dev_attach_bdev(c, &sb);
1750 	if (ret)
1751 		goto err;
1752 
1753 	ca = bch_dev_locked(c, dev_idx);
1754 
1755 	ret = bch2_trans_mark_dev_sb(c, ca);
1756 	if (ret) {
1757 		bch_err_msg(c, ret, "bringing %s online: error from bch2_trans_mark_dev_sb", path);
1758 		goto err;
1759 	}
1760 
1761 	if (ca->mi.state == BCH_MEMBER_STATE_rw)
1762 		__bch2_dev_read_write(c, ca);
1763 
1764 	if (!ca->mi.freespace_initialized) {
1765 		ret = bch2_dev_freespace_init(c, ca, 0, ca->mi.nbuckets);
1766 		bch_err_msg(ca, ret, "initializing free space");
1767 		if (ret)
1768 			goto err;
1769 	}
1770 
1771 	if (!ca->journal.nr) {
1772 		ret = bch2_dev_journal_alloc(ca);
1773 		bch_err_msg(ca, ret, "allocating journal");
1774 		if (ret)
1775 			goto err;
1776 	}
1777 
1778 	mutex_lock(&c->sb_lock);
1779 	bch2_members_v2_get_mut(c->disk_sb.sb, ca->dev_idx)->last_mount =
1780 		cpu_to_le64(ktime_get_real_seconds());
1781 	bch2_write_super(c);
1782 	mutex_unlock(&c->sb_lock);
1783 
1784 	up_write(&c->state_lock);
1785 	return 0;
1786 err:
1787 	up_write(&c->state_lock);
1788 	bch2_free_super(&sb);
1789 	return ret;
1790 }
1791 
1792 int bch2_dev_offline(struct bch_fs *c, struct bch_dev *ca, int flags)
1793 {
1794 	down_write(&c->state_lock);
1795 
1796 	if (!bch2_dev_is_online(ca)) {
1797 		bch_err(ca, "Already offline");
1798 		up_write(&c->state_lock);
1799 		return 0;
1800 	}
1801 
1802 	if (!bch2_dev_state_allowed(c, ca, BCH_MEMBER_STATE_failed, flags)) {
1803 		bch_err(ca, "Cannot offline required disk");
1804 		up_write(&c->state_lock);
1805 		return -BCH_ERR_device_state_not_allowed;
1806 	}
1807 
1808 	__bch2_dev_offline(c, ca);
1809 
1810 	up_write(&c->state_lock);
1811 	return 0;
1812 }
1813 
1814 int bch2_dev_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
1815 {
1816 	struct bch_member *m;
1817 	u64 old_nbuckets;
1818 	int ret = 0;
1819 
1820 	down_write(&c->state_lock);
1821 	old_nbuckets = ca->mi.nbuckets;
1822 
1823 	if (nbuckets < ca->mi.nbuckets) {
1824 		bch_err(ca, "Cannot shrink yet");
1825 		ret = -EINVAL;
1826 		goto err;
1827 	}
1828 
1829 	if (bch2_dev_is_online(ca) &&
1830 	    get_capacity(ca->disk_sb.bdev->bd_disk) <
1831 	    ca->mi.bucket_size * nbuckets) {
1832 		bch_err(ca, "New size larger than device");
1833 		ret = -BCH_ERR_device_size_too_small;
1834 		goto err;
1835 	}
1836 
1837 	ret = bch2_dev_buckets_resize(c, ca, nbuckets);
1838 	if (ret) {
1839 		bch_err_msg(ca, ret, "resizing buckets");
1840 		goto err;
1841 	}
1842 
1843 	ret = bch2_trans_mark_dev_sb(c, ca);
1844 	if (ret)
1845 		goto err;
1846 
1847 	mutex_lock(&c->sb_lock);
1848 	m = bch2_members_v2_get_mut(c->disk_sb.sb, ca->dev_idx);
1849 	m->nbuckets = cpu_to_le64(nbuckets);
1850 
1851 	bch2_write_super(c);
1852 	mutex_unlock(&c->sb_lock);
1853 
1854 	if (ca->mi.freespace_initialized) {
1855 		ret = bch2_dev_freespace_init(c, ca, old_nbuckets, nbuckets);
1856 		if (ret)
1857 			goto err;
1858 
1859 		/*
1860 		 * XXX: this is all wrong transactionally - we'll be able to do
1861 		 * this correctly after the disk space accounting rewrite
1862 		 */
1863 		ca->usage_base->d[BCH_DATA_free].buckets += nbuckets - old_nbuckets;
1864 	}
1865 
1866 	bch2_recalc_capacity(c);
1867 err:
1868 	up_write(&c->state_lock);
1869 	return ret;
1870 }
1871 
1872 /* return with ref on ca->ref: */
1873 struct bch_dev *bch2_dev_lookup(struct bch_fs *c, const char *name)
1874 {
1875 	struct bch_dev *ca;
1876 	unsigned i;
1877 
1878 	rcu_read_lock();
1879 	for_each_member_device_rcu(ca, c, i, NULL)
1880 		if (!strcmp(name, ca->name))
1881 			goto found;
1882 	ca = ERR_PTR(-BCH_ERR_ENOENT_dev_not_found);
1883 found:
1884 	rcu_read_unlock();
1885 
1886 	return ca;
1887 }
1888 
1889 /* Filesystem open: */
1890 
1891 struct bch_fs *bch2_fs_open(char * const *devices, unsigned nr_devices,
1892 			    struct bch_opts opts)
1893 {
1894 	DARRAY(struct bch_sb_handle) sbs = { 0 };
1895 	struct bch_fs *c = NULL;
1896 	struct bch_sb_handle *sb, *best = NULL;
1897 	struct printbuf errbuf = PRINTBUF;
1898 	int ret = 0;
1899 
1900 	if (!try_module_get(THIS_MODULE))
1901 		return ERR_PTR(-ENODEV);
1902 
1903 	if (!nr_devices) {
1904 		ret = -EINVAL;
1905 		goto err;
1906 	}
1907 
1908 	ret = darray_make_room(&sbs, nr_devices);
1909 	if (ret)
1910 		goto err;
1911 
1912 	for (unsigned i = 0; i < nr_devices; i++) {
1913 		struct bch_sb_handle sb = { NULL };
1914 
1915 		ret = bch2_read_super(devices[i], &opts, &sb);
1916 		if (ret)
1917 			goto err;
1918 
1919 		BUG_ON(darray_push(&sbs, sb));
1920 	}
1921 
1922 	darray_for_each(sbs, sb)
1923 		if (!best || le64_to_cpu(sb->sb->seq) > le64_to_cpu(best->sb->seq))
1924 			best = sb;
1925 
1926 	darray_for_each_reverse(sbs, sb) {
1927 		if (sb != best && !bch2_dev_exists(best->sb, sb->sb->dev_idx)) {
1928 			pr_info("%pg has been removed, skipping", sb->bdev);
1929 			bch2_free_super(sb);
1930 			darray_remove_item(&sbs, sb);
1931 			best -= best > sb;
1932 			continue;
1933 		}
1934 
1935 		ret = bch2_dev_in_fs(best->sb, sb->sb);
1936 		if (ret)
1937 			goto err_print;
1938 	}
1939 
1940 	c = bch2_fs_alloc(best->sb, opts);
1941 	ret = PTR_ERR_OR_ZERO(c);
1942 	if (ret)
1943 		goto err;
1944 
1945 	down_write(&c->state_lock);
1946 	darray_for_each(sbs, sb) {
1947 		ret = bch2_dev_attach_bdev(c, sb);
1948 		if (ret) {
1949 			up_write(&c->state_lock);
1950 			goto err;
1951 		}
1952 	}
1953 	up_write(&c->state_lock);
1954 
1955 	if (!bch2_fs_may_start(c)) {
1956 		ret = -BCH_ERR_insufficient_devices_to_start;
1957 		goto err_print;
1958 	}
1959 
1960 	if (!c->opts.nostart) {
1961 		ret = bch2_fs_start(c);
1962 		if (ret)
1963 			goto err;
1964 	}
1965 out:
1966 	darray_for_each(sbs, sb)
1967 		bch2_free_super(sb);
1968 	darray_exit(&sbs);
1969 	printbuf_exit(&errbuf);
1970 	module_put(THIS_MODULE);
1971 	return c;
1972 err_print:
1973 	pr_err("bch_fs_open err opening %s: %s",
1974 	       devices[0], bch2_err_str(ret));
1975 err:
1976 	if (!IS_ERR_OR_NULL(c))
1977 		bch2_fs_stop(c);
1978 	c = ERR_PTR(ret);
1979 	goto out;
1980 }
1981 
1982 /* Global interfaces/init */
1983 
1984 static void bcachefs_exit(void)
1985 {
1986 	bch2_debug_exit();
1987 	bch2_vfs_exit();
1988 	bch2_chardev_exit();
1989 	bch2_btree_key_cache_exit();
1990 	if (bcachefs_kset)
1991 		kset_unregister(bcachefs_kset);
1992 }
1993 
1994 static int __init bcachefs_init(void)
1995 {
1996 	bch2_bkey_pack_test();
1997 
1998 	if (!(bcachefs_kset = kset_create_and_add("bcachefs", NULL, fs_kobj)) ||
1999 	    bch2_btree_key_cache_init() ||
2000 	    bch2_chardev_init() ||
2001 	    bch2_vfs_init() ||
2002 	    bch2_debug_init())
2003 		goto err;
2004 
2005 	return 0;
2006 err:
2007 	bcachefs_exit();
2008 	return -ENOMEM;
2009 }
2010 
2011 #define BCH_DEBUG_PARAM(name, description)			\
2012 	bool bch2_##name;					\
2013 	module_param_named(name, bch2_##name, bool, 0644);	\
2014 	MODULE_PARM_DESC(name, description);
2015 BCH_DEBUG_PARAMS()
2016 #undef BCH_DEBUG_PARAM
2017 
2018 __maybe_unused
2019 static unsigned bch2_metadata_version = bcachefs_metadata_version_current;
2020 module_param_named(version, bch2_metadata_version, uint, 0400);
2021 
2022 module_exit(bcachefs_exit);
2023 module_init(bcachefs_init);
2024