xref: /linux/fs/bcachefs/super.c (revision 3027ce13e04eee76539ca65c2cb1028a01c8c508)
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_node_scan.h"
19 #include "btree_update_interior.h"
20 #include "btree_io.h"
21 #include "btree_write_buffer.h"
22 #include "buckets_waiting_for_journal.h"
23 #include "chardev.h"
24 #include "checksum.h"
25 #include "clock.h"
26 #include "compress.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-counters.h"
53 #include "sb-errors.h"
54 #include "sb-members.h"
55 #include "snapshot.h"
56 #include "subvolume.h"
57 #include "super.h"
58 #include "super-io.h"
59 #include "sysfs.h"
60 #include "thread_with_file.h"
61 #include "trace.h"
62 
63 #include <linux/backing-dev.h>
64 #include <linux/blkdev.h>
65 #include <linux/debugfs.h>
66 #include <linux/device.h>
67 #include <linux/idr.h>
68 #include <linux/module.h>
69 #include <linux/percpu.h>
70 #include <linux/random.h>
71 #include <linux/sysfs.h>
72 #include <crypto/hash.h>
73 
74 MODULE_LICENSE("GPL");
75 MODULE_AUTHOR("Kent Overstreet <kent.overstreet@gmail.com>");
76 MODULE_DESCRIPTION("bcachefs filesystem");
77 MODULE_SOFTDEP("pre: crc32c");
78 MODULE_SOFTDEP("pre: crc64");
79 MODULE_SOFTDEP("pre: sha256");
80 MODULE_SOFTDEP("pre: chacha20");
81 MODULE_SOFTDEP("pre: poly1305");
82 MODULE_SOFTDEP("pre: xxhash");
83 
84 const char * const bch2_fs_flag_strs[] = {
85 #define x(n)		#n,
86 	BCH_FS_FLAGS()
87 #undef x
88 	NULL
89 };
90 
91 __printf(2, 0)
92 static void bch2_print_maybe_redirect(struct stdio_redirect *stdio, const char *fmt, va_list args)
93 {
94 #ifdef __KERNEL__
95 	if (unlikely(stdio)) {
96 		if (fmt[0] == KERN_SOH[0])
97 			fmt += 2;
98 
99 		bch2_stdio_redirect_vprintf(stdio, true, fmt, args);
100 		return;
101 	}
102 #endif
103 	vprintk(fmt, args);
104 }
105 
106 void bch2_print_opts(struct bch_opts *opts, const char *fmt, ...)
107 {
108 	struct stdio_redirect *stdio = (void *)(unsigned long)opts->stdio;
109 
110 	va_list args;
111 	va_start(args, fmt);
112 	bch2_print_maybe_redirect(stdio, fmt, args);
113 	va_end(args);
114 }
115 
116 void __bch2_print(struct bch_fs *c, const char *fmt, ...)
117 {
118 	struct stdio_redirect *stdio = bch2_fs_stdio_redirect(c);
119 
120 	va_list args;
121 	va_start(args, fmt);
122 	bch2_print_maybe_redirect(stdio, fmt, args);
123 	va_end(args);
124 }
125 
126 #define KTYPE(type)							\
127 static const struct attribute_group type ## _group = {			\
128 	.attrs = type ## _files						\
129 };									\
130 									\
131 static const struct attribute_group *type ## _groups[] = {		\
132 	&type ## _group,						\
133 	NULL								\
134 };									\
135 									\
136 static const struct kobj_type type ## _ktype = {			\
137 	.release	= type ## _release,				\
138 	.sysfs_ops	= &type ## _sysfs_ops,				\
139 	.default_groups = type ## _groups				\
140 }
141 
142 static void bch2_fs_release(struct kobject *);
143 static void bch2_dev_release(struct kobject *);
144 static void bch2_fs_counters_release(struct kobject *k)
145 {
146 }
147 
148 static void bch2_fs_internal_release(struct kobject *k)
149 {
150 }
151 
152 static void bch2_fs_opts_dir_release(struct kobject *k)
153 {
154 }
155 
156 static void bch2_fs_time_stats_release(struct kobject *k)
157 {
158 }
159 
160 KTYPE(bch2_fs);
161 KTYPE(bch2_fs_counters);
162 KTYPE(bch2_fs_internal);
163 KTYPE(bch2_fs_opts_dir);
164 KTYPE(bch2_fs_time_stats);
165 KTYPE(bch2_dev);
166 
167 static struct kset *bcachefs_kset;
168 static LIST_HEAD(bch_fs_list);
169 static DEFINE_MUTEX(bch_fs_list_lock);
170 
171 DECLARE_WAIT_QUEUE_HEAD(bch2_read_only_wait);
172 
173 static void bch2_dev_free(struct bch_dev *);
174 static int bch2_dev_alloc(struct bch_fs *, unsigned);
175 static int bch2_dev_sysfs_online(struct bch_fs *, struct bch_dev *);
176 static void __bch2_dev_read_only(struct bch_fs *, struct bch_dev *);
177 
178 struct bch_fs *bch2_dev_to_fs(dev_t dev)
179 {
180 	struct bch_fs *c;
181 
182 	mutex_lock(&bch_fs_list_lock);
183 	rcu_read_lock();
184 
185 	list_for_each_entry(c, &bch_fs_list, list)
186 		for_each_member_device_rcu(c, ca, NULL)
187 			if (ca->disk_sb.bdev && ca->disk_sb.bdev->bd_dev == dev) {
188 				closure_get(&c->cl);
189 				goto found;
190 			}
191 	c = NULL;
192 found:
193 	rcu_read_unlock();
194 	mutex_unlock(&bch_fs_list_lock);
195 
196 	return c;
197 }
198 
199 static struct bch_fs *__bch2_uuid_to_fs(__uuid_t uuid)
200 {
201 	struct bch_fs *c;
202 
203 	lockdep_assert_held(&bch_fs_list_lock);
204 
205 	list_for_each_entry(c, &bch_fs_list, list)
206 		if (!memcmp(&c->disk_sb.sb->uuid, &uuid, sizeof(uuid)))
207 			return c;
208 
209 	return NULL;
210 }
211 
212 struct bch_fs *bch2_uuid_to_fs(__uuid_t uuid)
213 {
214 	struct bch_fs *c;
215 
216 	mutex_lock(&bch_fs_list_lock);
217 	c = __bch2_uuid_to_fs(uuid);
218 	if (c)
219 		closure_get(&c->cl);
220 	mutex_unlock(&bch_fs_list_lock);
221 
222 	return c;
223 }
224 
225 static void bch2_dev_usage_journal_reserve(struct bch_fs *c)
226 {
227 	unsigned nr = 0, u64s =
228 		((sizeof(struct jset_entry_dev_usage) +
229 		  sizeof(struct jset_entry_dev_usage_type) * BCH_DATA_NR)) /
230 		sizeof(u64);
231 
232 	rcu_read_lock();
233 	for_each_member_device_rcu(c, ca, NULL)
234 		nr++;
235 	rcu_read_unlock();
236 
237 	bch2_journal_entry_res_resize(&c->journal,
238 			&c->dev_usage_journal_res, u64s * nr);
239 }
240 
241 /* Filesystem RO/RW: */
242 
243 /*
244  * For startup/shutdown of RW stuff, the dependencies are:
245  *
246  * - foreground writes depend on copygc and rebalance (to free up space)
247  *
248  * - copygc and rebalance depend on mark and sweep gc (they actually probably
249  *   don't because they either reserve ahead of time or don't block if
250  *   allocations fail, but allocations can require mark and sweep gc to run
251  *   because of generation number wraparound)
252  *
253  * - all of the above depends on the allocator threads
254  *
255  * - allocator depends on the journal (when it rewrites prios and gens)
256  */
257 
258 static void __bch2_fs_read_only(struct bch_fs *c)
259 {
260 	unsigned clean_passes = 0;
261 	u64 seq = 0;
262 
263 	bch2_fs_ec_stop(c);
264 	bch2_open_buckets_stop(c, NULL, true);
265 	bch2_rebalance_stop(c);
266 	bch2_copygc_stop(c);
267 	bch2_gc_thread_stop(c);
268 	bch2_fs_ec_flush(c);
269 
270 	bch_verbose(c, "flushing journal and stopping allocators, journal seq %llu",
271 		    journal_cur_seq(&c->journal));
272 
273 	do {
274 		clean_passes++;
275 
276 		if (bch2_btree_interior_updates_flush(c) ||
277 		    bch2_journal_flush_all_pins(&c->journal) ||
278 		    bch2_btree_flush_all_writes(c) ||
279 		    seq != atomic64_read(&c->journal.seq)) {
280 			seq = atomic64_read(&c->journal.seq);
281 			clean_passes = 0;
282 		}
283 	} while (clean_passes < 2);
284 
285 	bch_verbose(c, "flushing journal and stopping allocators complete, journal seq %llu",
286 		    journal_cur_seq(&c->journal));
287 
288 	if (test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags) &&
289 	    !test_bit(BCH_FS_emergency_ro, &c->flags))
290 		set_bit(BCH_FS_clean_shutdown, &c->flags);
291 
292 	bch2_fs_journal_stop(&c->journal);
293 
294 	bch_info(c, "%sshutdown complete, journal seq %llu",
295 		 test_bit(BCH_FS_clean_shutdown, &c->flags) ? "" : "un",
296 		 c->journal.seq_ondisk);
297 
298 	/*
299 	 * After stopping journal:
300 	 */
301 	for_each_member_device(c, ca)
302 		bch2_dev_allocator_remove(c, ca);
303 }
304 
305 #ifndef BCH_WRITE_REF_DEBUG
306 static void bch2_writes_disabled(struct percpu_ref *writes)
307 {
308 	struct bch_fs *c = container_of(writes, struct bch_fs, writes);
309 
310 	set_bit(BCH_FS_write_disable_complete, &c->flags);
311 	wake_up(&bch2_read_only_wait);
312 }
313 #endif
314 
315 void bch2_fs_read_only(struct bch_fs *c)
316 {
317 	if (!test_bit(BCH_FS_rw, &c->flags)) {
318 		bch2_journal_reclaim_stop(&c->journal);
319 		return;
320 	}
321 
322 	BUG_ON(test_bit(BCH_FS_write_disable_complete, &c->flags));
323 
324 	bch_verbose(c, "going read-only");
325 
326 	/*
327 	 * Block new foreground-end write operations from starting - any new
328 	 * writes will return -EROFS:
329 	 */
330 	set_bit(BCH_FS_going_ro, &c->flags);
331 #ifndef BCH_WRITE_REF_DEBUG
332 	percpu_ref_kill(&c->writes);
333 #else
334 	for (unsigned i = 0; i < BCH_WRITE_REF_NR; i++)
335 		bch2_write_ref_put(c, i);
336 #endif
337 
338 	/*
339 	 * If we're not doing an emergency shutdown, we want to wait on
340 	 * outstanding writes to complete so they don't see spurious errors due
341 	 * to shutting down the allocator:
342 	 *
343 	 * If we are doing an emergency shutdown outstanding writes may
344 	 * hang until we shutdown the allocator so we don't want to wait
345 	 * on outstanding writes before shutting everything down - but
346 	 * we do need to wait on them before returning and signalling
347 	 * that going RO is complete:
348 	 */
349 	wait_event(bch2_read_only_wait,
350 		   test_bit(BCH_FS_write_disable_complete, &c->flags) ||
351 		   test_bit(BCH_FS_emergency_ro, &c->flags));
352 
353 	bool writes_disabled = test_bit(BCH_FS_write_disable_complete, &c->flags);
354 	if (writes_disabled)
355 		bch_verbose(c, "finished waiting for writes to stop");
356 
357 	__bch2_fs_read_only(c);
358 
359 	wait_event(bch2_read_only_wait,
360 		   test_bit(BCH_FS_write_disable_complete, &c->flags));
361 
362 	if (!writes_disabled)
363 		bch_verbose(c, "finished waiting for writes to stop");
364 
365 	clear_bit(BCH_FS_write_disable_complete, &c->flags);
366 	clear_bit(BCH_FS_going_ro, &c->flags);
367 	clear_bit(BCH_FS_rw, &c->flags);
368 
369 	if (!bch2_journal_error(&c->journal) &&
370 	    !test_bit(BCH_FS_error, &c->flags) &&
371 	    !test_bit(BCH_FS_emergency_ro, &c->flags) &&
372 	    test_bit(BCH_FS_started, &c->flags) &&
373 	    test_bit(BCH_FS_clean_shutdown, &c->flags) &&
374 	    c->recovery_pass_done >= BCH_RECOVERY_PASS_journal_replay) {
375 		BUG_ON(c->journal.last_empty_seq != journal_cur_seq(&c->journal));
376 		BUG_ON(atomic_read(&c->btree_cache.dirty));
377 		BUG_ON(atomic_long_read(&c->btree_key_cache.nr_dirty));
378 		BUG_ON(c->btree_write_buffer.inc.keys.nr);
379 		BUG_ON(c->btree_write_buffer.flushing.keys.nr);
380 
381 		bch_verbose(c, "marking filesystem clean");
382 		bch2_fs_mark_clean(c);
383 	} else {
384 		bch_verbose(c, "done going read-only, filesystem not clean");
385 	}
386 }
387 
388 static void bch2_fs_read_only_work(struct work_struct *work)
389 {
390 	struct bch_fs *c =
391 		container_of(work, struct bch_fs, read_only_work);
392 
393 	down_write(&c->state_lock);
394 	bch2_fs_read_only(c);
395 	up_write(&c->state_lock);
396 }
397 
398 static void bch2_fs_read_only_async(struct bch_fs *c)
399 {
400 	queue_work(system_long_wq, &c->read_only_work);
401 }
402 
403 bool bch2_fs_emergency_read_only(struct bch_fs *c)
404 {
405 	bool ret = !test_and_set_bit(BCH_FS_emergency_ro, &c->flags);
406 
407 	bch2_journal_halt(&c->journal);
408 	bch2_fs_read_only_async(c);
409 
410 	wake_up(&bch2_read_only_wait);
411 	return ret;
412 }
413 
414 static int bch2_fs_read_write_late(struct bch_fs *c)
415 {
416 	int ret;
417 
418 	/*
419 	 * Data move operations can't run until after check_snapshots has
420 	 * completed, and bch2_snapshot_is_ancestor() is available.
421 	 *
422 	 * Ideally we'd start copygc/rebalance earlier instead of waiting for
423 	 * all of recovery/fsck to complete:
424 	 */
425 	ret = bch2_copygc_start(c);
426 	if (ret) {
427 		bch_err(c, "error starting copygc thread");
428 		return ret;
429 	}
430 
431 	ret = bch2_rebalance_start(c);
432 	if (ret) {
433 		bch_err(c, "error starting rebalance thread");
434 		return ret;
435 	}
436 
437 	return 0;
438 }
439 
440 static int __bch2_fs_read_write(struct bch_fs *c, bool early)
441 {
442 	int ret;
443 
444 	if (test_bit(BCH_FS_initial_gc_unfixed, &c->flags)) {
445 		bch_err(c, "cannot go rw, unfixed btree errors");
446 		return -BCH_ERR_erofs_unfixed_errors;
447 	}
448 
449 	if (test_bit(BCH_FS_rw, &c->flags))
450 		return 0;
451 
452 	bch_info(c, "going read-write");
453 
454 	ret = bch2_sb_members_v2_init(c);
455 	if (ret)
456 		goto err;
457 
458 	ret = bch2_fs_mark_dirty(c);
459 	if (ret)
460 		goto err;
461 
462 	clear_bit(BCH_FS_clean_shutdown, &c->flags);
463 
464 	/*
465 	 * First journal write must be a flush write: after a clean shutdown we
466 	 * don't read the journal, so the first journal write may end up
467 	 * overwriting whatever was there previously, and there must always be
468 	 * at least one non-flush write in the journal or recovery will fail:
469 	 */
470 	set_bit(JOURNAL_NEED_FLUSH_WRITE, &c->journal.flags);
471 
472 	for_each_rw_member(c, ca)
473 		bch2_dev_allocator_add(c, ca);
474 	bch2_recalc_capacity(c);
475 
476 	set_bit(BCH_FS_rw, &c->flags);
477 	set_bit(BCH_FS_was_rw, &c->flags);
478 
479 #ifndef BCH_WRITE_REF_DEBUG
480 	percpu_ref_reinit(&c->writes);
481 #else
482 	for (unsigned i = 0; i < BCH_WRITE_REF_NR; i++) {
483 		BUG_ON(atomic_long_read(&c->writes[i]));
484 		atomic_long_inc(&c->writes[i]);
485 	}
486 #endif
487 
488 	ret = bch2_gc_thread_start(c);
489 	if (ret) {
490 		bch_err(c, "error starting gc thread");
491 		return ret;
492 	}
493 
494 	ret = bch2_journal_reclaim_start(&c->journal);
495 	if (ret)
496 		goto err;
497 
498 	if (!early) {
499 		ret = bch2_fs_read_write_late(c);
500 		if (ret)
501 			goto err;
502 	}
503 
504 	bch2_do_discards(c);
505 	bch2_do_invalidates(c);
506 	bch2_do_stripe_deletes(c);
507 	bch2_do_pending_node_rewrites(c);
508 	return 0;
509 err:
510 	if (test_bit(BCH_FS_rw, &c->flags))
511 		bch2_fs_read_only(c);
512 	else
513 		__bch2_fs_read_only(c);
514 	return ret;
515 }
516 
517 int bch2_fs_read_write(struct bch_fs *c)
518 {
519 	if (c->opts.recovery_pass_last &&
520 	    c->opts.recovery_pass_last < BCH_RECOVERY_PASS_journal_replay)
521 		return -BCH_ERR_erofs_norecovery;
522 
523 	if (c->opts.nochanges)
524 		return -BCH_ERR_erofs_nochanges;
525 
526 	return __bch2_fs_read_write(c, false);
527 }
528 
529 int bch2_fs_read_write_early(struct bch_fs *c)
530 {
531 	lockdep_assert_held(&c->state_lock);
532 
533 	return __bch2_fs_read_write(c, true);
534 }
535 
536 /* Filesystem startup/shutdown: */
537 
538 static void __bch2_fs_free(struct bch_fs *c)
539 {
540 	unsigned i;
541 
542 	for (i = 0; i < BCH_TIME_STAT_NR; i++)
543 		bch2_time_stats_exit(&c->times[i]);
544 
545 	bch2_find_btree_nodes_exit(&c->found_btree_nodes);
546 	bch2_free_pending_node_rewrites(c);
547 	bch2_fs_sb_errors_exit(c);
548 	bch2_fs_counters_exit(c);
549 	bch2_fs_snapshots_exit(c);
550 	bch2_fs_quota_exit(c);
551 	bch2_fs_fs_io_direct_exit(c);
552 	bch2_fs_fs_io_buffered_exit(c);
553 	bch2_fs_fsio_exit(c);
554 	bch2_fs_ec_exit(c);
555 	bch2_fs_encryption_exit(c);
556 	bch2_fs_nocow_locking_exit(c);
557 	bch2_fs_io_write_exit(c);
558 	bch2_fs_io_read_exit(c);
559 	bch2_fs_buckets_waiting_for_journal_exit(c);
560 	bch2_fs_btree_interior_update_exit(c);
561 	bch2_fs_btree_iter_exit(c);
562 	bch2_fs_btree_key_cache_exit(&c->btree_key_cache);
563 	bch2_fs_btree_cache_exit(c);
564 	bch2_fs_replicas_exit(c);
565 	bch2_fs_journal_exit(&c->journal);
566 	bch2_io_clock_exit(&c->io_clock[WRITE]);
567 	bch2_io_clock_exit(&c->io_clock[READ]);
568 	bch2_fs_compress_exit(c);
569 	bch2_journal_keys_put_initial(c);
570 	bch2_find_btree_nodes_exit(&c->found_btree_nodes);
571 	BUG_ON(atomic_read(&c->journal_keys.ref));
572 	bch2_fs_btree_write_buffer_exit(c);
573 	percpu_free_rwsem(&c->mark_lock);
574 	free_percpu(c->online_reserved);
575 
576 	darray_exit(&c->btree_roots_extra);
577 	free_percpu(c->pcpu);
578 	mempool_exit(&c->large_bkey_pool);
579 	mempool_exit(&c->btree_bounce_pool);
580 	bioset_exit(&c->btree_bio);
581 	mempool_exit(&c->fill_iter);
582 #ifndef BCH_WRITE_REF_DEBUG
583 	percpu_ref_exit(&c->writes);
584 #endif
585 	kfree(rcu_dereference_protected(c->disk_groups, 1));
586 	kfree(c->journal_seq_blacklist_table);
587 	kfree(c->unused_inode_hints);
588 
589 	if (c->write_ref_wq)
590 		destroy_workqueue(c->write_ref_wq);
591 	if (c->io_complete_wq)
592 		destroy_workqueue(c->io_complete_wq);
593 	if (c->copygc_wq)
594 		destroy_workqueue(c->copygc_wq);
595 	if (c->btree_io_complete_wq)
596 		destroy_workqueue(c->btree_io_complete_wq);
597 	if (c->btree_update_wq)
598 		destroy_workqueue(c->btree_update_wq);
599 
600 	bch2_free_super(&c->disk_sb);
601 	kvfree(c);
602 	module_put(THIS_MODULE);
603 }
604 
605 static void bch2_fs_release(struct kobject *kobj)
606 {
607 	struct bch_fs *c = container_of(kobj, struct bch_fs, kobj);
608 
609 	__bch2_fs_free(c);
610 }
611 
612 void __bch2_fs_stop(struct bch_fs *c)
613 {
614 	bch_verbose(c, "shutting down");
615 
616 	set_bit(BCH_FS_stopping, &c->flags);
617 
618 	cancel_work_sync(&c->journal_seq_blacklist_gc_work);
619 
620 	down_write(&c->state_lock);
621 	bch2_fs_read_only(c);
622 	up_write(&c->state_lock);
623 
624 	for_each_member_device(c, ca)
625 		if (ca->kobj.state_in_sysfs &&
626 		    ca->disk_sb.bdev)
627 			sysfs_remove_link(bdev_kobj(ca->disk_sb.bdev), "bcachefs");
628 
629 	if (c->kobj.state_in_sysfs)
630 		kobject_del(&c->kobj);
631 
632 	bch2_fs_debug_exit(c);
633 	bch2_fs_chardev_exit(c);
634 
635 	bch2_ro_ref_put(c);
636 	wait_event(c->ro_ref_wait, !refcount_read(&c->ro_ref));
637 
638 	kobject_put(&c->counters_kobj);
639 	kobject_put(&c->time_stats);
640 	kobject_put(&c->opts_dir);
641 	kobject_put(&c->internal);
642 
643 	/* btree prefetch might have kicked off reads in the background: */
644 	bch2_btree_flush_all_reads(c);
645 
646 	for_each_member_device(c, ca)
647 		cancel_work_sync(&ca->io_error_work);
648 
649 	cancel_work_sync(&c->read_only_work);
650 }
651 
652 void bch2_fs_free(struct bch_fs *c)
653 {
654 	unsigned i;
655 
656 	mutex_lock(&bch_fs_list_lock);
657 	list_del(&c->list);
658 	mutex_unlock(&bch_fs_list_lock);
659 
660 	closure_sync(&c->cl);
661 	closure_debug_destroy(&c->cl);
662 
663 	for (i = 0; i < c->sb.nr_devices; i++) {
664 		struct bch_dev *ca = rcu_dereference_protected(c->devs[i], true);
665 
666 		if (ca) {
667 			bch2_free_super(&ca->disk_sb);
668 			bch2_dev_free(ca);
669 		}
670 	}
671 
672 	bch_verbose(c, "shutdown complete");
673 
674 	kobject_put(&c->kobj);
675 }
676 
677 void bch2_fs_stop(struct bch_fs *c)
678 {
679 	__bch2_fs_stop(c);
680 	bch2_fs_free(c);
681 }
682 
683 static int bch2_fs_online(struct bch_fs *c)
684 {
685 	int ret = 0;
686 
687 	lockdep_assert_held(&bch_fs_list_lock);
688 
689 	if (__bch2_uuid_to_fs(c->sb.uuid)) {
690 		bch_err(c, "filesystem UUID already open");
691 		return -EINVAL;
692 	}
693 
694 	ret = bch2_fs_chardev_init(c);
695 	if (ret) {
696 		bch_err(c, "error creating character device");
697 		return ret;
698 	}
699 
700 	bch2_fs_debug_init(c);
701 
702 	ret = kobject_add(&c->kobj, NULL, "%pU", c->sb.user_uuid.b) ?:
703 	    kobject_add(&c->internal, &c->kobj, "internal") ?:
704 	    kobject_add(&c->opts_dir, &c->kobj, "options") ?:
705 #ifndef CONFIG_BCACHEFS_NO_LATENCY_ACCT
706 	    kobject_add(&c->time_stats, &c->kobj, "time_stats") ?:
707 #endif
708 	    kobject_add(&c->counters_kobj, &c->kobj, "counters") ?:
709 	    bch2_opts_create_sysfs_files(&c->opts_dir);
710 	if (ret) {
711 		bch_err(c, "error creating sysfs objects");
712 		return ret;
713 	}
714 
715 	down_write(&c->state_lock);
716 
717 	for_each_member_device(c, ca) {
718 		ret = bch2_dev_sysfs_online(c, ca);
719 		if (ret) {
720 			bch_err(c, "error creating sysfs objects");
721 			percpu_ref_put(&ca->ref);
722 			goto err;
723 		}
724 	}
725 
726 	BUG_ON(!list_empty(&c->list));
727 	list_add(&c->list, &bch_fs_list);
728 err:
729 	up_write(&c->state_lock);
730 	return ret;
731 }
732 
733 static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
734 {
735 	struct bch_fs *c;
736 	struct printbuf name = PRINTBUF;
737 	unsigned i, iter_size;
738 	int ret = 0;
739 
740 	c = kvmalloc(sizeof(struct bch_fs), GFP_KERNEL|__GFP_ZERO);
741 	if (!c) {
742 		c = ERR_PTR(-BCH_ERR_ENOMEM_fs_alloc);
743 		goto out;
744 	}
745 
746 	c->stdio = (void *)(unsigned long) opts.stdio;
747 
748 	__module_get(THIS_MODULE);
749 
750 	closure_init(&c->cl, NULL);
751 
752 	c->kobj.kset = bcachefs_kset;
753 	kobject_init(&c->kobj, &bch2_fs_ktype);
754 	kobject_init(&c->internal, &bch2_fs_internal_ktype);
755 	kobject_init(&c->opts_dir, &bch2_fs_opts_dir_ktype);
756 	kobject_init(&c->time_stats, &bch2_fs_time_stats_ktype);
757 	kobject_init(&c->counters_kobj, &bch2_fs_counters_ktype);
758 
759 	c->minor		= -1;
760 	c->disk_sb.fs_sb	= true;
761 
762 	init_rwsem(&c->state_lock);
763 	mutex_init(&c->sb_lock);
764 	mutex_init(&c->replicas_gc_lock);
765 	mutex_init(&c->btree_root_lock);
766 	INIT_WORK(&c->read_only_work, bch2_fs_read_only_work);
767 
768 	refcount_set(&c->ro_ref, 1);
769 	init_waitqueue_head(&c->ro_ref_wait);
770 	sema_init(&c->online_fsck_mutex, 1);
771 
772 	init_rwsem(&c->gc_lock);
773 	mutex_init(&c->gc_gens_lock);
774 	atomic_set(&c->journal_keys.ref, 1);
775 	c->journal_keys.initial_ref_held = true;
776 
777 	for (i = 0; i < BCH_TIME_STAT_NR; i++)
778 		bch2_time_stats_init(&c->times[i]);
779 
780 	bch2_fs_copygc_init(c);
781 	bch2_fs_btree_key_cache_init_early(&c->btree_key_cache);
782 	bch2_fs_btree_iter_init_early(c);
783 	bch2_fs_btree_interior_update_init_early(c);
784 	bch2_fs_allocator_background_init(c);
785 	bch2_fs_allocator_foreground_init(c);
786 	bch2_fs_rebalance_init(c);
787 	bch2_fs_quota_init(c);
788 	bch2_fs_ec_init_early(c);
789 	bch2_fs_move_init(c);
790 	bch2_fs_sb_errors_init_early(c);
791 
792 	INIT_LIST_HEAD(&c->list);
793 
794 	mutex_init(&c->usage_scratch_lock);
795 
796 	mutex_init(&c->bio_bounce_pages_lock);
797 	mutex_init(&c->snapshot_table_lock);
798 	init_rwsem(&c->snapshot_create_lock);
799 
800 	spin_lock_init(&c->btree_write_error_lock);
801 
802 	INIT_WORK(&c->journal_seq_blacklist_gc_work,
803 		  bch2_blacklist_entries_gc);
804 
805 	INIT_LIST_HEAD(&c->journal_iters);
806 
807 	INIT_LIST_HEAD(&c->fsck_error_msgs);
808 	mutex_init(&c->fsck_error_msgs_lock);
809 
810 	seqcount_init(&c->gc_pos_lock);
811 
812 	seqcount_init(&c->usage_lock);
813 
814 	sema_init(&c->io_in_flight, 128);
815 
816 	INIT_LIST_HEAD(&c->vfs_inodes_list);
817 	mutex_init(&c->vfs_inodes_lock);
818 
819 	c->copy_gc_enabled		= 1;
820 	c->rebalance.enabled		= 1;
821 	c->promote_whole_extents	= true;
822 
823 	c->journal.flush_write_time	= &c->times[BCH_TIME_journal_flush_write];
824 	c->journal.noflush_write_time	= &c->times[BCH_TIME_journal_noflush_write];
825 	c->journal.flush_seq_time	= &c->times[BCH_TIME_journal_flush_seq];
826 
827 	bch2_fs_btree_cache_init_early(&c->btree_cache);
828 
829 	mutex_init(&c->sectors_available_lock);
830 
831 	ret = percpu_init_rwsem(&c->mark_lock);
832 	if (ret)
833 		goto err;
834 
835 	mutex_lock(&c->sb_lock);
836 	ret = bch2_sb_to_fs(c, sb);
837 	mutex_unlock(&c->sb_lock);
838 
839 	if (ret)
840 		goto err;
841 
842 	pr_uuid(&name, c->sb.user_uuid.b);
843 	ret = name.allocation_failure ? -BCH_ERR_ENOMEM_fs_name_alloc : 0;
844 	if (ret)
845 		goto err;
846 
847 	strscpy(c->name, name.buf, sizeof(c->name));
848 	printbuf_exit(&name);
849 
850 	/* Compat: */
851 	if (le16_to_cpu(sb->version) <= bcachefs_metadata_version_inode_v2 &&
852 	    !BCH_SB_JOURNAL_FLUSH_DELAY(sb))
853 		SET_BCH_SB_JOURNAL_FLUSH_DELAY(sb, 1000);
854 
855 	if (le16_to_cpu(sb->version) <= bcachefs_metadata_version_inode_v2 &&
856 	    !BCH_SB_JOURNAL_RECLAIM_DELAY(sb))
857 		SET_BCH_SB_JOURNAL_RECLAIM_DELAY(sb, 100);
858 
859 	c->opts = bch2_opts_default;
860 	ret = bch2_opts_from_sb(&c->opts, sb);
861 	if (ret)
862 		goto err;
863 
864 	bch2_opts_apply(&c->opts, opts);
865 
866 	c->btree_key_cache_btrees |= 1U << BTREE_ID_alloc;
867 	if (c->opts.inodes_use_key_cache)
868 		c->btree_key_cache_btrees |= 1U << BTREE_ID_inodes;
869 	c->btree_key_cache_btrees |= 1U << BTREE_ID_logged_ops;
870 
871 	c->block_bits		= ilog2(block_sectors(c));
872 	c->btree_foreground_merge_threshold = BTREE_FOREGROUND_MERGE_THRESHOLD(c);
873 
874 	if (bch2_fs_init_fault("fs_alloc")) {
875 		bch_err(c, "fs_alloc fault injected");
876 		ret = -EFAULT;
877 		goto err;
878 	}
879 
880 	iter_size = sizeof(struct sort_iter) +
881 		(btree_blocks(c) + 1) * 2 *
882 		sizeof(struct sort_iter_set);
883 
884 	c->inode_shard_bits = ilog2(roundup_pow_of_two(num_possible_cpus()));
885 
886 	if (!(c->btree_update_wq = alloc_workqueue("bcachefs",
887 				WQ_HIGHPRI|WQ_FREEZABLE|WQ_MEM_RECLAIM|WQ_UNBOUND, 512)) ||
888 	    !(c->btree_io_complete_wq = alloc_workqueue("bcachefs_btree_io",
889 				WQ_HIGHPRI|WQ_FREEZABLE|WQ_MEM_RECLAIM, 1)) ||
890 	    !(c->copygc_wq = alloc_workqueue("bcachefs_copygc",
891 				WQ_HIGHPRI|WQ_FREEZABLE|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE, 1)) ||
892 	    !(c->io_complete_wq = alloc_workqueue("bcachefs_io",
893 				WQ_HIGHPRI|WQ_FREEZABLE|WQ_MEM_RECLAIM, 512)) ||
894 	    !(c->write_ref_wq = alloc_workqueue("bcachefs_write_ref",
895 				WQ_FREEZABLE, 0)) ||
896 #ifndef BCH_WRITE_REF_DEBUG
897 	    percpu_ref_init(&c->writes, bch2_writes_disabled,
898 			    PERCPU_REF_INIT_DEAD, GFP_KERNEL) ||
899 #endif
900 	    mempool_init_kmalloc_pool(&c->fill_iter, 1, iter_size) ||
901 	    bioset_init(&c->btree_bio, 1,
902 			max(offsetof(struct btree_read_bio, bio),
903 			    offsetof(struct btree_write_bio, wbio.bio)),
904 			BIOSET_NEED_BVECS) ||
905 	    !(c->pcpu = alloc_percpu(struct bch_fs_pcpu)) ||
906 	    !(c->online_reserved = alloc_percpu(u64)) ||
907 	    mempool_init_kvmalloc_pool(&c->btree_bounce_pool, 1,
908 				       c->opts.btree_node_size) ||
909 	    mempool_init_kmalloc_pool(&c->large_bkey_pool, 1, 2048) ||
910 	    !(c->unused_inode_hints = kcalloc(1U << c->inode_shard_bits,
911 					      sizeof(u64), GFP_KERNEL))) {
912 		ret = -BCH_ERR_ENOMEM_fs_other_alloc;
913 		goto err;
914 	}
915 
916 	ret = bch2_fs_counters_init(c) ?:
917 	    bch2_fs_sb_errors_init(c) ?:
918 	    bch2_io_clock_init(&c->io_clock[READ]) ?:
919 	    bch2_io_clock_init(&c->io_clock[WRITE]) ?:
920 	    bch2_fs_journal_init(&c->journal) ?:
921 	    bch2_fs_replicas_init(c) ?:
922 	    bch2_fs_btree_cache_init(c) ?:
923 	    bch2_fs_btree_key_cache_init(&c->btree_key_cache) ?:
924 	    bch2_fs_btree_iter_init(c) ?:
925 	    bch2_fs_btree_interior_update_init(c) ?:
926 	    bch2_fs_buckets_waiting_for_journal_init(c) ?:
927 	    bch2_fs_btree_write_buffer_init(c) ?:
928 	    bch2_fs_subvolumes_init(c) ?:
929 	    bch2_fs_io_read_init(c) ?:
930 	    bch2_fs_io_write_init(c) ?:
931 	    bch2_fs_nocow_locking_init(c) ?:
932 	    bch2_fs_encryption_init(c) ?:
933 	    bch2_fs_compress_init(c) ?:
934 	    bch2_fs_ec_init(c) ?:
935 	    bch2_fs_fsio_init(c) ?:
936 	    bch2_fs_fs_io_buffered_init(c) ?:
937 	    bch2_fs_fs_io_direct_init(c);
938 	if (ret)
939 		goto err;
940 
941 	for (i = 0; i < c->sb.nr_devices; i++)
942 		if (bch2_dev_exists(c->disk_sb.sb, i) &&
943 		    bch2_dev_alloc(c, i)) {
944 			ret = -EEXIST;
945 			goto err;
946 		}
947 
948 	bch2_journal_entry_res_resize(&c->journal,
949 			&c->btree_root_journal_res,
950 			BTREE_ID_NR * (JSET_KEYS_U64s + BKEY_BTREE_PTR_U64s_MAX));
951 	bch2_dev_usage_journal_reserve(c);
952 	bch2_journal_entry_res_resize(&c->journal,
953 			&c->clock_journal_res,
954 			(sizeof(struct jset_entry_clock) / sizeof(u64)) * 2);
955 
956 	mutex_lock(&bch_fs_list_lock);
957 	ret = bch2_fs_online(c);
958 	mutex_unlock(&bch_fs_list_lock);
959 
960 	if (ret)
961 		goto err;
962 out:
963 	return c;
964 err:
965 	bch2_fs_free(c);
966 	c = ERR_PTR(ret);
967 	goto out;
968 }
969 
970 noinline_for_stack
971 static void print_mount_opts(struct bch_fs *c)
972 {
973 	enum bch_opt_id i;
974 	struct printbuf p = PRINTBUF;
975 	bool first = true;
976 
977 	prt_str(&p, "mounting version ");
978 	bch2_version_to_text(&p, c->sb.version);
979 
980 	if (c->opts.read_only) {
981 		prt_str(&p, " opts=");
982 		first = false;
983 		prt_printf(&p, "ro");
984 	}
985 
986 	for (i = 0; i < bch2_opts_nr; i++) {
987 		const struct bch_option *opt = &bch2_opt_table[i];
988 		u64 v = bch2_opt_get_by_id(&c->opts, i);
989 
990 		if (!(opt->flags & OPT_MOUNT))
991 			continue;
992 
993 		if (v == bch2_opt_get_by_id(&bch2_opts_default, i))
994 			continue;
995 
996 		prt_str(&p, first ? " opts=" : ",");
997 		first = false;
998 		bch2_opt_to_text(&p, c, c->disk_sb.sb, opt, v, OPT_SHOW_MOUNT_STYLE);
999 	}
1000 
1001 	bch_info(c, "%s", p.buf);
1002 	printbuf_exit(&p);
1003 }
1004 
1005 int bch2_fs_start(struct bch_fs *c)
1006 {
1007 	time64_t now = ktime_get_real_seconds();
1008 	int ret;
1009 
1010 	print_mount_opts(c);
1011 
1012 	down_write(&c->state_lock);
1013 
1014 	BUG_ON(test_bit(BCH_FS_started, &c->flags));
1015 
1016 	mutex_lock(&c->sb_lock);
1017 
1018 	ret = bch2_sb_members_v2_init(c);
1019 	if (ret) {
1020 		mutex_unlock(&c->sb_lock);
1021 		goto err;
1022 	}
1023 
1024 	for_each_online_member(c, ca)
1025 		bch2_members_v2_get_mut(c->disk_sb.sb, ca->dev_idx)->last_mount = cpu_to_le64(now);
1026 
1027 	struct bch_sb_field_ext *ext =
1028 		bch2_sb_field_get_minsize(&c->disk_sb, ext, sizeof(*ext) / sizeof(u64));
1029 	mutex_unlock(&c->sb_lock);
1030 
1031 	if (!ext) {
1032 		bch_err(c, "insufficient space in superblock for sb_field_ext");
1033 		ret = -BCH_ERR_ENOSPC_sb;
1034 		goto err;
1035 	}
1036 
1037 	for_each_rw_member(c, ca)
1038 		bch2_dev_allocator_add(c, ca);
1039 	bch2_recalc_capacity(c);
1040 
1041 	ret = BCH_SB_INITIALIZED(c->disk_sb.sb)
1042 		? bch2_fs_recovery(c)
1043 		: bch2_fs_initialize(c);
1044 	if (ret)
1045 		goto err;
1046 
1047 	ret = bch2_opts_check_may_set(c);
1048 	if (ret)
1049 		goto err;
1050 
1051 	if (bch2_fs_init_fault("fs_start")) {
1052 		bch_err(c, "fs_start fault injected");
1053 		ret = -EINVAL;
1054 		goto err;
1055 	}
1056 
1057 	set_bit(BCH_FS_started, &c->flags);
1058 
1059 	if (c->opts.read_only) {
1060 		bch2_fs_read_only(c);
1061 	} else {
1062 		ret = !test_bit(BCH_FS_rw, &c->flags)
1063 			? bch2_fs_read_write(c)
1064 			: bch2_fs_read_write_late(c);
1065 		if (ret)
1066 			goto err;
1067 	}
1068 
1069 	ret = 0;
1070 err:
1071 	if (ret)
1072 		bch_err_msg(c, ret, "starting filesystem");
1073 	else
1074 		bch_verbose(c, "done starting filesystem");
1075 	up_write(&c->state_lock);
1076 	return ret;
1077 }
1078 
1079 static int bch2_dev_may_add(struct bch_sb *sb, struct bch_fs *c)
1080 {
1081 	struct bch_member m = bch2_sb_member_get(sb, sb->dev_idx);
1082 
1083 	if (le16_to_cpu(sb->block_size) != block_sectors(c))
1084 		return -BCH_ERR_mismatched_block_size;
1085 
1086 	if (le16_to_cpu(m.bucket_size) <
1087 	    BCH_SB_BTREE_NODE_SIZE(c->disk_sb.sb))
1088 		return -BCH_ERR_bucket_size_too_small;
1089 
1090 	return 0;
1091 }
1092 
1093 static int bch2_dev_in_fs(struct bch_sb_handle *fs,
1094 			  struct bch_sb_handle *sb,
1095 			  struct bch_opts *opts)
1096 {
1097 	if (fs == sb)
1098 		return 0;
1099 
1100 	if (!uuid_equal(&fs->sb->uuid, &sb->sb->uuid))
1101 		return -BCH_ERR_device_not_a_member_of_filesystem;
1102 
1103 	if (!bch2_dev_exists(fs->sb, sb->sb->dev_idx))
1104 		return -BCH_ERR_device_has_been_removed;
1105 
1106 	if (fs->sb->block_size != sb->sb->block_size)
1107 		return -BCH_ERR_mismatched_block_size;
1108 
1109 	if (le16_to_cpu(fs->sb->version) < bcachefs_metadata_version_member_seq ||
1110 	    le16_to_cpu(sb->sb->version) < bcachefs_metadata_version_member_seq)
1111 		return 0;
1112 
1113 	if (fs->sb->seq == sb->sb->seq &&
1114 	    fs->sb->write_time != sb->sb->write_time) {
1115 		struct printbuf buf = PRINTBUF;
1116 
1117 		prt_str(&buf, "Split brain detected between ");
1118 		prt_bdevname(&buf, sb->bdev);
1119 		prt_str(&buf, " and ");
1120 		prt_bdevname(&buf, fs->bdev);
1121 		prt_char(&buf, ':');
1122 		prt_newline(&buf);
1123 		prt_printf(&buf, "seq=%llu but write_time different, got", le64_to_cpu(sb->sb->seq));
1124 		prt_newline(&buf);
1125 
1126 		prt_bdevname(&buf, fs->bdev);
1127 		prt_char(&buf, ' ');
1128 		bch2_prt_datetime(&buf, le64_to_cpu(fs->sb->write_time));;
1129 		prt_newline(&buf);
1130 
1131 		prt_bdevname(&buf, sb->bdev);
1132 		prt_char(&buf, ' ');
1133 		bch2_prt_datetime(&buf, le64_to_cpu(sb->sb->write_time));;
1134 		prt_newline(&buf);
1135 
1136 		if (!opts->no_splitbrain_check)
1137 			prt_printf(&buf, "Not using older sb");
1138 
1139 		pr_err("%s", buf.buf);
1140 		printbuf_exit(&buf);
1141 
1142 		if (!opts->no_splitbrain_check)
1143 			return -BCH_ERR_device_splitbrain;
1144 	}
1145 
1146 	struct bch_member m = bch2_sb_member_get(fs->sb, sb->sb->dev_idx);
1147 	u64 seq_from_fs		= le64_to_cpu(m.seq);
1148 	u64 seq_from_member	= le64_to_cpu(sb->sb->seq);
1149 
1150 	if (seq_from_fs && seq_from_fs < seq_from_member) {
1151 		struct printbuf buf = PRINTBUF;
1152 
1153 		prt_str(&buf, "Split brain detected between ");
1154 		prt_bdevname(&buf, sb->bdev);
1155 		prt_str(&buf, " and ");
1156 		prt_bdevname(&buf, fs->bdev);
1157 		prt_char(&buf, ':');
1158 		prt_newline(&buf);
1159 
1160 		prt_bdevname(&buf, fs->bdev);
1161 		prt_str(&buf, " believes seq of ");
1162 		prt_bdevname(&buf, sb->bdev);
1163 		prt_printf(&buf, " to be %llu, but ", seq_from_fs);
1164 		prt_bdevname(&buf, sb->bdev);
1165 		prt_printf(&buf, " has %llu\n", seq_from_member);
1166 
1167 		if (!opts->no_splitbrain_check) {
1168 			prt_str(&buf, "Not using ");
1169 			prt_bdevname(&buf, sb->bdev);
1170 		}
1171 
1172 		pr_err("%s", buf.buf);
1173 		printbuf_exit(&buf);
1174 
1175 		if (!opts->no_splitbrain_check)
1176 			return -BCH_ERR_device_splitbrain;
1177 	}
1178 
1179 	return 0;
1180 }
1181 
1182 /* Device startup/shutdown: */
1183 
1184 static void bch2_dev_release(struct kobject *kobj)
1185 {
1186 	struct bch_dev *ca = container_of(kobj, struct bch_dev, kobj);
1187 
1188 	kfree(ca);
1189 }
1190 
1191 static void bch2_dev_free(struct bch_dev *ca)
1192 {
1193 	cancel_work_sync(&ca->io_error_work);
1194 
1195 	if (ca->kobj.state_in_sysfs &&
1196 	    ca->disk_sb.bdev)
1197 		sysfs_remove_link(bdev_kobj(ca->disk_sb.bdev), "bcachefs");
1198 
1199 	if (ca->kobj.state_in_sysfs)
1200 		kobject_del(&ca->kobj);
1201 
1202 	bch2_free_super(&ca->disk_sb);
1203 	bch2_dev_journal_exit(ca);
1204 
1205 	free_percpu(ca->io_done);
1206 	bioset_exit(&ca->replica_set);
1207 	bch2_dev_buckets_free(ca);
1208 	free_page((unsigned long) ca->sb_read_scratch);
1209 
1210 	bch2_time_stats_quantiles_exit(&ca->io_latency[WRITE]);
1211 	bch2_time_stats_quantiles_exit(&ca->io_latency[READ]);
1212 
1213 	percpu_ref_exit(&ca->io_ref);
1214 	percpu_ref_exit(&ca->ref);
1215 	kobject_put(&ca->kobj);
1216 }
1217 
1218 static void __bch2_dev_offline(struct bch_fs *c, struct bch_dev *ca)
1219 {
1220 
1221 	lockdep_assert_held(&c->state_lock);
1222 
1223 	if (percpu_ref_is_zero(&ca->io_ref))
1224 		return;
1225 
1226 	__bch2_dev_read_only(c, ca);
1227 
1228 	reinit_completion(&ca->io_ref_completion);
1229 	percpu_ref_kill(&ca->io_ref);
1230 	wait_for_completion(&ca->io_ref_completion);
1231 
1232 	if (ca->kobj.state_in_sysfs) {
1233 		sysfs_remove_link(bdev_kobj(ca->disk_sb.bdev), "bcachefs");
1234 		sysfs_remove_link(&ca->kobj, "block");
1235 	}
1236 
1237 	bch2_free_super(&ca->disk_sb);
1238 	bch2_dev_journal_exit(ca);
1239 }
1240 
1241 static void bch2_dev_ref_complete(struct percpu_ref *ref)
1242 {
1243 	struct bch_dev *ca = container_of(ref, struct bch_dev, ref);
1244 
1245 	complete(&ca->ref_completion);
1246 }
1247 
1248 static void bch2_dev_io_ref_complete(struct percpu_ref *ref)
1249 {
1250 	struct bch_dev *ca = container_of(ref, struct bch_dev, io_ref);
1251 
1252 	complete(&ca->io_ref_completion);
1253 }
1254 
1255 static int bch2_dev_sysfs_online(struct bch_fs *c, struct bch_dev *ca)
1256 {
1257 	int ret;
1258 
1259 	if (!c->kobj.state_in_sysfs)
1260 		return 0;
1261 
1262 	if (!ca->kobj.state_in_sysfs) {
1263 		ret = kobject_add(&ca->kobj, &c->kobj,
1264 				  "dev-%u", ca->dev_idx);
1265 		if (ret)
1266 			return ret;
1267 	}
1268 
1269 	if (ca->disk_sb.bdev) {
1270 		struct kobject *block = bdev_kobj(ca->disk_sb.bdev);
1271 
1272 		ret = sysfs_create_link(block, &ca->kobj, "bcachefs");
1273 		if (ret)
1274 			return ret;
1275 
1276 		ret = sysfs_create_link(&ca->kobj, block, "block");
1277 		if (ret)
1278 			return ret;
1279 	}
1280 
1281 	return 0;
1282 }
1283 
1284 static struct bch_dev *__bch2_dev_alloc(struct bch_fs *c,
1285 					struct bch_member *member)
1286 {
1287 	struct bch_dev *ca;
1288 	unsigned i;
1289 
1290 	ca = kzalloc(sizeof(*ca), GFP_KERNEL);
1291 	if (!ca)
1292 		return NULL;
1293 
1294 	kobject_init(&ca->kobj, &bch2_dev_ktype);
1295 	init_completion(&ca->ref_completion);
1296 	init_completion(&ca->io_ref_completion);
1297 
1298 	init_rwsem(&ca->bucket_lock);
1299 
1300 	INIT_WORK(&ca->io_error_work, bch2_io_error_work);
1301 
1302 	bch2_time_stats_quantiles_init(&ca->io_latency[READ]);
1303 	bch2_time_stats_quantiles_init(&ca->io_latency[WRITE]);
1304 
1305 	ca->mi = bch2_mi_to_cpu(member);
1306 
1307 	for (i = 0; i < ARRAY_SIZE(member->errors); i++)
1308 		atomic64_set(&ca->errors[i], le64_to_cpu(member->errors[i]));
1309 
1310 	ca->uuid = member->uuid;
1311 
1312 	ca->nr_btree_reserve = DIV_ROUND_UP(BTREE_NODE_RESERVE,
1313 			     ca->mi.bucket_size / btree_sectors(c));
1314 
1315 	if (percpu_ref_init(&ca->ref, bch2_dev_ref_complete,
1316 			    0, GFP_KERNEL) ||
1317 	    percpu_ref_init(&ca->io_ref, bch2_dev_io_ref_complete,
1318 			    PERCPU_REF_INIT_DEAD, GFP_KERNEL) ||
1319 	    !(ca->sb_read_scratch = (void *) __get_free_page(GFP_KERNEL)) ||
1320 	    bch2_dev_buckets_alloc(c, ca) ||
1321 	    bioset_init(&ca->replica_set, 4,
1322 			offsetof(struct bch_write_bio, bio), 0) ||
1323 	    !(ca->io_done	= alloc_percpu(*ca->io_done)))
1324 		goto err;
1325 
1326 	return ca;
1327 err:
1328 	bch2_dev_free(ca);
1329 	return NULL;
1330 }
1331 
1332 static void bch2_dev_attach(struct bch_fs *c, struct bch_dev *ca,
1333 			    unsigned dev_idx)
1334 {
1335 	ca->dev_idx = dev_idx;
1336 	__set_bit(ca->dev_idx, ca->self.d);
1337 	scnprintf(ca->name, sizeof(ca->name), "dev-%u", dev_idx);
1338 
1339 	ca->fs = c;
1340 	rcu_assign_pointer(c->devs[ca->dev_idx], ca);
1341 
1342 	if (bch2_dev_sysfs_online(c, ca))
1343 		pr_warn("error creating sysfs objects");
1344 }
1345 
1346 static int bch2_dev_alloc(struct bch_fs *c, unsigned dev_idx)
1347 {
1348 	struct bch_member member = bch2_sb_member_get(c->disk_sb.sb, dev_idx);
1349 	struct bch_dev *ca = NULL;
1350 	int ret = 0;
1351 
1352 	if (bch2_fs_init_fault("dev_alloc"))
1353 		goto err;
1354 
1355 	ca = __bch2_dev_alloc(c, &member);
1356 	if (!ca)
1357 		goto err;
1358 
1359 	ca->fs = c;
1360 
1361 	bch2_dev_attach(c, ca, dev_idx);
1362 	return ret;
1363 err:
1364 	if (ca)
1365 		bch2_dev_free(ca);
1366 	return -BCH_ERR_ENOMEM_dev_alloc;
1367 }
1368 
1369 static int __bch2_dev_attach_bdev(struct bch_dev *ca, struct bch_sb_handle *sb)
1370 {
1371 	unsigned ret;
1372 
1373 	if (bch2_dev_is_online(ca)) {
1374 		bch_err(ca, "already have device online in slot %u",
1375 			sb->sb->dev_idx);
1376 		return -BCH_ERR_device_already_online;
1377 	}
1378 
1379 	if (get_capacity(sb->bdev->bd_disk) <
1380 	    ca->mi.bucket_size * ca->mi.nbuckets) {
1381 		bch_err(ca, "cannot online: device too small");
1382 		return -BCH_ERR_device_size_too_small;
1383 	}
1384 
1385 	BUG_ON(!percpu_ref_is_zero(&ca->io_ref));
1386 
1387 	ret = bch2_dev_journal_init(ca, sb->sb);
1388 	if (ret)
1389 		return ret;
1390 
1391 	/* Commit: */
1392 	ca->disk_sb = *sb;
1393 	memset(sb, 0, sizeof(*sb));
1394 
1395 	ca->dev = ca->disk_sb.bdev->bd_dev;
1396 
1397 	percpu_ref_reinit(&ca->io_ref);
1398 
1399 	return 0;
1400 }
1401 
1402 static int bch2_dev_attach_bdev(struct bch_fs *c, struct bch_sb_handle *sb)
1403 {
1404 	struct bch_dev *ca;
1405 	int ret;
1406 
1407 	lockdep_assert_held(&c->state_lock);
1408 
1409 	if (le64_to_cpu(sb->sb->seq) >
1410 	    le64_to_cpu(c->disk_sb.sb->seq))
1411 		bch2_sb_to_fs(c, sb->sb);
1412 
1413 	BUG_ON(sb->sb->dev_idx >= c->sb.nr_devices ||
1414 	       !c->devs[sb->sb->dev_idx]);
1415 
1416 	ca = bch_dev_locked(c, sb->sb->dev_idx);
1417 
1418 	ret = __bch2_dev_attach_bdev(ca, sb);
1419 	if (ret)
1420 		return ret;
1421 
1422 	bch2_dev_sysfs_online(c, ca);
1423 
1424 	struct printbuf name = PRINTBUF;
1425 	prt_bdevname(&name, ca->disk_sb.bdev);
1426 
1427 	if (c->sb.nr_devices == 1)
1428 		strscpy(c->name, name.buf, sizeof(c->name));
1429 	strscpy(ca->name, name.buf, sizeof(ca->name));
1430 
1431 	printbuf_exit(&name);
1432 
1433 	rebalance_wakeup(c);
1434 	return 0;
1435 }
1436 
1437 /* Device management: */
1438 
1439 /*
1440  * Note: this function is also used by the error paths - when a particular
1441  * device sees an error, we call it to determine whether we can just set the
1442  * device RO, or - if this function returns false - we'll set the whole
1443  * filesystem RO:
1444  *
1445  * XXX: maybe we should be more explicit about whether we're changing state
1446  * because we got an error or what have you?
1447  */
1448 bool bch2_dev_state_allowed(struct bch_fs *c, struct bch_dev *ca,
1449 			    enum bch_member_state new_state, int flags)
1450 {
1451 	struct bch_devs_mask new_online_devs;
1452 	int nr_rw = 0, required;
1453 
1454 	lockdep_assert_held(&c->state_lock);
1455 
1456 	switch (new_state) {
1457 	case BCH_MEMBER_STATE_rw:
1458 		return true;
1459 	case BCH_MEMBER_STATE_ro:
1460 		if (ca->mi.state != BCH_MEMBER_STATE_rw)
1461 			return true;
1462 
1463 		/* do we have enough devices to write to?  */
1464 		for_each_member_device(c, ca2)
1465 			if (ca2 != ca)
1466 				nr_rw += ca2->mi.state == BCH_MEMBER_STATE_rw;
1467 
1468 		required = max(!(flags & BCH_FORCE_IF_METADATA_DEGRADED)
1469 			       ? c->opts.metadata_replicas
1470 			       : metadata_replicas_required(c),
1471 			       !(flags & BCH_FORCE_IF_DATA_DEGRADED)
1472 			       ? c->opts.data_replicas
1473 			       : data_replicas_required(c));
1474 
1475 		return nr_rw >= required;
1476 	case BCH_MEMBER_STATE_failed:
1477 	case BCH_MEMBER_STATE_spare:
1478 		if (ca->mi.state != BCH_MEMBER_STATE_rw &&
1479 		    ca->mi.state != BCH_MEMBER_STATE_ro)
1480 			return true;
1481 
1482 		/* do we have enough devices to read from?  */
1483 		new_online_devs = bch2_online_devs(c);
1484 		__clear_bit(ca->dev_idx, new_online_devs.d);
1485 
1486 		return bch2_have_enough_devs(c, new_online_devs, flags, false);
1487 	default:
1488 		BUG();
1489 	}
1490 }
1491 
1492 static bool bch2_fs_may_start(struct bch_fs *c)
1493 {
1494 	struct bch_dev *ca;
1495 	unsigned i, flags = 0;
1496 
1497 	if (c->opts.very_degraded)
1498 		flags |= BCH_FORCE_IF_DEGRADED|BCH_FORCE_IF_LOST;
1499 
1500 	if (c->opts.degraded)
1501 		flags |= BCH_FORCE_IF_DEGRADED;
1502 
1503 	if (!c->opts.degraded &&
1504 	    !c->opts.very_degraded) {
1505 		mutex_lock(&c->sb_lock);
1506 
1507 		for (i = 0; i < c->disk_sb.sb->nr_devices; i++) {
1508 			if (!bch2_dev_exists(c->disk_sb.sb, i))
1509 				continue;
1510 
1511 			ca = bch_dev_locked(c, i);
1512 
1513 			if (!bch2_dev_is_online(ca) &&
1514 			    (ca->mi.state == BCH_MEMBER_STATE_rw ||
1515 			     ca->mi.state == BCH_MEMBER_STATE_ro)) {
1516 				mutex_unlock(&c->sb_lock);
1517 				return false;
1518 			}
1519 		}
1520 		mutex_unlock(&c->sb_lock);
1521 	}
1522 
1523 	return bch2_have_enough_devs(c, bch2_online_devs(c), flags, true);
1524 }
1525 
1526 static void __bch2_dev_read_only(struct bch_fs *c, struct bch_dev *ca)
1527 {
1528 	/*
1529 	 * The allocator thread itself allocates btree nodes, so stop it first:
1530 	 */
1531 	bch2_dev_allocator_remove(c, ca);
1532 	bch2_dev_journal_stop(&c->journal, ca);
1533 }
1534 
1535 static void __bch2_dev_read_write(struct bch_fs *c, struct bch_dev *ca)
1536 {
1537 	lockdep_assert_held(&c->state_lock);
1538 
1539 	BUG_ON(ca->mi.state != BCH_MEMBER_STATE_rw);
1540 
1541 	bch2_dev_allocator_add(c, ca);
1542 	bch2_recalc_capacity(c);
1543 }
1544 
1545 int __bch2_dev_set_state(struct bch_fs *c, struct bch_dev *ca,
1546 			 enum bch_member_state new_state, int flags)
1547 {
1548 	struct bch_member *m;
1549 	int ret = 0;
1550 
1551 	if (ca->mi.state == new_state)
1552 		return 0;
1553 
1554 	if (!bch2_dev_state_allowed(c, ca, new_state, flags))
1555 		return -BCH_ERR_device_state_not_allowed;
1556 
1557 	if (new_state != BCH_MEMBER_STATE_rw)
1558 		__bch2_dev_read_only(c, ca);
1559 
1560 	bch_notice(ca, "%s", bch2_member_states[new_state]);
1561 
1562 	mutex_lock(&c->sb_lock);
1563 	m = bch2_members_v2_get_mut(c->disk_sb.sb, ca->dev_idx);
1564 	SET_BCH_MEMBER_STATE(m, new_state);
1565 	bch2_write_super(c);
1566 	mutex_unlock(&c->sb_lock);
1567 
1568 	if (new_state == BCH_MEMBER_STATE_rw)
1569 		__bch2_dev_read_write(c, ca);
1570 
1571 	rebalance_wakeup(c);
1572 
1573 	return ret;
1574 }
1575 
1576 int bch2_dev_set_state(struct bch_fs *c, struct bch_dev *ca,
1577 		       enum bch_member_state new_state, int flags)
1578 {
1579 	int ret;
1580 
1581 	down_write(&c->state_lock);
1582 	ret = __bch2_dev_set_state(c, ca, new_state, flags);
1583 	up_write(&c->state_lock);
1584 
1585 	return ret;
1586 }
1587 
1588 /* Device add/removal: */
1589 
1590 static int bch2_dev_remove_alloc(struct bch_fs *c, struct bch_dev *ca)
1591 {
1592 	struct bpos start	= POS(ca->dev_idx, 0);
1593 	struct bpos end		= POS(ca->dev_idx, U64_MAX);
1594 	int ret;
1595 
1596 	/*
1597 	 * We clear the LRU and need_discard btrees first so that we don't race
1598 	 * with bch2_do_invalidates() and bch2_do_discards()
1599 	 */
1600 	ret =   bch2_btree_delete_range(c, BTREE_ID_lru, start, end,
1601 					BTREE_TRIGGER_NORUN, NULL) ?:
1602 		bch2_btree_delete_range(c, BTREE_ID_need_discard, start, end,
1603 					BTREE_TRIGGER_NORUN, NULL) ?:
1604 		bch2_btree_delete_range(c, BTREE_ID_freespace, start, end,
1605 					BTREE_TRIGGER_NORUN, NULL) ?:
1606 		bch2_btree_delete_range(c, BTREE_ID_backpointers, start, end,
1607 					BTREE_TRIGGER_NORUN, NULL) ?:
1608 		bch2_btree_delete_range(c, BTREE_ID_alloc, start, end,
1609 					BTREE_TRIGGER_NORUN, NULL) ?:
1610 		bch2_btree_delete_range(c, BTREE_ID_bucket_gens, start, end,
1611 					BTREE_TRIGGER_NORUN, NULL);
1612 	bch_err_msg(c, ret, "removing dev alloc info");
1613 	return ret;
1614 }
1615 
1616 int bch2_dev_remove(struct bch_fs *c, struct bch_dev *ca, int flags)
1617 {
1618 	struct bch_member *m;
1619 	unsigned dev_idx = ca->dev_idx, data;
1620 	int ret;
1621 
1622 	down_write(&c->state_lock);
1623 
1624 	/*
1625 	 * We consume a reference to ca->ref, regardless of whether we succeed
1626 	 * or fail:
1627 	 */
1628 	percpu_ref_put(&ca->ref);
1629 
1630 	if (!bch2_dev_state_allowed(c, ca, BCH_MEMBER_STATE_failed, flags)) {
1631 		bch_err(ca, "Cannot remove without losing data");
1632 		ret = -BCH_ERR_device_state_not_allowed;
1633 		goto err;
1634 	}
1635 
1636 	__bch2_dev_read_only(c, ca);
1637 
1638 	ret = bch2_dev_data_drop(c, ca->dev_idx, flags);
1639 	bch_err_msg(ca, ret, "bch2_dev_data_drop()");
1640 	if (ret)
1641 		goto err;
1642 
1643 	ret = bch2_dev_remove_alloc(c, ca);
1644 	bch_err_msg(ca, ret, "bch2_dev_remove_alloc()");
1645 	if (ret)
1646 		goto err;
1647 
1648 	ret = bch2_journal_flush_device_pins(&c->journal, ca->dev_idx);
1649 	bch_err_msg(ca, ret, "bch2_journal_flush_device_pins()");
1650 	if (ret)
1651 		goto err;
1652 
1653 	ret = bch2_journal_flush(&c->journal);
1654 	bch_err_msg(ca, ret, "bch2_journal_flush()");
1655 	if (ret)
1656 		goto err;
1657 
1658 	ret = bch2_replicas_gc2(c);
1659 	bch_err_msg(ca, ret, "bch2_replicas_gc2()");
1660 	if (ret)
1661 		goto err;
1662 
1663 	data = bch2_dev_has_data(c, ca);
1664 	if (data) {
1665 		struct printbuf data_has = PRINTBUF;
1666 
1667 		prt_bitflags(&data_has, __bch2_data_types, data);
1668 		bch_err(ca, "Remove failed, still has data (%s)", data_has.buf);
1669 		printbuf_exit(&data_has);
1670 		ret = -EBUSY;
1671 		goto err;
1672 	}
1673 
1674 	__bch2_dev_offline(c, ca);
1675 
1676 	mutex_lock(&c->sb_lock);
1677 	rcu_assign_pointer(c->devs[ca->dev_idx], NULL);
1678 	mutex_unlock(&c->sb_lock);
1679 
1680 	percpu_ref_kill(&ca->ref);
1681 	wait_for_completion(&ca->ref_completion);
1682 
1683 	bch2_dev_free(ca);
1684 
1685 	/*
1686 	 * At this point the device object has been removed in-core, but the
1687 	 * on-disk journal might still refer to the device index via sb device
1688 	 * usage entries. Recovery fails if it sees usage information for an
1689 	 * invalid device. Flush journal pins to push the back of the journal
1690 	 * past now invalid device index references before we update the
1691 	 * superblock, but after the device object has been removed so any
1692 	 * further journal writes elide usage info for the device.
1693 	 */
1694 	bch2_journal_flush_all_pins(&c->journal);
1695 
1696 	/*
1697 	 * Free this device's slot in the bch_member array - all pointers to
1698 	 * this device must be gone:
1699 	 */
1700 	mutex_lock(&c->sb_lock);
1701 	m = bch2_members_v2_get_mut(c->disk_sb.sb, dev_idx);
1702 	memset(&m->uuid, 0, sizeof(m->uuid));
1703 
1704 	bch2_write_super(c);
1705 
1706 	mutex_unlock(&c->sb_lock);
1707 	up_write(&c->state_lock);
1708 
1709 	bch2_dev_usage_journal_reserve(c);
1710 	return 0;
1711 err:
1712 	if (ca->mi.state == BCH_MEMBER_STATE_rw &&
1713 	    !percpu_ref_is_zero(&ca->io_ref))
1714 		__bch2_dev_read_write(c, ca);
1715 	up_write(&c->state_lock);
1716 	return ret;
1717 }
1718 
1719 /* Add new device to running filesystem: */
1720 int bch2_dev_add(struct bch_fs *c, const char *path)
1721 {
1722 	struct bch_opts opts = bch2_opts_empty();
1723 	struct bch_sb_handle sb;
1724 	struct bch_dev *ca = NULL;
1725 	struct bch_sb_field_members_v2 *mi;
1726 	struct bch_member dev_mi;
1727 	unsigned dev_idx, nr_devices, u64s;
1728 	struct printbuf errbuf = PRINTBUF;
1729 	struct printbuf label = PRINTBUF;
1730 	int ret;
1731 
1732 	ret = bch2_read_super(path, &opts, &sb);
1733 	bch_err_msg(c, ret, "reading super");
1734 	if (ret)
1735 		goto err;
1736 
1737 	dev_mi = bch2_sb_member_get(sb.sb, sb.sb->dev_idx);
1738 
1739 	if (BCH_MEMBER_GROUP(&dev_mi)) {
1740 		bch2_disk_path_to_text_sb(&label, sb.sb, BCH_MEMBER_GROUP(&dev_mi) - 1);
1741 		if (label.allocation_failure) {
1742 			ret = -ENOMEM;
1743 			goto err;
1744 		}
1745 	}
1746 
1747 	ret = bch2_dev_may_add(sb.sb, c);
1748 	if (ret)
1749 		goto err;
1750 
1751 	ca = __bch2_dev_alloc(c, &dev_mi);
1752 	if (!ca) {
1753 		ret = -ENOMEM;
1754 		goto err;
1755 	}
1756 
1757 	bch2_dev_usage_init(ca);
1758 
1759 	ret = __bch2_dev_attach_bdev(ca, &sb);
1760 	if (ret)
1761 		goto err;
1762 
1763 	ret = bch2_dev_journal_alloc(ca);
1764 	bch_err_msg(c, ret, "allocating journal");
1765 	if (ret)
1766 		goto err;
1767 
1768 	down_write(&c->state_lock);
1769 	mutex_lock(&c->sb_lock);
1770 
1771 	ret = bch2_sb_from_fs(c, ca);
1772 	bch_err_msg(c, ret, "setting up new superblock");
1773 	if (ret)
1774 		goto err_unlock;
1775 
1776 	if (dynamic_fault("bcachefs:add:no_slot"))
1777 		goto no_slot;
1778 
1779 	for (dev_idx = 0; dev_idx < BCH_SB_MEMBERS_MAX; dev_idx++)
1780 		if (!bch2_dev_exists(c->disk_sb.sb, dev_idx))
1781 			goto have_slot;
1782 no_slot:
1783 	ret = -BCH_ERR_ENOSPC_sb_members;
1784 	bch_err_msg(c, ret, "setting up new superblock");
1785 	goto err_unlock;
1786 
1787 have_slot:
1788 	nr_devices = max_t(unsigned, dev_idx + 1, c->sb.nr_devices);
1789 
1790 	mi = bch2_sb_field_get(c->disk_sb.sb, members_v2);
1791 	u64s = DIV_ROUND_UP(sizeof(struct bch_sb_field_members_v2) +
1792 			    le16_to_cpu(mi->member_bytes) * nr_devices, sizeof(u64));
1793 
1794 	mi = bch2_sb_field_resize(&c->disk_sb, members_v2, u64s);
1795 	if (!mi) {
1796 		ret = -BCH_ERR_ENOSPC_sb_members;
1797 		bch_err_msg(c, ret, "setting up new superblock");
1798 		goto err_unlock;
1799 	}
1800 	struct bch_member *m = bch2_members_v2_get_mut(c->disk_sb.sb, dev_idx);
1801 
1802 	/* success: */
1803 
1804 	*m = dev_mi;
1805 	m->last_mount = cpu_to_le64(ktime_get_real_seconds());
1806 	c->disk_sb.sb->nr_devices	= nr_devices;
1807 
1808 	ca->disk_sb.sb->dev_idx	= dev_idx;
1809 	bch2_dev_attach(c, ca, dev_idx);
1810 
1811 	if (BCH_MEMBER_GROUP(&dev_mi)) {
1812 		ret = __bch2_dev_group_set(c, ca, label.buf);
1813 		bch_err_msg(c, ret, "creating new label");
1814 		if (ret)
1815 			goto err_unlock;
1816 	}
1817 
1818 	bch2_write_super(c);
1819 	mutex_unlock(&c->sb_lock);
1820 
1821 	bch2_dev_usage_journal_reserve(c);
1822 
1823 	ret = bch2_trans_mark_dev_sb(c, ca);
1824 	bch_err_msg(ca, ret, "marking new superblock");
1825 	if (ret)
1826 		goto err_late;
1827 
1828 	ret = bch2_fs_freespace_init(c);
1829 	bch_err_msg(ca, ret, "initializing free space");
1830 	if (ret)
1831 		goto err_late;
1832 
1833 	ca->new_fs_bucket_idx = 0;
1834 
1835 	if (ca->mi.state == BCH_MEMBER_STATE_rw)
1836 		__bch2_dev_read_write(c, ca);
1837 
1838 	up_write(&c->state_lock);
1839 	return 0;
1840 
1841 err_unlock:
1842 	mutex_unlock(&c->sb_lock);
1843 	up_write(&c->state_lock);
1844 err:
1845 	if (ca)
1846 		bch2_dev_free(ca);
1847 	bch2_free_super(&sb);
1848 	printbuf_exit(&label);
1849 	printbuf_exit(&errbuf);
1850 	bch_err_fn(c, ret);
1851 	return ret;
1852 err_late:
1853 	up_write(&c->state_lock);
1854 	ca = NULL;
1855 	goto err;
1856 }
1857 
1858 /* Hot add existing device to running filesystem: */
1859 int bch2_dev_online(struct bch_fs *c, const char *path)
1860 {
1861 	struct bch_opts opts = bch2_opts_empty();
1862 	struct bch_sb_handle sb = { NULL };
1863 	struct bch_dev *ca;
1864 	unsigned dev_idx;
1865 	int ret;
1866 
1867 	down_write(&c->state_lock);
1868 
1869 	ret = bch2_read_super(path, &opts, &sb);
1870 	if (ret) {
1871 		up_write(&c->state_lock);
1872 		return ret;
1873 	}
1874 
1875 	dev_idx = sb.sb->dev_idx;
1876 
1877 	ret = bch2_dev_in_fs(&c->disk_sb, &sb, &c->opts);
1878 	bch_err_msg(c, ret, "bringing %s online", path);
1879 	if (ret)
1880 		goto err;
1881 
1882 	ret = bch2_dev_attach_bdev(c, &sb);
1883 	if (ret)
1884 		goto err;
1885 
1886 	ca = bch_dev_locked(c, dev_idx);
1887 
1888 	ret = bch2_trans_mark_dev_sb(c, ca);
1889 	bch_err_msg(c, ret, "bringing %s online: error from bch2_trans_mark_dev_sb", path);
1890 	if (ret)
1891 		goto err;
1892 
1893 	if (ca->mi.state == BCH_MEMBER_STATE_rw)
1894 		__bch2_dev_read_write(c, ca);
1895 
1896 	if (!ca->mi.freespace_initialized) {
1897 		ret = bch2_dev_freespace_init(c, ca, 0, ca->mi.nbuckets);
1898 		bch_err_msg(ca, ret, "initializing free space");
1899 		if (ret)
1900 			goto err;
1901 	}
1902 
1903 	if (!ca->journal.nr) {
1904 		ret = bch2_dev_journal_alloc(ca);
1905 		bch_err_msg(ca, ret, "allocating journal");
1906 		if (ret)
1907 			goto err;
1908 	}
1909 
1910 	mutex_lock(&c->sb_lock);
1911 	bch2_members_v2_get_mut(c->disk_sb.sb, ca->dev_idx)->last_mount =
1912 		cpu_to_le64(ktime_get_real_seconds());
1913 	bch2_write_super(c);
1914 	mutex_unlock(&c->sb_lock);
1915 
1916 	up_write(&c->state_lock);
1917 	return 0;
1918 err:
1919 	up_write(&c->state_lock);
1920 	bch2_free_super(&sb);
1921 	return ret;
1922 }
1923 
1924 int bch2_dev_offline(struct bch_fs *c, struct bch_dev *ca, int flags)
1925 {
1926 	down_write(&c->state_lock);
1927 
1928 	if (!bch2_dev_is_online(ca)) {
1929 		bch_err(ca, "Already offline");
1930 		up_write(&c->state_lock);
1931 		return 0;
1932 	}
1933 
1934 	if (!bch2_dev_state_allowed(c, ca, BCH_MEMBER_STATE_failed, flags)) {
1935 		bch_err(ca, "Cannot offline required disk");
1936 		up_write(&c->state_lock);
1937 		return -BCH_ERR_device_state_not_allowed;
1938 	}
1939 
1940 	__bch2_dev_offline(c, ca);
1941 
1942 	up_write(&c->state_lock);
1943 	return 0;
1944 }
1945 
1946 int bch2_dev_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
1947 {
1948 	struct bch_member *m;
1949 	u64 old_nbuckets;
1950 	int ret = 0;
1951 
1952 	down_write(&c->state_lock);
1953 	old_nbuckets = ca->mi.nbuckets;
1954 
1955 	if (nbuckets < ca->mi.nbuckets) {
1956 		bch_err(ca, "Cannot shrink yet");
1957 		ret = -EINVAL;
1958 		goto err;
1959 	}
1960 
1961 	if (bch2_dev_is_online(ca) &&
1962 	    get_capacity(ca->disk_sb.bdev->bd_disk) <
1963 	    ca->mi.bucket_size * nbuckets) {
1964 		bch_err(ca, "New size larger than device");
1965 		ret = -BCH_ERR_device_size_too_small;
1966 		goto err;
1967 	}
1968 
1969 	ret = bch2_dev_buckets_resize(c, ca, nbuckets);
1970 	bch_err_msg(ca, ret, "resizing buckets");
1971 	if (ret)
1972 		goto err;
1973 
1974 	ret = bch2_trans_mark_dev_sb(c, ca);
1975 	if (ret)
1976 		goto err;
1977 
1978 	mutex_lock(&c->sb_lock);
1979 	m = bch2_members_v2_get_mut(c->disk_sb.sb, ca->dev_idx);
1980 	m->nbuckets = cpu_to_le64(nbuckets);
1981 
1982 	bch2_write_super(c);
1983 	mutex_unlock(&c->sb_lock);
1984 
1985 	if (ca->mi.freespace_initialized) {
1986 		ret = bch2_dev_freespace_init(c, ca, old_nbuckets, nbuckets);
1987 		if (ret)
1988 			goto err;
1989 
1990 		/*
1991 		 * XXX: this is all wrong transactionally - we'll be able to do
1992 		 * this correctly after the disk space accounting rewrite
1993 		 */
1994 		ca->usage_base->d[BCH_DATA_free].buckets += nbuckets - old_nbuckets;
1995 	}
1996 
1997 	bch2_recalc_capacity(c);
1998 err:
1999 	up_write(&c->state_lock);
2000 	return ret;
2001 }
2002 
2003 /* return with ref on ca->ref: */
2004 struct bch_dev *bch2_dev_lookup(struct bch_fs *c, const char *name)
2005 {
2006 	rcu_read_lock();
2007 	for_each_member_device_rcu(c, ca, NULL)
2008 		if (!strcmp(name, ca->name)) {
2009 			rcu_read_unlock();
2010 			return ca;
2011 		}
2012 	rcu_read_unlock();
2013 	return ERR_PTR(-BCH_ERR_ENOENT_dev_not_found);
2014 }
2015 
2016 /* Filesystem open: */
2017 
2018 static inline int sb_cmp(struct bch_sb *l, struct bch_sb *r)
2019 {
2020 	return  cmp_int(le64_to_cpu(l->seq), le64_to_cpu(r->seq)) ?:
2021 		cmp_int(le64_to_cpu(l->write_time), le64_to_cpu(r->write_time));
2022 }
2023 
2024 struct bch_fs *bch2_fs_open(char * const *devices, unsigned nr_devices,
2025 			    struct bch_opts opts)
2026 {
2027 	DARRAY(struct bch_sb_handle) sbs = { 0 };
2028 	struct bch_fs *c = NULL;
2029 	struct bch_sb_handle *best = NULL;
2030 	struct printbuf errbuf = PRINTBUF;
2031 	int ret = 0;
2032 
2033 	if (!try_module_get(THIS_MODULE))
2034 		return ERR_PTR(-ENODEV);
2035 
2036 	if (!nr_devices) {
2037 		ret = -EINVAL;
2038 		goto err;
2039 	}
2040 
2041 	ret = darray_make_room(&sbs, nr_devices);
2042 	if (ret)
2043 		goto err;
2044 
2045 	for (unsigned i = 0; i < nr_devices; i++) {
2046 		struct bch_sb_handle sb = { NULL };
2047 
2048 		ret = bch2_read_super(devices[i], &opts, &sb);
2049 		if (ret)
2050 			goto err;
2051 
2052 		BUG_ON(darray_push(&sbs, sb));
2053 	}
2054 
2055 	if (opts.nochanges && !opts.read_only) {
2056 		ret = -BCH_ERR_erofs_nochanges;
2057 		goto err_print;
2058 	}
2059 
2060 	darray_for_each(sbs, sb)
2061 		if (!best || sb_cmp(sb->sb, best->sb) > 0)
2062 			best = sb;
2063 
2064 	darray_for_each_reverse(sbs, sb) {
2065 		ret = bch2_dev_in_fs(best, sb, &opts);
2066 
2067 		if (ret == -BCH_ERR_device_has_been_removed ||
2068 		    ret == -BCH_ERR_device_splitbrain) {
2069 			bch2_free_super(sb);
2070 			darray_remove_item(&sbs, sb);
2071 			best -= best > sb;
2072 			ret = 0;
2073 			continue;
2074 		}
2075 
2076 		if (ret)
2077 			goto err_print;
2078 	}
2079 
2080 	c = bch2_fs_alloc(best->sb, opts);
2081 	ret = PTR_ERR_OR_ZERO(c);
2082 	if (ret)
2083 		goto err;
2084 
2085 	down_write(&c->state_lock);
2086 	darray_for_each(sbs, sb) {
2087 		ret = bch2_dev_attach_bdev(c, sb);
2088 		if (ret) {
2089 			up_write(&c->state_lock);
2090 			goto err;
2091 		}
2092 	}
2093 	up_write(&c->state_lock);
2094 
2095 	if (!bch2_fs_may_start(c)) {
2096 		ret = -BCH_ERR_insufficient_devices_to_start;
2097 		goto err_print;
2098 	}
2099 
2100 	if (!c->opts.nostart) {
2101 		ret = bch2_fs_start(c);
2102 		if (ret)
2103 			goto err;
2104 	}
2105 out:
2106 	darray_for_each(sbs, sb)
2107 		bch2_free_super(sb);
2108 	darray_exit(&sbs);
2109 	printbuf_exit(&errbuf);
2110 	module_put(THIS_MODULE);
2111 	return c;
2112 err_print:
2113 	pr_err("bch_fs_open err opening %s: %s",
2114 	       devices[0], bch2_err_str(ret));
2115 err:
2116 	if (!IS_ERR_OR_NULL(c))
2117 		bch2_fs_stop(c);
2118 	c = ERR_PTR(ret);
2119 	goto out;
2120 }
2121 
2122 /* Global interfaces/init */
2123 
2124 static void bcachefs_exit(void)
2125 {
2126 	bch2_debug_exit();
2127 	bch2_vfs_exit();
2128 	bch2_chardev_exit();
2129 	bch2_btree_key_cache_exit();
2130 	if (bcachefs_kset)
2131 		kset_unregister(bcachefs_kset);
2132 }
2133 
2134 static int __init bcachefs_init(void)
2135 {
2136 	bch2_bkey_pack_test();
2137 
2138 	if (!(bcachefs_kset = kset_create_and_add("bcachefs", NULL, fs_kobj)) ||
2139 	    bch2_btree_key_cache_init() ||
2140 	    bch2_chardev_init() ||
2141 	    bch2_vfs_init() ||
2142 	    bch2_debug_init())
2143 		goto err;
2144 
2145 	return 0;
2146 err:
2147 	bcachefs_exit();
2148 	return -ENOMEM;
2149 }
2150 
2151 #define BCH_DEBUG_PARAM(name, description)			\
2152 	bool bch2_##name;					\
2153 	module_param_named(name, bch2_##name, bool, 0644);	\
2154 	MODULE_PARM_DESC(name, description);
2155 BCH_DEBUG_PARAMS()
2156 #undef BCH_DEBUG_PARAM
2157 
2158 __maybe_unused
2159 static unsigned bch2_metadata_version = bcachefs_metadata_version_current;
2160 module_param_named(version, bch2_metadata_version, uint, 0400);
2161 
2162 module_exit(bcachefs_exit);
2163 module_init(bcachefs_init);
2164