xref: /linux/fs/bcachefs/super.c (revision 0678df8271820bcf8fb4f877129f05d68a237de4)
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 	ret = bch2_gc_thread_start(c);
427 	if (ret) {
428 		bch_err(c, "error starting gc thread");
429 		return ret;
430 	}
431 
432 	ret = bch2_journal_reclaim_start(&c->journal);
433 	if (ret)
434 		goto err;
435 
436 	if (!early) {
437 		ret = bch2_fs_read_write_late(c);
438 		if (ret)
439 			goto err;
440 	}
441 
442 #ifndef BCH_WRITE_REF_DEBUG
443 	percpu_ref_reinit(&c->writes);
444 #else
445 	for (i = 0; i < BCH_WRITE_REF_NR; i++) {
446 		BUG_ON(atomic_long_read(&c->writes[i]));
447 		atomic_long_inc(&c->writes[i]);
448 	}
449 #endif
450 	set_bit(BCH_FS_RW, &c->flags);
451 	set_bit(BCH_FS_WAS_RW, &c->flags);
452 
453 	bch2_do_discards(c);
454 	bch2_do_invalidates(c);
455 	bch2_do_stripe_deletes(c);
456 	bch2_do_pending_node_rewrites(c);
457 	return 0;
458 err:
459 	__bch2_fs_read_only(c);
460 	return ret;
461 }
462 
463 int bch2_fs_read_write(struct bch_fs *c)
464 {
465 	return __bch2_fs_read_write(c, false);
466 }
467 
468 int bch2_fs_read_write_early(struct bch_fs *c)
469 {
470 	lockdep_assert_held(&c->state_lock);
471 
472 	return __bch2_fs_read_write(c, true);
473 }
474 
475 /* Filesystem startup/shutdown: */
476 
477 static void __bch2_fs_free(struct bch_fs *c)
478 {
479 	unsigned i;
480 
481 	for (i = 0; i < BCH_TIME_STAT_NR; i++)
482 		bch2_time_stats_exit(&c->times[i]);
483 
484 	bch2_free_pending_node_rewrites(c);
485 	bch2_fs_sb_errors_exit(c);
486 	bch2_fs_counters_exit(c);
487 	bch2_fs_snapshots_exit(c);
488 	bch2_fs_quota_exit(c);
489 	bch2_fs_fs_io_direct_exit(c);
490 	bch2_fs_fs_io_buffered_exit(c);
491 	bch2_fs_fsio_exit(c);
492 	bch2_fs_ec_exit(c);
493 	bch2_fs_encryption_exit(c);
494 	bch2_fs_nocow_locking_exit(c);
495 	bch2_fs_io_write_exit(c);
496 	bch2_fs_io_read_exit(c);
497 	bch2_fs_buckets_waiting_for_journal_exit(c);
498 	bch2_fs_btree_interior_update_exit(c);
499 	bch2_fs_btree_iter_exit(c);
500 	bch2_fs_btree_key_cache_exit(&c->btree_key_cache);
501 	bch2_fs_btree_cache_exit(c);
502 	bch2_fs_replicas_exit(c);
503 	bch2_fs_journal_exit(&c->journal);
504 	bch2_io_clock_exit(&c->io_clock[WRITE]);
505 	bch2_io_clock_exit(&c->io_clock[READ]);
506 	bch2_fs_compress_exit(c);
507 	bch2_journal_keys_free(&c->journal_keys);
508 	bch2_journal_entries_free(c);
509 	bch2_fs_btree_write_buffer_exit(c);
510 	percpu_free_rwsem(&c->mark_lock);
511 	free_percpu(c->online_reserved);
512 
513 	darray_exit(&c->btree_roots_extra);
514 	free_percpu(c->pcpu);
515 	mempool_exit(&c->large_bkey_pool);
516 	mempool_exit(&c->btree_bounce_pool);
517 	bioset_exit(&c->btree_bio);
518 	mempool_exit(&c->fill_iter);
519 #ifndef BCH_WRITE_REF_DEBUG
520 	percpu_ref_exit(&c->writes);
521 #endif
522 	kfree(rcu_dereference_protected(c->disk_groups, 1));
523 	kfree(c->journal_seq_blacklist_table);
524 	kfree(c->unused_inode_hints);
525 
526 	if (c->write_ref_wq)
527 		destroy_workqueue(c->write_ref_wq);
528 	if (c->io_complete_wq)
529 		destroy_workqueue(c->io_complete_wq);
530 	if (c->copygc_wq)
531 		destroy_workqueue(c->copygc_wq);
532 	if (c->btree_io_complete_wq)
533 		destroy_workqueue(c->btree_io_complete_wq);
534 	if (c->btree_update_wq)
535 		destroy_workqueue(c->btree_update_wq);
536 
537 	bch2_free_super(&c->disk_sb);
538 	kvpfree(c, sizeof(*c));
539 	module_put(THIS_MODULE);
540 }
541 
542 static void bch2_fs_release(struct kobject *kobj)
543 {
544 	struct bch_fs *c = container_of(kobj, struct bch_fs, kobj);
545 
546 	__bch2_fs_free(c);
547 }
548 
549 void __bch2_fs_stop(struct bch_fs *c)
550 {
551 	struct bch_dev *ca;
552 	unsigned i;
553 
554 	bch_verbose(c, "shutting down");
555 
556 	set_bit(BCH_FS_STOPPING, &c->flags);
557 
558 	cancel_work_sync(&c->journal_seq_blacklist_gc_work);
559 
560 	down_write(&c->state_lock);
561 	bch2_fs_read_only(c);
562 	up_write(&c->state_lock);
563 
564 	for_each_member_device(ca, c, i)
565 		if (ca->kobj.state_in_sysfs &&
566 		    ca->disk_sb.bdev)
567 			sysfs_remove_link(bdev_kobj(ca->disk_sb.bdev), "bcachefs");
568 
569 	if (c->kobj.state_in_sysfs)
570 		kobject_del(&c->kobj);
571 
572 	bch2_fs_debug_exit(c);
573 	bch2_fs_chardev_exit(c);
574 
575 	kobject_put(&c->counters_kobj);
576 	kobject_put(&c->time_stats);
577 	kobject_put(&c->opts_dir);
578 	kobject_put(&c->internal);
579 
580 	/* btree prefetch might have kicked off reads in the background: */
581 	bch2_btree_flush_all_reads(c);
582 
583 	for_each_member_device(ca, c, i)
584 		cancel_work_sync(&ca->io_error_work);
585 
586 	cancel_work_sync(&c->read_only_work);
587 }
588 
589 void bch2_fs_free(struct bch_fs *c)
590 {
591 	unsigned i;
592 
593 	mutex_lock(&bch_fs_list_lock);
594 	list_del(&c->list);
595 	mutex_unlock(&bch_fs_list_lock);
596 
597 	closure_sync(&c->cl);
598 	closure_debug_destroy(&c->cl);
599 
600 	for (i = 0; i < c->sb.nr_devices; i++) {
601 		struct bch_dev *ca = rcu_dereference_protected(c->devs[i], true);
602 
603 		if (ca) {
604 			bch2_free_super(&ca->disk_sb);
605 			bch2_dev_free(ca);
606 		}
607 	}
608 
609 	bch_verbose(c, "shutdown complete");
610 
611 	kobject_put(&c->kobj);
612 }
613 
614 void bch2_fs_stop(struct bch_fs *c)
615 {
616 	__bch2_fs_stop(c);
617 	bch2_fs_free(c);
618 }
619 
620 static int bch2_fs_online(struct bch_fs *c)
621 {
622 	struct bch_dev *ca;
623 	unsigned i;
624 	int ret = 0;
625 
626 	lockdep_assert_held(&bch_fs_list_lock);
627 
628 	if (__bch2_uuid_to_fs(c->sb.uuid)) {
629 		bch_err(c, "filesystem UUID already open");
630 		return -EINVAL;
631 	}
632 
633 	ret = bch2_fs_chardev_init(c);
634 	if (ret) {
635 		bch_err(c, "error creating character device");
636 		return ret;
637 	}
638 
639 	bch2_fs_debug_init(c);
640 
641 	ret = kobject_add(&c->kobj, NULL, "%pU", c->sb.user_uuid.b) ?:
642 	    kobject_add(&c->internal, &c->kobj, "internal") ?:
643 	    kobject_add(&c->opts_dir, &c->kobj, "options") ?:
644 	    kobject_add(&c->time_stats, &c->kobj, "time_stats") ?:
645 	    kobject_add(&c->counters_kobj, &c->kobj, "counters") ?:
646 	    bch2_opts_create_sysfs_files(&c->opts_dir);
647 	if (ret) {
648 		bch_err(c, "error creating sysfs objects");
649 		return ret;
650 	}
651 
652 	down_write(&c->state_lock);
653 
654 	for_each_member_device(ca, c, i) {
655 		ret = bch2_dev_sysfs_online(c, ca);
656 		if (ret) {
657 			bch_err(c, "error creating sysfs objects");
658 			percpu_ref_put(&ca->ref);
659 			goto err;
660 		}
661 	}
662 
663 	BUG_ON(!list_empty(&c->list));
664 	list_add(&c->list, &bch_fs_list);
665 err:
666 	up_write(&c->state_lock);
667 	return ret;
668 }
669 
670 static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
671 {
672 	struct bch_fs *c;
673 	struct printbuf name = PRINTBUF;
674 	unsigned i, iter_size;
675 	int ret = 0;
676 
677 	c = kvpmalloc(sizeof(struct bch_fs), GFP_KERNEL|__GFP_ZERO);
678 	if (!c) {
679 		c = ERR_PTR(-BCH_ERR_ENOMEM_fs_alloc);
680 		goto out;
681 	}
682 
683 	__module_get(THIS_MODULE);
684 
685 	closure_init(&c->cl, NULL);
686 
687 	c->kobj.kset = bcachefs_kset;
688 	kobject_init(&c->kobj, &bch2_fs_ktype);
689 	kobject_init(&c->internal, &bch2_fs_internal_ktype);
690 	kobject_init(&c->opts_dir, &bch2_fs_opts_dir_ktype);
691 	kobject_init(&c->time_stats, &bch2_fs_time_stats_ktype);
692 	kobject_init(&c->counters_kobj, &bch2_fs_counters_ktype);
693 
694 	c->minor		= -1;
695 	c->disk_sb.fs_sb	= true;
696 
697 	init_rwsem(&c->state_lock);
698 	mutex_init(&c->sb_lock);
699 	mutex_init(&c->replicas_gc_lock);
700 	mutex_init(&c->btree_root_lock);
701 	INIT_WORK(&c->read_only_work, bch2_fs_read_only_work);
702 
703 	init_rwsem(&c->gc_lock);
704 	mutex_init(&c->gc_gens_lock);
705 
706 	for (i = 0; i < BCH_TIME_STAT_NR; i++)
707 		bch2_time_stats_init(&c->times[i]);
708 
709 	bch2_fs_copygc_init(c);
710 	bch2_fs_btree_key_cache_init_early(&c->btree_key_cache);
711 	bch2_fs_btree_interior_update_init_early(c);
712 	bch2_fs_allocator_background_init(c);
713 	bch2_fs_allocator_foreground_init(c);
714 	bch2_fs_rebalance_init(c);
715 	bch2_fs_quota_init(c);
716 	bch2_fs_ec_init_early(c);
717 	bch2_fs_move_init(c);
718 	bch2_fs_sb_errors_init_early(c);
719 
720 	INIT_LIST_HEAD(&c->list);
721 
722 	mutex_init(&c->usage_scratch_lock);
723 
724 	mutex_init(&c->bio_bounce_pages_lock);
725 	mutex_init(&c->snapshot_table_lock);
726 	init_rwsem(&c->snapshot_create_lock);
727 
728 	spin_lock_init(&c->btree_write_error_lock);
729 
730 	INIT_WORK(&c->journal_seq_blacklist_gc_work,
731 		  bch2_blacklist_entries_gc);
732 
733 	INIT_LIST_HEAD(&c->journal_iters);
734 
735 	INIT_LIST_HEAD(&c->fsck_error_msgs);
736 	mutex_init(&c->fsck_error_msgs_lock);
737 
738 	seqcount_init(&c->gc_pos_lock);
739 
740 	seqcount_init(&c->usage_lock);
741 
742 	sema_init(&c->io_in_flight, 128);
743 
744 	INIT_LIST_HEAD(&c->vfs_inodes_list);
745 	mutex_init(&c->vfs_inodes_lock);
746 
747 	c->copy_gc_enabled		= 1;
748 	c->rebalance.enabled		= 1;
749 	c->promote_whole_extents	= true;
750 
751 	c->journal.flush_write_time	= &c->times[BCH_TIME_journal_flush_write];
752 	c->journal.noflush_write_time	= &c->times[BCH_TIME_journal_noflush_write];
753 	c->journal.blocked_time		= &c->times[BCH_TIME_blocked_journal];
754 	c->journal.flush_seq_time	= &c->times[BCH_TIME_journal_flush_seq];
755 
756 	bch2_fs_btree_cache_init_early(&c->btree_cache);
757 
758 	mutex_init(&c->sectors_available_lock);
759 
760 	ret = percpu_init_rwsem(&c->mark_lock);
761 	if (ret)
762 		goto err;
763 
764 	mutex_lock(&c->sb_lock);
765 	ret = bch2_sb_to_fs(c, sb);
766 	mutex_unlock(&c->sb_lock);
767 
768 	if (ret)
769 		goto err;
770 
771 	pr_uuid(&name, c->sb.user_uuid.b);
772 	strscpy(c->name, name.buf, sizeof(c->name));
773 	printbuf_exit(&name);
774 
775 	ret = name.allocation_failure ? -BCH_ERR_ENOMEM_fs_name_alloc : 0;
776 	if (ret)
777 		goto err;
778 
779 	/* Compat: */
780 	if (le16_to_cpu(sb->version) <= bcachefs_metadata_version_inode_v2 &&
781 	    !BCH_SB_JOURNAL_FLUSH_DELAY(sb))
782 		SET_BCH_SB_JOURNAL_FLUSH_DELAY(sb, 1000);
783 
784 	if (le16_to_cpu(sb->version) <= bcachefs_metadata_version_inode_v2 &&
785 	    !BCH_SB_JOURNAL_RECLAIM_DELAY(sb))
786 		SET_BCH_SB_JOURNAL_RECLAIM_DELAY(sb, 100);
787 
788 	c->opts = bch2_opts_default;
789 	ret = bch2_opts_from_sb(&c->opts, sb);
790 	if (ret)
791 		goto err;
792 
793 	bch2_opts_apply(&c->opts, opts);
794 
795 	c->btree_key_cache_btrees |= 1U << BTREE_ID_alloc;
796 	if (c->opts.inodes_use_key_cache)
797 		c->btree_key_cache_btrees |= 1U << BTREE_ID_inodes;
798 	c->btree_key_cache_btrees |= 1U << BTREE_ID_logged_ops;
799 
800 	c->block_bits		= ilog2(block_sectors(c));
801 	c->btree_foreground_merge_threshold = BTREE_FOREGROUND_MERGE_THRESHOLD(c);
802 
803 	if (bch2_fs_init_fault("fs_alloc")) {
804 		bch_err(c, "fs_alloc fault injected");
805 		ret = -EFAULT;
806 		goto err;
807 	}
808 
809 	iter_size = sizeof(struct sort_iter) +
810 		(btree_blocks(c) + 1) * 2 *
811 		sizeof(struct sort_iter_set);
812 
813 	c->inode_shard_bits = ilog2(roundup_pow_of_two(num_possible_cpus()));
814 
815 	if (!(c->btree_update_wq = alloc_workqueue("bcachefs",
816 				WQ_FREEZABLE|WQ_UNBOUND|WQ_MEM_RECLAIM, 512)) ||
817 	    !(c->btree_io_complete_wq = alloc_workqueue("bcachefs_btree_io",
818 				WQ_FREEZABLE|WQ_MEM_RECLAIM, 1)) ||
819 	    !(c->copygc_wq = alloc_workqueue("bcachefs_copygc",
820 				WQ_FREEZABLE|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE, 1)) ||
821 	    !(c->io_complete_wq = alloc_workqueue("bcachefs_io",
822 				WQ_FREEZABLE|WQ_HIGHPRI|WQ_MEM_RECLAIM, 1)) ||
823 	    !(c->write_ref_wq = alloc_workqueue("bcachefs_write_ref",
824 				WQ_FREEZABLE, 0)) ||
825 #ifndef BCH_WRITE_REF_DEBUG
826 	    percpu_ref_init(&c->writes, bch2_writes_disabled,
827 			    PERCPU_REF_INIT_DEAD, GFP_KERNEL) ||
828 #endif
829 	    mempool_init_kmalloc_pool(&c->fill_iter, 1, iter_size) ||
830 	    bioset_init(&c->btree_bio, 1,
831 			max(offsetof(struct btree_read_bio, bio),
832 			    offsetof(struct btree_write_bio, wbio.bio)),
833 			BIOSET_NEED_BVECS) ||
834 	    !(c->pcpu = alloc_percpu(struct bch_fs_pcpu)) ||
835 	    !(c->online_reserved = alloc_percpu(u64)) ||
836 	    mempool_init_kvpmalloc_pool(&c->btree_bounce_pool, 1,
837 					btree_bytes(c)) ||
838 	    mempool_init_kmalloc_pool(&c->large_bkey_pool, 1, 2048) ||
839 	    !(c->unused_inode_hints = kcalloc(1U << c->inode_shard_bits,
840 					      sizeof(u64), GFP_KERNEL))) {
841 		ret = -BCH_ERR_ENOMEM_fs_other_alloc;
842 		goto err;
843 	}
844 
845 	ret = bch2_fs_counters_init(c) ?:
846 	    bch2_fs_sb_errors_init(c) ?:
847 	    bch2_io_clock_init(&c->io_clock[READ]) ?:
848 	    bch2_io_clock_init(&c->io_clock[WRITE]) ?:
849 	    bch2_fs_journal_init(&c->journal) ?:
850 	    bch2_fs_replicas_init(c) ?:
851 	    bch2_fs_btree_cache_init(c) ?:
852 	    bch2_fs_btree_key_cache_init(&c->btree_key_cache) ?:
853 	    bch2_fs_btree_iter_init(c) ?:
854 	    bch2_fs_btree_interior_update_init(c) ?:
855 	    bch2_fs_buckets_waiting_for_journal_init(c) ?:
856 	    bch2_fs_btree_write_buffer_init(c) ?:
857 	    bch2_fs_subvolumes_init(c) ?:
858 	    bch2_fs_io_read_init(c) ?:
859 	    bch2_fs_io_write_init(c) ?:
860 	    bch2_fs_nocow_locking_init(c) ?:
861 	    bch2_fs_encryption_init(c) ?:
862 	    bch2_fs_compress_init(c) ?:
863 	    bch2_fs_ec_init(c) ?:
864 	    bch2_fs_fsio_init(c) ?:
865 	    bch2_fs_fs_io_buffered_init(c) ?:
866 	    bch2_fs_fs_io_direct_init(c);
867 	if (ret)
868 		goto err;
869 
870 	for (i = 0; i < c->sb.nr_devices; i++)
871 		if (bch2_dev_exists(c->disk_sb.sb, i) &&
872 		    bch2_dev_alloc(c, i)) {
873 			ret = -EEXIST;
874 			goto err;
875 		}
876 
877 	bch2_journal_entry_res_resize(&c->journal,
878 			&c->btree_root_journal_res,
879 			BTREE_ID_NR * (JSET_KEYS_U64s + BKEY_BTREE_PTR_U64s_MAX));
880 	bch2_dev_usage_journal_reserve(c);
881 	bch2_journal_entry_res_resize(&c->journal,
882 			&c->clock_journal_res,
883 			(sizeof(struct jset_entry_clock) / sizeof(u64)) * 2);
884 
885 	mutex_lock(&bch_fs_list_lock);
886 	ret = bch2_fs_online(c);
887 	mutex_unlock(&bch_fs_list_lock);
888 
889 	if (ret)
890 		goto err;
891 out:
892 	return c;
893 err:
894 	bch2_fs_free(c);
895 	c = ERR_PTR(ret);
896 	goto out;
897 }
898 
899 noinline_for_stack
900 static void print_mount_opts(struct bch_fs *c)
901 {
902 	enum bch_opt_id i;
903 	struct printbuf p = PRINTBUF;
904 	bool first = true;
905 
906 	prt_str(&p, "mounting version ");
907 	bch2_version_to_text(&p, c->sb.version);
908 
909 	if (c->opts.read_only) {
910 		prt_str(&p, " opts=");
911 		first = false;
912 		prt_printf(&p, "ro");
913 	}
914 
915 	for (i = 0; i < bch2_opts_nr; i++) {
916 		const struct bch_option *opt = &bch2_opt_table[i];
917 		u64 v = bch2_opt_get_by_id(&c->opts, i);
918 
919 		if (!(opt->flags & OPT_MOUNT))
920 			continue;
921 
922 		if (v == bch2_opt_get_by_id(&bch2_opts_default, i))
923 			continue;
924 
925 		prt_str(&p, first ? " opts=" : ",");
926 		first = false;
927 		bch2_opt_to_text(&p, c, c->disk_sb.sb, opt, v, OPT_SHOW_MOUNT_STYLE);
928 	}
929 
930 	bch_info(c, "%s", p.buf);
931 	printbuf_exit(&p);
932 }
933 
934 int bch2_fs_start(struct bch_fs *c)
935 {
936 	struct bch_dev *ca;
937 	time64_t now = ktime_get_real_seconds();
938 	unsigned i;
939 	int ret;
940 
941 	print_mount_opts(c);
942 
943 	down_write(&c->state_lock);
944 
945 	BUG_ON(test_bit(BCH_FS_STARTED, &c->flags));
946 
947 	mutex_lock(&c->sb_lock);
948 
949 	ret = bch2_sb_members_v2_init(c);
950 	if (ret) {
951 		mutex_unlock(&c->sb_lock);
952 		goto err;
953 	}
954 
955 	for_each_online_member(ca, c, i)
956 		bch2_members_v2_get_mut(c->disk_sb.sb, i)->last_mount = cpu_to_le64(now);
957 
958 	mutex_unlock(&c->sb_lock);
959 
960 	for_each_rw_member(ca, c, i)
961 		bch2_dev_allocator_add(c, ca);
962 	bch2_recalc_capacity(c);
963 
964 	ret = BCH_SB_INITIALIZED(c->disk_sb.sb)
965 		? bch2_fs_recovery(c)
966 		: bch2_fs_initialize(c);
967 	if (ret)
968 		goto err;
969 
970 	ret = bch2_opts_check_may_set(c);
971 	if (ret)
972 		goto err;
973 
974 	if (bch2_fs_init_fault("fs_start")) {
975 		bch_err(c, "fs_start fault injected");
976 		ret = -EINVAL;
977 		goto err;
978 	}
979 
980 	set_bit(BCH_FS_STARTED, &c->flags);
981 
982 	if (c->opts.read_only || c->opts.nochanges) {
983 		bch2_fs_read_only(c);
984 	} else {
985 		ret = !test_bit(BCH_FS_RW, &c->flags)
986 			? bch2_fs_read_write(c)
987 			: bch2_fs_read_write_late(c);
988 		if (ret)
989 			goto err;
990 	}
991 
992 	ret = 0;
993 out:
994 	up_write(&c->state_lock);
995 	return ret;
996 err:
997 	bch_err_msg(c, ret, "starting filesystem");
998 	goto out;
999 }
1000 
1001 static int bch2_dev_may_add(struct bch_sb *sb, struct bch_fs *c)
1002 {
1003 	struct bch_member m = bch2_sb_member_get(sb, sb->dev_idx);
1004 
1005 	if (le16_to_cpu(sb->block_size) != block_sectors(c))
1006 		return -BCH_ERR_mismatched_block_size;
1007 
1008 	if (le16_to_cpu(m.bucket_size) <
1009 	    BCH_SB_BTREE_NODE_SIZE(c->disk_sb.sb))
1010 		return -BCH_ERR_bucket_size_too_small;
1011 
1012 	return 0;
1013 }
1014 
1015 static int bch2_dev_in_fs(struct bch_sb *fs, struct bch_sb *sb)
1016 {
1017 	struct bch_sb *newest =
1018 		le64_to_cpu(fs->seq) > le64_to_cpu(sb->seq) ? fs : sb;
1019 
1020 	if (!uuid_equal(&fs->uuid, &sb->uuid))
1021 		return -BCH_ERR_device_not_a_member_of_filesystem;
1022 
1023 	if (!bch2_dev_exists(newest, sb->dev_idx))
1024 		return -BCH_ERR_device_has_been_removed;
1025 
1026 	if (fs->block_size != sb->block_size)
1027 		return -BCH_ERR_mismatched_block_size;
1028 
1029 	return 0;
1030 }
1031 
1032 /* Device startup/shutdown: */
1033 
1034 static void bch2_dev_release(struct kobject *kobj)
1035 {
1036 	struct bch_dev *ca = container_of(kobj, struct bch_dev, kobj);
1037 
1038 	kfree(ca);
1039 }
1040 
1041 static void bch2_dev_free(struct bch_dev *ca)
1042 {
1043 	cancel_work_sync(&ca->io_error_work);
1044 
1045 	if (ca->kobj.state_in_sysfs &&
1046 	    ca->disk_sb.bdev)
1047 		sysfs_remove_link(bdev_kobj(ca->disk_sb.bdev), "bcachefs");
1048 
1049 	if (ca->kobj.state_in_sysfs)
1050 		kobject_del(&ca->kobj);
1051 
1052 	bch2_free_super(&ca->disk_sb);
1053 	bch2_dev_journal_exit(ca);
1054 
1055 	free_percpu(ca->io_done);
1056 	bioset_exit(&ca->replica_set);
1057 	bch2_dev_buckets_free(ca);
1058 	free_page((unsigned long) ca->sb_read_scratch);
1059 
1060 	bch2_time_stats_exit(&ca->io_latency[WRITE]);
1061 	bch2_time_stats_exit(&ca->io_latency[READ]);
1062 
1063 	percpu_ref_exit(&ca->io_ref);
1064 	percpu_ref_exit(&ca->ref);
1065 	kobject_put(&ca->kobj);
1066 }
1067 
1068 static void __bch2_dev_offline(struct bch_fs *c, struct bch_dev *ca)
1069 {
1070 
1071 	lockdep_assert_held(&c->state_lock);
1072 
1073 	if (percpu_ref_is_zero(&ca->io_ref))
1074 		return;
1075 
1076 	__bch2_dev_read_only(c, ca);
1077 
1078 	reinit_completion(&ca->io_ref_completion);
1079 	percpu_ref_kill(&ca->io_ref);
1080 	wait_for_completion(&ca->io_ref_completion);
1081 
1082 	if (ca->kobj.state_in_sysfs) {
1083 		sysfs_remove_link(bdev_kobj(ca->disk_sb.bdev), "bcachefs");
1084 		sysfs_remove_link(&ca->kobj, "block");
1085 	}
1086 
1087 	bch2_free_super(&ca->disk_sb);
1088 	bch2_dev_journal_exit(ca);
1089 }
1090 
1091 static void bch2_dev_ref_complete(struct percpu_ref *ref)
1092 {
1093 	struct bch_dev *ca = container_of(ref, struct bch_dev, ref);
1094 
1095 	complete(&ca->ref_completion);
1096 }
1097 
1098 static void bch2_dev_io_ref_complete(struct percpu_ref *ref)
1099 {
1100 	struct bch_dev *ca = container_of(ref, struct bch_dev, io_ref);
1101 
1102 	complete(&ca->io_ref_completion);
1103 }
1104 
1105 static int bch2_dev_sysfs_online(struct bch_fs *c, struct bch_dev *ca)
1106 {
1107 	int ret;
1108 
1109 	if (!c->kobj.state_in_sysfs)
1110 		return 0;
1111 
1112 	if (!ca->kobj.state_in_sysfs) {
1113 		ret = kobject_add(&ca->kobj, &c->kobj,
1114 				  "dev-%u", ca->dev_idx);
1115 		if (ret)
1116 			return ret;
1117 	}
1118 
1119 	if (ca->disk_sb.bdev) {
1120 		struct kobject *block = bdev_kobj(ca->disk_sb.bdev);
1121 
1122 		ret = sysfs_create_link(block, &ca->kobj, "bcachefs");
1123 		if (ret)
1124 			return ret;
1125 
1126 		ret = sysfs_create_link(&ca->kobj, block, "block");
1127 		if (ret)
1128 			return ret;
1129 	}
1130 
1131 	return 0;
1132 }
1133 
1134 static struct bch_dev *__bch2_dev_alloc(struct bch_fs *c,
1135 					struct bch_member *member)
1136 {
1137 	struct bch_dev *ca;
1138 	unsigned i;
1139 
1140 	ca = kzalloc(sizeof(*ca), GFP_KERNEL);
1141 	if (!ca)
1142 		return NULL;
1143 
1144 	kobject_init(&ca->kobj, &bch2_dev_ktype);
1145 	init_completion(&ca->ref_completion);
1146 	init_completion(&ca->io_ref_completion);
1147 
1148 	init_rwsem(&ca->bucket_lock);
1149 
1150 	INIT_WORK(&ca->io_error_work, bch2_io_error_work);
1151 
1152 	bch2_time_stats_init(&ca->io_latency[READ]);
1153 	bch2_time_stats_init(&ca->io_latency[WRITE]);
1154 
1155 	ca->mi = bch2_mi_to_cpu(member);
1156 
1157 	for (i = 0; i < ARRAY_SIZE(member->errors); i++)
1158 		atomic64_set(&ca->errors[i], le64_to_cpu(member->errors[i]));
1159 
1160 	ca->uuid = member->uuid;
1161 
1162 	ca->nr_btree_reserve = DIV_ROUND_UP(BTREE_NODE_RESERVE,
1163 			     ca->mi.bucket_size / btree_sectors(c));
1164 
1165 	if (percpu_ref_init(&ca->ref, bch2_dev_ref_complete,
1166 			    0, GFP_KERNEL) ||
1167 	    percpu_ref_init(&ca->io_ref, bch2_dev_io_ref_complete,
1168 			    PERCPU_REF_INIT_DEAD, GFP_KERNEL) ||
1169 	    !(ca->sb_read_scratch = (void *) __get_free_page(GFP_KERNEL)) ||
1170 	    bch2_dev_buckets_alloc(c, ca) ||
1171 	    bioset_init(&ca->replica_set, 4,
1172 			offsetof(struct bch_write_bio, bio), 0) ||
1173 	    !(ca->io_done	= alloc_percpu(*ca->io_done)))
1174 		goto err;
1175 
1176 	return ca;
1177 err:
1178 	bch2_dev_free(ca);
1179 	return NULL;
1180 }
1181 
1182 static void bch2_dev_attach(struct bch_fs *c, struct bch_dev *ca,
1183 			    unsigned dev_idx)
1184 {
1185 	ca->dev_idx = dev_idx;
1186 	__set_bit(ca->dev_idx, ca->self.d);
1187 	scnprintf(ca->name, sizeof(ca->name), "dev-%u", dev_idx);
1188 
1189 	ca->fs = c;
1190 	rcu_assign_pointer(c->devs[ca->dev_idx], ca);
1191 
1192 	if (bch2_dev_sysfs_online(c, ca))
1193 		pr_warn("error creating sysfs objects");
1194 }
1195 
1196 static int bch2_dev_alloc(struct bch_fs *c, unsigned dev_idx)
1197 {
1198 	struct bch_member member = bch2_sb_member_get(c->disk_sb.sb, dev_idx);
1199 	struct bch_dev *ca = NULL;
1200 	int ret = 0;
1201 
1202 	if (bch2_fs_init_fault("dev_alloc"))
1203 		goto err;
1204 
1205 	ca = __bch2_dev_alloc(c, &member);
1206 	if (!ca)
1207 		goto err;
1208 
1209 	ca->fs = c;
1210 
1211 	bch2_dev_attach(c, ca, dev_idx);
1212 	return ret;
1213 err:
1214 	if (ca)
1215 		bch2_dev_free(ca);
1216 	return -BCH_ERR_ENOMEM_dev_alloc;
1217 }
1218 
1219 static int __bch2_dev_attach_bdev(struct bch_dev *ca, struct bch_sb_handle *sb)
1220 {
1221 	unsigned ret;
1222 
1223 	if (bch2_dev_is_online(ca)) {
1224 		bch_err(ca, "already have device online in slot %u",
1225 			sb->sb->dev_idx);
1226 		return -BCH_ERR_device_already_online;
1227 	}
1228 
1229 	if (get_capacity(sb->bdev->bd_disk) <
1230 	    ca->mi.bucket_size * ca->mi.nbuckets) {
1231 		bch_err(ca, "cannot online: device too small");
1232 		return -BCH_ERR_device_size_too_small;
1233 	}
1234 
1235 	BUG_ON(!percpu_ref_is_zero(&ca->io_ref));
1236 
1237 	ret = bch2_dev_journal_init(ca, sb->sb);
1238 	if (ret)
1239 		return ret;
1240 
1241 	/* Commit: */
1242 	ca->disk_sb = *sb;
1243 	memset(sb, 0, sizeof(*sb));
1244 
1245 	ca->dev = ca->disk_sb.bdev->bd_dev;
1246 
1247 	percpu_ref_reinit(&ca->io_ref);
1248 
1249 	return 0;
1250 }
1251 
1252 static int bch2_dev_attach_bdev(struct bch_fs *c, struct bch_sb_handle *sb)
1253 {
1254 	struct bch_dev *ca;
1255 	int ret;
1256 
1257 	lockdep_assert_held(&c->state_lock);
1258 
1259 	if (le64_to_cpu(sb->sb->seq) >
1260 	    le64_to_cpu(c->disk_sb.sb->seq))
1261 		bch2_sb_to_fs(c, sb->sb);
1262 
1263 	BUG_ON(sb->sb->dev_idx >= c->sb.nr_devices ||
1264 	       !c->devs[sb->sb->dev_idx]);
1265 
1266 	ca = bch_dev_locked(c, sb->sb->dev_idx);
1267 
1268 	ret = __bch2_dev_attach_bdev(ca, sb);
1269 	if (ret)
1270 		return ret;
1271 
1272 	bch2_dev_sysfs_online(c, ca);
1273 
1274 	if (c->sb.nr_devices == 1)
1275 		snprintf(c->name, sizeof(c->name), "%pg", ca->disk_sb.bdev);
1276 	snprintf(ca->name, sizeof(ca->name), "%pg", ca->disk_sb.bdev);
1277 
1278 	rebalance_wakeup(c);
1279 	return 0;
1280 }
1281 
1282 /* Device management: */
1283 
1284 /*
1285  * Note: this function is also used by the error paths - when a particular
1286  * device sees an error, we call it to determine whether we can just set the
1287  * device RO, or - if this function returns false - we'll set the whole
1288  * filesystem RO:
1289  *
1290  * XXX: maybe we should be more explicit about whether we're changing state
1291  * because we got an error or what have you?
1292  */
1293 bool bch2_dev_state_allowed(struct bch_fs *c, struct bch_dev *ca,
1294 			    enum bch_member_state new_state, int flags)
1295 {
1296 	struct bch_devs_mask new_online_devs;
1297 	struct bch_dev *ca2;
1298 	int i, nr_rw = 0, required;
1299 
1300 	lockdep_assert_held(&c->state_lock);
1301 
1302 	switch (new_state) {
1303 	case BCH_MEMBER_STATE_rw:
1304 		return true;
1305 	case BCH_MEMBER_STATE_ro:
1306 		if (ca->mi.state != BCH_MEMBER_STATE_rw)
1307 			return true;
1308 
1309 		/* do we have enough devices to write to?  */
1310 		for_each_member_device(ca2, c, i)
1311 			if (ca2 != ca)
1312 				nr_rw += ca2->mi.state == BCH_MEMBER_STATE_rw;
1313 
1314 		required = max(!(flags & BCH_FORCE_IF_METADATA_DEGRADED)
1315 			       ? c->opts.metadata_replicas
1316 			       : c->opts.metadata_replicas_required,
1317 			       !(flags & BCH_FORCE_IF_DATA_DEGRADED)
1318 			       ? c->opts.data_replicas
1319 			       : c->opts.data_replicas_required);
1320 
1321 		return nr_rw >= required;
1322 	case BCH_MEMBER_STATE_failed:
1323 	case BCH_MEMBER_STATE_spare:
1324 		if (ca->mi.state != BCH_MEMBER_STATE_rw &&
1325 		    ca->mi.state != BCH_MEMBER_STATE_ro)
1326 			return true;
1327 
1328 		/* do we have enough devices to read from?  */
1329 		new_online_devs = bch2_online_devs(c);
1330 		__clear_bit(ca->dev_idx, new_online_devs.d);
1331 
1332 		return bch2_have_enough_devs(c, new_online_devs, flags, false);
1333 	default:
1334 		BUG();
1335 	}
1336 }
1337 
1338 static bool bch2_fs_may_start(struct bch_fs *c)
1339 {
1340 	struct bch_dev *ca;
1341 	unsigned i, flags = 0;
1342 
1343 	if (c->opts.very_degraded)
1344 		flags |= BCH_FORCE_IF_DEGRADED|BCH_FORCE_IF_LOST;
1345 
1346 	if (c->opts.degraded)
1347 		flags |= BCH_FORCE_IF_DEGRADED;
1348 
1349 	if (!c->opts.degraded &&
1350 	    !c->opts.very_degraded) {
1351 		mutex_lock(&c->sb_lock);
1352 
1353 		for (i = 0; i < c->disk_sb.sb->nr_devices; i++) {
1354 			if (!bch2_dev_exists(c->disk_sb.sb, i))
1355 				continue;
1356 
1357 			ca = bch_dev_locked(c, i);
1358 
1359 			if (!bch2_dev_is_online(ca) &&
1360 			    (ca->mi.state == BCH_MEMBER_STATE_rw ||
1361 			     ca->mi.state == BCH_MEMBER_STATE_ro)) {
1362 				mutex_unlock(&c->sb_lock);
1363 				return false;
1364 			}
1365 		}
1366 		mutex_unlock(&c->sb_lock);
1367 	}
1368 
1369 	return bch2_have_enough_devs(c, bch2_online_devs(c), flags, true);
1370 }
1371 
1372 static void __bch2_dev_read_only(struct bch_fs *c, struct bch_dev *ca)
1373 {
1374 	/*
1375 	 * The allocator thread itself allocates btree nodes, so stop it first:
1376 	 */
1377 	bch2_dev_allocator_remove(c, ca);
1378 	bch2_dev_journal_stop(&c->journal, ca);
1379 }
1380 
1381 static void __bch2_dev_read_write(struct bch_fs *c, struct bch_dev *ca)
1382 {
1383 	lockdep_assert_held(&c->state_lock);
1384 
1385 	BUG_ON(ca->mi.state != BCH_MEMBER_STATE_rw);
1386 
1387 	bch2_dev_allocator_add(c, ca);
1388 	bch2_recalc_capacity(c);
1389 }
1390 
1391 int __bch2_dev_set_state(struct bch_fs *c, struct bch_dev *ca,
1392 			 enum bch_member_state new_state, int flags)
1393 {
1394 	struct bch_member *m;
1395 	int ret = 0;
1396 
1397 	if (ca->mi.state == new_state)
1398 		return 0;
1399 
1400 	if (!bch2_dev_state_allowed(c, ca, new_state, flags))
1401 		return -BCH_ERR_device_state_not_allowed;
1402 
1403 	if (new_state != BCH_MEMBER_STATE_rw)
1404 		__bch2_dev_read_only(c, ca);
1405 
1406 	bch_notice(ca, "%s", bch2_member_states[new_state]);
1407 
1408 	mutex_lock(&c->sb_lock);
1409 	m = bch2_members_v2_get_mut(c->disk_sb.sb, ca->dev_idx);
1410 	SET_BCH_MEMBER_STATE(m, new_state);
1411 	bch2_write_super(c);
1412 	mutex_unlock(&c->sb_lock);
1413 
1414 	if (new_state == BCH_MEMBER_STATE_rw)
1415 		__bch2_dev_read_write(c, ca);
1416 
1417 	rebalance_wakeup(c);
1418 
1419 	return ret;
1420 }
1421 
1422 int bch2_dev_set_state(struct bch_fs *c, struct bch_dev *ca,
1423 		       enum bch_member_state new_state, int flags)
1424 {
1425 	int ret;
1426 
1427 	down_write(&c->state_lock);
1428 	ret = __bch2_dev_set_state(c, ca, new_state, flags);
1429 	up_write(&c->state_lock);
1430 
1431 	return ret;
1432 }
1433 
1434 /* Device add/removal: */
1435 
1436 static int bch2_dev_remove_alloc(struct bch_fs *c, struct bch_dev *ca)
1437 {
1438 	struct bpos start	= POS(ca->dev_idx, 0);
1439 	struct bpos end		= POS(ca->dev_idx, U64_MAX);
1440 	int ret;
1441 
1442 	/*
1443 	 * We clear the LRU and need_discard btrees first so that we don't race
1444 	 * with bch2_do_invalidates() and bch2_do_discards()
1445 	 */
1446 	ret =   bch2_btree_delete_range(c, BTREE_ID_lru, start, end,
1447 					BTREE_TRIGGER_NORUN, NULL) ?:
1448 		bch2_btree_delete_range(c, BTREE_ID_need_discard, start, end,
1449 					BTREE_TRIGGER_NORUN, NULL) ?:
1450 		bch2_btree_delete_range(c, BTREE_ID_freespace, start, end,
1451 					BTREE_TRIGGER_NORUN, NULL) ?:
1452 		bch2_btree_delete_range(c, BTREE_ID_backpointers, start, end,
1453 					BTREE_TRIGGER_NORUN, NULL) ?:
1454 		bch2_btree_delete_range(c, BTREE_ID_alloc, start, end,
1455 					BTREE_TRIGGER_NORUN, NULL) ?:
1456 		bch2_btree_delete_range(c, BTREE_ID_bucket_gens, start, end,
1457 					BTREE_TRIGGER_NORUN, NULL);
1458 	if (ret)
1459 		bch_err_msg(c, ret, "removing dev alloc info");
1460 
1461 	return ret;
1462 }
1463 
1464 int bch2_dev_remove(struct bch_fs *c, struct bch_dev *ca, int flags)
1465 {
1466 	struct bch_member *m;
1467 	unsigned dev_idx = ca->dev_idx, data;
1468 	int ret;
1469 
1470 	down_write(&c->state_lock);
1471 
1472 	/*
1473 	 * We consume a reference to ca->ref, regardless of whether we succeed
1474 	 * or fail:
1475 	 */
1476 	percpu_ref_put(&ca->ref);
1477 
1478 	if (!bch2_dev_state_allowed(c, ca, BCH_MEMBER_STATE_failed, flags)) {
1479 		bch_err(ca, "Cannot remove without losing data");
1480 		ret = -BCH_ERR_device_state_not_allowed;
1481 		goto err;
1482 	}
1483 
1484 	__bch2_dev_read_only(c, ca);
1485 
1486 	ret = bch2_dev_data_drop(c, ca->dev_idx, flags);
1487 	if (ret) {
1488 		bch_err_msg(ca, ret, "dropping data");
1489 		goto err;
1490 	}
1491 
1492 	ret = bch2_dev_remove_alloc(c, ca);
1493 	if (ret) {
1494 		bch_err_msg(ca, ret, "deleting alloc info");
1495 		goto err;
1496 	}
1497 
1498 	ret = bch2_journal_flush_device_pins(&c->journal, ca->dev_idx);
1499 	if (ret) {
1500 		bch_err_msg(ca, ret, "flushing journal");
1501 		goto err;
1502 	}
1503 
1504 	ret = bch2_journal_flush(&c->journal);
1505 	if (ret) {
1506 		bch_err(ca, "journal error");
1507 		goto err;
1508 	}
1509 
1510 	ret = bch2_replicas_gc2(c);
1511 	if (ret) {
1512 		bch_err_msg(ca, ret, "in replicas_gc2()");
1513 		goto err;
1514 	}
1515 
1516 	data = bch2_dev_has_data(c, ca);
1517 	if (data) {
1518 		struct printbuf data_has = PRINTBUF;
1519 
1520 		prt_bitflags(&data_has, bch2_data_types, data);
1521 		bch_err(ca, "Remove failed, still has data (%s)", data_has.buf);
1522 		printbuf_exit(&data_has);
1523 		ret = -EBUSY;
1524 		goto err;
1525 	}
1526 
1527 	__bch2_dev_offline(c, ca);
1528 
1529 	mutex_lock(&c->sb_lock);
1530 	rcu_assign_pointer(c->devs[ca->dev_idx], NULL);
1531 	mutex_unlock(&c->sb_lock);
1532 
1533 	percpu_ref_kill(&ca->ref);
1534 	wait_for_completion(&ca->ref_completion);
1535 
1536 	bch2_dev_free(ca);
1537 
1538 	/*
1539 	 * At this point the device object has been removed in-core, but the
1540 	 * on-disk journal might still refer to the device index via sb device
1541 	 * usage entries. Recovery fails if it sees usage information for an
1542 	 * invalid device. Flush journal pins to push the back of the journal
1543 	 * past now invalid device index references before we update the
1544 	 * superblock, but after the device object has been removed so any
1545 	 * further journal writes elide usage info for the device.
1546 	 */
1547 	bch2_journal_flush_all_pins(&c->journal);
1548 
1549 	/*
1550 	 * Free this device's slot in the bch_member array - all pointers to
1551 	 * this device must be gone:
1552 	 */
1553 	mutex_lock(&c->sb_lock);
1554 	m = bch2_members_v2_get_mut(c->disk_sb.sb, dev_idx);
1555 	memset(&m->uuid, 0, sizeof(m->uuid));
1556 
1557 	bch2_write_super(c);
1558 
1559 	mutex_unlock(&c->sb_lock);
1560 	up_write(&c->state_lock);
1561 
1562 	bch2_dev_usage_journal_reserve(c);
1563 	return 0;
1564 err:
1565 	if (ca->mi.state == BCH_MEMBER_STATE_rw &&
1566 	    !percpu_ref_is_zero(&ca->io_ref))
1567 		__bch2_dev_read_write(c, ca);
1568 	up_write(&c->state_lock);
1569 	return ret;
1570 }
1571 
1572 /* Add new device to running filesystem: */
1573 int bch2_dev_add(struct bch_fs *c, const char *path)
1574 {
1575 	struct bch_opts opts = bch2_opts_empty();
1576 	struct bch_sb_handle sb;
1577 	struct bch_dev *ca = NULL;
1578 	struct bch_sb_field_members_v2 *mi;
1579 	struct bch_member dev_mi;
1580 	unsigned dev_idx, nr_devices, u64s;
1581 	struct printbuf errbuf = PRINTBUF;
1582 	struct printbuf label = PRINTBUF;
1583 	int ret;
1584 
1585 	ret = bch2_read_super(path, &opts, &sb);
1586 	if (ret) {
1587 		bch_err_msg(c, ret, "reading super");
1588 		goto err;
1589 	}
1590 
1591 	dev_mi = bch2_sb_member_get(sb.sb, sb.sb->dev_idx);
1592 
1593 	if (BCH_MEMBER_GROUP(&dev_mi)) {
1594 		bch2_disk_path_to_text_sb(&label, sb.sb, BCH_MEMBER_GROUP(&dev_mi) - 1);
1595 		if (label.allocation_failure) {
1596 			ret = -ENOMEM;
1597 			goto err;
1598 		}
1599 	}
1600 
1601 	ret = bch2_dev_may_add(sb.sb, c);
1602 	if (ret) {
1603 		bch_err_fn(c, ret);
1604 		goto err;
1605 	}
1606 
1607 	ca = __bch2_dev_alloc(c, &dev_mi);
1608 	if (!ca) {
1609 		ret = -ENOMEM;
1610 		goto err;
1611 	}
1612 
1613 	bch2_dev_usage_init(ca);
1614 
1615 	ret = __bch2_dev_attach_bdev(ca, &sb);
1616 	if (ret)
1617 		goto err;
1618 
1619 	ret = bch2_dev_journal_alloc(ca);
1620 	if (ret) {
1621 		bch_err_msg(c, ret, "allocating journal");
1622 		goto err;
1623 	}
1624 
1625 	down_write(&c->state_lock);
1626 	mutex_lock(&c->sb_lock);
1627 
1628 	ret = bch2_sb_from_fs(c, ca);
1629 	if (ret) {
1630 		bch_err_msg(c, ret, "setting up new superblock");
1631 		goto err_unlock;
1632 	}
1633 
1634 	if (dynamic_fault("bcachefs:add:no_slot"))
1635 		goto no_slot;
1636 
1637 	for (dev_idx = 0; dev_idx < BCH_SB_MEMBERS_MAX; dev_idx++)
1638 		if (!bch2_dev_exists(c->disk_sb.sb, dev_idx))
1639 			goto have_slot;
1640 no_slot:
1641 	ret = -BCH_ERR_ENOSPC_sb_members;
1642 	bch_err_msg(c, ret, "setting up new superblock");
1643 	goto err_unlock;
1644 
1645 have_slot:
1646 	nr_devices = max_t(unsigned, dev_idx + 1, c->sb.nr_devices);
1647 
1648 	mi = bch2_sb_field_get(c->disk_sb.sb, members_v2);
1649 	u64s = DIV_ROUND_UP(sizeof(struct bch_sb_field_members_v2) +
1650 			    le16_to_cpu(mi->member_bytes) * nr_devices, sizeof(u64));
1651 
1652 	mi = bch2_sb_field_resize(&c->disk_sb, members_v2, u64s);
1653 	if (!mi) {
1654 		ret = -BCH_ERR_ENOSPC_sb_members;
1655 		bch_err_msg(c, ret, "setting up new superblock");
1656 		goto err_unlock;
1657 	}
1658 	struct bch_member *m = bch2_members_v2_get_mut(c->disk_sb.sb, dev_idx);
1659 
1660 	/* success: */
1661 
1662 	*m = dev_mi;
1663 	m->last_mount = cpu_to_le64(ktime_get_real_seconds());
1664 	c->disk_sb.sb->nr_devices	= nr_devices;
1665 
1666 	ca->disk_sb.sb->dev_idx	= dev_idx;
1667 	bch2_dev_attach(c, ca, dev_idx);
1668 
1669 	if (BCH_MEMBER_GROUP(&dev_mi)) {
1670 		ret = __bch2_dev_group_set(c, ca, label.buf);
1671 		if (ret) {
1672 			bch_err_msg(c, ret, "creating new label");
1673 			goto err_unlock;
1674 		}
1675 	}
1676 
1677 	bch2_write_super(c);
1678 	mutex_unlock(&c->sb_lock);
1679 
1680 	bch2_dev_usage_journal_reserve(c);
1681 
1682 	ret = bch2_trans_mark_dev_sb(c, ca);
1683 	if (ret) {
1684 		bch_err_msg(ca, ret, "marking new superblock");
1685 		goto err_late;
1686 	}
1687 
1688 	ret = bch2_fs_freespace_init(c);
1689 	if (ret) {
1690 		bch_err_msg(ca, ret, "initializing free space");
1691 		goto err_late;
1692 	}
1693 
1694 	ca->new_fs_bucket_idx = 0;
1695 
1696 	if (ca->mi.state == BCH_MEMBER_STATE_rw)
1697 		__bch2_dev_read_write(c, ca);
1698 
1699 	up_write(&c->state_lock);
1700 	return 0;
1701 
1702 err_unlock:
1703 	mutex_unlock(&c->sb_lock);
1704 	up_write(&c->state_lock);
1705 err:
1706 	if (ca)
1707 		bch2_dev_free(ca);
1708 	bch2_free_super(&sb);
1709 	printbuf_exit(&label);
1710 	printbuf_exit(&errbuf);
1711 	return ret;
1712 err_late:
1713 	up_write(&c->state_lock);
1714 	ca = NULL;
1715 	goto err;
1716 }
1717 
1718 /* Hot add existing device to running filesystem: */
1719 int bch2_dev_online(struct bch_fs *c, const char *path)
1720 {
1721 	struct bch_opts opts = bch2_opts_empty();
1722 	struct bch_sb_handle sb = { NULL };
1723 	struct bch_dev *ca;
1724 	unsigned dev_idx;
1725 	int ret;
1726 
1727 	down_write(&c->state_lock);
1728 
1729 	ret = bch2_read_super(path, &opts, &sb);
1730 	if (ret) {
1731 		up_write(&c->state_lock);
1732 		return ret;
1733 	}
1734 
1735 	dev_idx = sb.sb->dev_idx;
1736 
1737 	ret = bch2_dev_in_fs(c->disk_sb.sb, sb.sb);
1738 	if (ret) {
1739 		bch_err_msg(c, ret, "bringing %s online", path);
1740 		goto err;
1741 	}
1742 
1743 	ret = bch2_dev_attach_bdev(c, &sb);
1744 	if (ret)
1745 		goto err;
1746 
1747 	ca = bch_dev_locked(c, dev_idx);
1748 
1749 	ret = bch2_trans_mark_dev_sb(c, ca);
1750 	if (ret) {
1751 		bch_err_msg(c, ret, "bringing %s online: error from bch2_trans_mark_dev_sb", path);
1752 		goto err;
1753 	}
1754 
1755 	if (ca->mi.state == BCH_MEMBER_STATE_rw)
1756 		__bch2_dev_read_write(c, ca);
1757 
1758 	if (!ca->mi.freespace_initialized) {
1759 		ret = bch2_dev_freespace_init(c, ca, 0, ca->mi.nbuckets);
1760 		bch_err_msg(ca, ret, "initializing free space");
1761 		if (ret)
1762 			goto err;
1763 	}
1764 
1765 	if (!ca->journal.nr) {
1766 		ret = bch2_dev_journal_alloc(ca);
1767 		bch_err_msg(ca, ret, "allocating journal");
1768 		if (ret)
1769 			goto err;
1770 	}
1771 
1772 	mutex_lock(&c->sb_lock);
1773 	bch2_members_v2_get_mut(c->disk_sb.sb, ca->dev_idx)->last_mount =
1774 		cpu_to_le64(ktime_get_real_seconds());
1775 	bch2_write_super(c);
1776 	mutex_unlock(&c->sb_lock);
1777 
1778 	up_write(&c->state_lock);
1779 	return 0;
1780 err:
1781 	up_write(&c->state_lock);
1782 	bch2_free_super(&sb);
1783 	return ret;
1784 }
1785 
1786 int bch2_dev_offline(struct bch_fs *c, struct bch_dev *ca, int flags)
1787 {
1788 	down_write(&c->state_lock);
1789 
1790 	if (!bch2_dev_is_online(ca)) {
1791 		bch_err(ca, "Already offline");
1792 		up_write(&c->state_lock);
1793 		return 0;
1794 	}
1795 
1796 	if (!bch2_dev_state_allowed(c, ca, BCH_MEMBER_STATE_failed, flags)) {
1797 		bch_err(ca, "Cannot offline required disk");
1798 		up_write(&c->state_lock);
1799 		return -BCH_ERR_device_state_not_allowed;
1800 	}
1801 
1802 	__bch2_dev_offline(c, ca);
1803 
1804 	up_write(&c->state_lock);
1805 	return 0;
1806 }
1807 
1808 int bch2_dev_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
1809 {
1810 	struct bch_member *m;
1811 	u64 old_nbuckets;
1812 	int ret = 0;
1813 
1814 	down_write(&c->state_lock);
1815 	old_nbuckets = ca->mi.nbuckets;
1816 
1817 	if (nbuckets < ca->mi.nbuckets) {
1818 		bch_err(ca, "Cannot shrink yet");
1819 		ret = -EINVAL;
1820 		goto err;
1821 	}
1822 
1823 	if (bch2_dev_is_online(ca) &&
1824 	    get_capacity(ca->disk_sb.bdev->bd_disk) <
1825 	    ca->mi.bucket_size * nbuckets) {
1826 		bch_err(ca, "New size larger than device");
1827 		ret = -BCH_ERR_device_size_too_small;
1828 		goto err;
1829 	}
1830 
1831 	ret = bch2_dev_buckets_resize(c, ca, nbuckets);
1832 	if (ret) {
1833 		bch_err_msg(ca, ret, "resizing buckets");
1834 		goto err;
1835 	}
1836 
1837 	ret = bch2_trans_mark_dev_sb(c, ca);
1838 	if (ret)
1839 		goto err;
1840 
1841 	mutex_lock(&c->sb_lock);
1842 	m = bch2_members_v2_get_mut(c->disk_sb.sb, ca->dev_idx);
1843 	m->nbuckets = cpu_to_le64(nbuckets);
1844 
1845 	bch2_write_super(c);
1846 	mutex_unlock(&c->sb_lock);
1847 
1848 	if (ca->mi.freespace_initialized) {
1849 		ret = bch2_dev_freespace_init(c, ca, old_nbuckets, nbuckets);
1850 		if (ret)
1851 			goto err;
1852 
1853 		/*
1854 		 * XXX: this is all wrong transactionally - we'll be able to do
1855 		 * this correctly after the disk space accounting rewrite
1856 		 */
1857 		ca->usage_base->d[BCH_DATA_free].buckets += nbuckets - old_nbuckets;
1858 	}
1859 
1860 	bch2_recalc_capacity(c);
1861 err:
1862 	up_write(&c->state_lock);
1863 	return ret;
1864 }
1865 
1866 /* return with ref on ca->ref: */
1867 struct bch_dev *bch2_dev_lookup(struct bch_fs *c, const char *name)
1868 {
1869 	struct bch_dev *ca;
1870 	unsigned i;
1871 
1872 	rcu_read_lock();
1873 	for_each_member_device_rcu(ca, c, i, NULL)
1874 		if (!strcmp(name, ca->name))
1875 			goto found;
1876 	ca = ERR_PTR(-BCH_ERR_ENOENT_dev_not_found);
1877 found:
1878 	rcu_read_unlock();
1879 
1880 	return ca;
1881 }
1882 
1883 /* Filesystem open: */
1884 
1885 struct bch_fs *bch2_fs_open(char * const *devices, unsigned nr_devices,
1886 			    struct bch_opts opts)
1887 {
1888 	DARRAY(struct bch_sb_handle) sbs = { 0 };
1889 	struct bch_fs *c = NULL;
1890 	struct bch_sb_handle *sb, *best = NULL;
1891 	struct printbuf errbuf = PRINTBUF;
1892 	int ret = 0;
1893 
1894 	if (!try_module_get(THIS_MODULE))
1895 		return ERR_PTR(-ENODEV);
1896 
1897 	if (!nr_devices) {
1898 		ret = -EINVAL;
1899 		goto err;
1900 	}
1901 
1902 	ret = darray_make_room(&sbs, nr_devices);
1903 	if (ret)
1904 		goto err;
1905 
1906 	for (unsigned i = 0; i < nr_devices; i++) {
1907 		struct bch_sb_handle sb = { NULL };
1908 
1909 		ret = bch2_read_super(devices[i], &opts, &sb);
1910 		if (ret)
1911 			goto err;
1912 
1913 		BUG_ON(darray_push(&sbs, sb));
1914 	}
1915 
1916 	darray_for_each(sbs, sb)
1917 		if (!best || le64_to_cpu(sb->sb->seq) > le64_to_cpu(best->sb->seq))
1918 			best = sb;
1919 
1920 	darray_for_each_reverse(sbs, sb) {
1921 		if (sb != best && !bch2_dev_exists(best->sb, sb->sb->dev_idx)) {
1922 			pr_info("%pg has been removed, skipping", sb->bdev);
1923 			bch2_free_super(sb);
1924 			darray_remove_item(&sbs, sb);
1925 			best -= best > sb;
1926 			continue;
1927 		}
1928 
1929 		ret = bch2_dev_in_fs(best->sb, sb->sb);
1930 		if (ret)
1931 			goto err_print;
1932 	}
1933 
1934 	c = bch2_fs_alloc(best->sb, opts);
1935 	ret = PTR_ERR_OR_ZERO(c);
1936 	if (ret)
1937 		goto err;
1938 
1939 	down_write(&c->state_lock);
1940 	darray_for_each(sbs, sb) {
1941 		ret = bch2_dev_attach_bdev(c, sb);
1942 		if (ret) {
1943 			up_write(&c->state_lock);
1944 			goto err;
1945 		}
1946 	}
1947 	up_write(&c->state_lock);
1948 
1949 	if (!bch2_fs_may_start(c)) {
1950 		ret = -BCH_ERR_insufficient_devices_to_start;
1951 		goto err_print;
1952 	}
1953 
1954 	if (!c->opts.nostart) {
1955 		ret = bch2_fs_start(c);
1956 		if (ret)
1957 			goto err;
1958 	}
1959 out:
1960 	darray_for_each(sbs, sb)
1961 		bch2_free_super(sb);
1962 	darray_exit(&sbs);
1963 	printbuf_exit(&errbuf);
1964 	module_put(THIS_MODULE);
1965 	return c;
1966 err_print:
1967 	pr_err("bch_fs_open err opening %s: %s",
1968 	       devices[0], bch2_err_str(ret));
1969 err:
1970 	if (!IS_ERR_OR_NULL(c))
1971 		bch2_fs_stop(c);
1972 	c = ERR_PTR(ret);
1973 	goto out;
1974 }
1975 
1976 /* Global interfaces/init */
1977 
1978 static void bcachefs_exit(void)
1979 {
1980 	bch2_debug_exit();
1981 	bch2_vfs_exit();
1982 	bch2_chardev_exit();
1983 	bch2_btree_key_cache_exit();
1984 	if (bcachefs_kset)
1985 		kset_unregister(bcachefs_kset);
1986 }
1987 
1988 static int __init bcachefs_init(void)
1989 {
1990 	bch2_bkey_pack_test();
1991 
1992 	if (!(bcachefs_kset = kset_create_and_add("bcachefs", NULL, fs_kobj)) ||
1993 	    bch2_btree_key_cache_init() ||
1994 	    bch2_chardev_init() ||
1995 	    bch2_vfs_init() ||
1996 	    bch2_debug_init())
1997 		goto err;
1998 
1999 	return 0;
2000 err:
2001 	bcachefs_exit();
2002 	return -ENOMEM;
2003 }
2004 
2005 #define BCH_DEBUG_PARAM(name, description)			\
2006 	bool bch2_##name;					\
2007 	module_param_named(name, bch2_##name, bool, 0644);	\
2008 	MODULE_PARM_DESC(name, description);
2009 BCH_DEBUG_PARAMS()
2010 #undef BCH_DEBUG_PARAM
2011 
2012 __maybe_unused
2013 static unsigned bch2_metadata_version = bcachefs_metadata_version_current;
2014 module_param_named(version, bch2_metadata_version, uint, 0400);
2015 
2016 module_exit(bcachefs_exit);
2017 module_init(bcachefs_init);
2018