xref: /linux/fs/bcachefs/super.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
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_accounting.h"
29 #include "disk_groups.h"
30 #include "ec.h"
31 #include "errcode.h"
32 #include "error.h"
33 #include "fs.h"
34 #include "fs-io.h"
35 #include "fs-io-buffered.h"
36 #include "fs-io-direct.h"
37 #include "fsck.h"
38 #include "inode.h"
39 #include "io_read.h"
40 #include "io_write.h"
41 #include "journal.h"
42 #include "journal_reclaim.h"
43 #include "journal_seq_blacklist.h"
44 #include "move.h"
45 #include "migrate.h"
46 #include "movinggc.h"
47 #include "nocow_locking.h"
48 #include "quota.h"
49 #include "rebalance.h"
50 #include "recovery.h"
51 #include "replicas.h"
52 #include "sb-clean.h"
53 #include "sb-counters.h"
54 #include "sb-errors.h"
55 #include "sb-members.h"
56 #include "snapshot.h"
57 #include "subvolume.h"
58 #include "super.h"
59 #include "super-io.h"
60 #include "sysfs.h"
61 #include "thread_with_file.h"
62 #include "trace.h"
63 
64 #include <linux/backing-dev.h>
65 #include <linux/blkdev.h>
66 #include <linux/debugfs.h>
67 #include <linux/device.h>
68 #include <linux/idr.h>
69 #include <linux/module.h>
70 #include <linux/percpu.h>
71 #include <linux/random.h>
72 #include <linux/sysfs.h>
73 #include <crypto/hash.h>
74 
75 MODULE_LICENSE("GPL");
76 MODULE_AUTHOR("Kent Overstreet <kent.overstreet@gmail.com>");
77 MODULE_DESCRIPTION("bcachefs filesystem");
78 MODULE_SOFTDEP("pre: crc32c");
79 MODULE_SOFTDEP("pre: crc64");
80 MODULE_SOFTDEP("pre: sha256");
81 MODULE_SOFTDEP("pre: chacha20");
82 MODULE_SOFTDEP("pre: poly1305");
83 MODULE_SOFTDEP("pre: xxhash");
84 
85 const char * const bch2_fs_flag_strs[] = {
86 #define x(n)		#n,
87 	BCH_FS_FLAGS()
88 #undef x
89 	NULL
90 };
91 
92 void bch2_print_str(struct bch_fs *c, const char *str)
93 {
94 #ifdef __KERNEL__
95 	struct stdio_redirect *stdio = bch2_fs_stdio_redirect(c);
96 
97 	if (unlikely(stdio)) {
98 		bch2_stdio_redirect_printf(stdio, true, "%s", str);
99 		return;
100 	}
101 #endif
102 	bch2_print_string_as_lines(KERN_ERR, str);
103 }
104 
105 __printf(2, 0)
106 static void bch2_print_maybe_redirect(struct stdio_redirect *stdio, const char *fmt, va_list args)
107 {
108 #ifdef __KERNEL__
109 	if (unlikely(stdio)) {
110 		if (fmt[0] == KERN_SOH[0])
111 			fmt += 2;
112 
113 		bch2_stdio_redirect_vprintf(stdio, true, fmt, args);
114 		return;
115 	}
116 #endif
117 	vprintk(fmt, args);
118 }
119 
120 void bch2_print_opts(struct bch_opts *opts, const char *fmt, ...)
121 {
122 	struct stdio_redirect *stdio = (void *)(unsigned long)opts->stdio;
123 
124 	va_list args;
125 	va_start(args, fmt);
126 	bch2_print_maybe_redirect(stdio, fmt, args);
127 	va_end(args);
128 }
129 
130 void __bch2_print(struct bch_fs *c, const char *fmt, ...)
131 {
132 	struct stdio_redirect *stdio = bch2_fs_stdio_redirect(c);
133 
134 	va_list args;
135 	va_start(args, fmt);
136 	bch2_print_maybe_redirect(stdio, fmt, args);
137 	va_end(args);
138 }
139 
140 #define KTYPE(type)							\
141 static const struct attribute_group type ## _group = {			\
142 	.attrs = type ## _files						\
143 };									\
144 									\
145 static const struct attribute_group *type ## _groups[] = {		\
146 	&type ## _group,						\
147 	NULL								\
148 };									\
149 									\
150 static const struct kobj_type type ## _ktype = {			\
151 	.release	= type ## _release,				\
152 	.sysfs_ops	= &type ## _sysfs_ops,				\
153 	.default_groups = type ## _groups				\
154 }
155 
156 static void bch2_fs_release(struct kobject *);
157 static void bch2_dev_release(struct kobject *);
158 static void bch2_fs_counters_release(struct kobject *k)
159 {
160 }
161 
162 static void bch2_fs_internal_release(struct kobject *k)
163 {
164 }
165 
166 static void bch2_fs_opts_dir_release(struct kobject *k)
167 {
168 }
169 
170 static void bch2_fs_time_stats_release(struct kobject *k)
171 {
172 }
173 
174 KTYPE(bch2_fs);
175 KTYPE(bch2_fs_counters);
176 KTYPE(bch2_fs_internal);
177 KTYPE(bch2_fs_opts_dir);
178 KTYPE(bch2_fs_time_stats);
179 KTYPE(bch2_dev);
180 
181 static struct kset *bcachefs_kset;
182 static LIST_HEAD(bch_fs_list);
183 static DEFINE_MUTEX(bch_fs_list_lock);
184 
185 DECLARE_WAIT_QUEUE_HEAD(bch2_read_only_wait);
186 
187 static void bch2_dev_unlink(struct bch_dev *);
188 static void bch2_dev_free(struct bch_dev *);
189 static int bch2_dev_alloc(struct bch_fs *, unsigned);
190 static int bch2_dev_sysfs_online(struct bch_fs *, struct bch_dev *);
191 static void __bch2_dev_read_only(struct bch_fs *, struct bch_dev *);
192 
193 struct bch_fs *bch2_dev_to_fs(dev_t dev)
194 {
195 	struct bch_fs *c;
196 
197 	mutex_lock(&bch_fs_list_lock);
198 	rcu_read_lock();
199 
200 	list_for_each_entry(c, &bch_fs_list, list)
201 		for_each_member_device_rcu(c, ca, NULL)
202 			if (ca->disk_sb.bdev && ca->disk_sb.bdev->bd_dev == dev) {
203 				closure_get(&c->cl);
204 				goto found;
205 			}
206 	c = NULL;
207 found:
208 	rcu_read_unlock();
209 	mutex_unlock(&bch_fs_list_lock);
210 
211 	return c;
212 }
213 
214 static struct bch_fs *__bch2_uuid_to_fs(__uuid_t uuid)
215 {
216 	struct bch_fs *c;
217 
218 	lockdep_assert_held(&bch_fs_list_lock);
219 
220 	list_for_each_entry(c, &bch_fs_list, list)
221 		if (!memcmp(&c->disk_sb.sb->uuid, &uuid, sizeof(uuid)))
222 			return c;
223 
224 	return NULL;
225 }
226 
227 struct bch_fs *bch2_uuid_to_fs(__uuid_t uuid)
228 {
229 	struct bch_fs *c;
230 
231 	mutex_lock(&bch_fs_list_lock);
232 	c = __bch2_uuid_to_fs(uuid);
233 	if (c)
234 		closure_get(&c->cl);
235 	mutex_unlock(&bch_fs_list_lock);
236 
237 	return c;
238 }
239 
240 /* Filesystem RO/RW: */
241 
242 /*
243  * For startup/shutdown of RW stuff, the dependencies are:
244  *
245  * - foreground writes depend on copygc and rebalance (to free up space)
246  *
247  * - copygc and rebalance depend on mark and sweep gc (they actually probably
248  *   don't because they either reserve ahead of time or don't block if
249  *   allocations fail, but allocations can require mark and sweep gc to run
250  *   because of generation number wraparound)
251  *
252  * - all of the above depends on the allocator threads
253  *
254  * - allocator depends on the journal (when it rewrites prios and gens)
255  */
256 
257 static void __bch2_fs_read_only(struct bch_fs *c)
258 {
259 	unsigned clean_passes = 0;
260 	u64 seq = 0;
261 
262 	bch2_fs_ec_stop(c);
263 	bch2_open_buckets_stop(c, NULL, true);
264 	bch2_rebalance_stop(c);
265 	bch2_copygc_stop(c);
266 	bch2_fs_ec_flush(c);
267 
268 	bch_verbose(c, "flushing journal and stopping allocators, journal seq %llu",
269 		    journal_cur_seq(&c->journal));
270 
271 	do {
272 		clean_passes++;
273 
274 		if (bch2_btree_interior_updates_flush(c) ||
275 		    bch2_journal_flush_all_pins(&c->journal) ||
276 		    bch2_btree_flush_all_writes(c) ||
277 		    seq != atomic64_read(&c->journal.seq)) {
278 			seq = atomic64_read(&c->journal.seq);
279 			clean_passes = 0;
280 		}
281 	} while (clean_passes < 2);
282 
283 	bch_verbose(c, "flushing journal and stopping allocators complete, journal seq %llu",
284 		    journal_cur_seq(&c->journal));
285 
286 	if (test_bit(JOURNAL_replay_done, &c->journal.flags) &&
287 	    !test_bit(BCH_FS_emergency_ro, &c->flags))
288 		set_bit(BCH_FS_clean_shutdown, &c->flags);
289 
290 	bch2_fs_journal_stop(&c->journal);
291 
292 	bch_info(c, "%sshutdown complete, journal seq %llu",
293 		 test_bit(BCH_FS_clean_shutdown, &c->flags) ? "" : "un",
294 		 c->journal.seq_ondisk);
295 
296 	/*
297 	 * After stopping journal:
298 	 */
299 	for_each_member_device(c, ca)
300 		bch2_dev_allocator_remove(c, ca);
301 }
302 
303 #ifndef BCH_WRITE_REF_DEBUG
304 static void bch2_writes_disabled(struct percpu_ref *writes)
305 {
306 	struct bch_fs *c = container_of(writes, struct bch_fs, writes);
307 
308 	set_bit(BCH_FS_write_disable_complete, &c->flags);
309 	wake_up(&bch2_read_only_wait);
310 }
311 #endif
312 
313 void bch2_fs_read_only(struct bch_fs *c)
314 {
315 	if (!test_bit(BCH_FS_rw, &c->flags)) {
316 		bch2_journal_reclaim_stop(&c->journal);
317 		return;
318 	}
319 
320 	BUG_ON(test_bit(BCH_FS_write_disable_complete, &c->flags));
321 
322 	bch_verbose(c, "going read-only");
323 
324 	/*
325 	 * Block new foreground-end write operations from starting - any new
326 	 * writes will return -EROFS:
327 	 */
328 	set_bit(BCH_FS_going_ro, &c->flags);
329 #ifndef BCH_WRITE_REF_DEBUG
330 	percpu_ref_kill(&c->writes);
331 #else
332 	for (unsigned i = 0; i < BCH_WRITE_REF_NR; i++)
333 		bch2_write_ref_put(c, i);
334 #endif
335 
336 	/*
337 	 * If we're not doing an emergency shutdown, we want to wait on
338 	 * outstanding writes to complete so they don't see spurious errors due
339 	 * to shutting down the allocator:
340 	 *
341 	 * If we are doing an emergency shutdown outstanding writes may
342 	 * hang until we shutdown the allocator so we don't want to wait
343 	 * on outstanding writes before shutting everything down - but
344 	 * we do need to wait on them before returning and signalling
345 	 * that going RO is complete:
346 	 */
347 	wait_event(bch2_read_only_wait,
348 		   test_bit(BCH_FS_write_disable_complete, &c->flags) ||
349 		   test_bit(BCH_FS_emergency_ro, &c->flags));
350 
351 	bool writes_disabled = test_bit(BCH_FS_write_disable_complete, &c->flags);
352 	if (writes_disabled)
353 		bch_verbose(c, "finished waiting for writes to stop");
354 
355 	__bch2_fs_read_only(c);
356 
357 	wait_event(bch2_read_only_wait,
358 		   test_bit(BCH_FS_write_disable_complete, &c->flags));
359 
360 	if (!writes_disabled)
361 		bch_verbose(c, "finished waiting for writes to stop");
362 
363 	clear_bit(BCH_FS_write_disable_complete, &c->flags);
364 	clear_bit(BCH_FS_going_ro, &c->flags);
365 	clear_bit(BCH_FS_rw, &c->flags);
366 
367 	if (!bch2_journal_error(&c->journal) &&
368 	    !test_bit(BCH_FS_error, &c->flags) &&
369 	    !test_bit(BCH_FS_emergency_ro, &c->flags) &&
370 	    test_bit(BCH_FS_started, &c->flags) &&
371 	    test_bit(BCH_FS_clean_shutdown, &c->flags) &&
372 	    c->recovery_pass_done >= BCH_RECOVERY_PASS_journal_replay) {
373 		BUG_ON(c->journal.last_empty_seq != journal_cur_seq(&c->journal));
374 		BUG_ON(atomic_long_read(&c->btree_cache.nr_dirty));
375 		BUG_ON(atomic_long_read(&c->btree_key_cache.nr_dirty));
376 		BUG_ON(c->btree_write_buffer.inc.keys.nr);
377 		BUG_ON(c->btree_write_buffer.flushing.keys.nr);
378 		bch2_verify_accounting_clean(c);
379 
380 		bch_verbose(c, "marking filesystem clean");
381 		bch2_fs_mark_clean(c);
382 	} else {
383 		bch_verbose(c, "done going read-only, filesystem not clean");
384 	}
385 }
386 
387 static void bch2_fs_read_only_work(struct work_struct *work)
388 {
389 	struct bch_fs *c =
390 		container_of(work, struct bch_fs, read_only_work);
391 
392 	down_write(&c->state_lock);
393 	bch2_fs_read_only(c);
394 	up_write(&c->state_lock);
395 }
396 
397 static void bch2_fs_read_only_async(struct bch_fs *c)
398 {
399 	queue_work(system_long_wq, &c->read_only_work);
400 }
401 
402 bool bch2_fs_emergency_read_only(struct bch_fs *c)
403 {
404 	bool ret = !test_and_set_bit(BCH_FS_emergency_ro, &c->flags);
405 
406 	bch2_journal_halt(&c->journal);
407 	bch2_fs_read_only_async(c);
408 
409 	wake_up(&bch2_read_only_wait);
410 	return ret;
411 }
412 
413 static int bch2_fs_read_write_late(struct bch_fs *c)
414 {
415 	int ret;
416 
417 	/*
418 	 * Data move operations can't run until after check_snapshots has
419 	 * completed, and bch2_snapshot_is_ancestor() is available.
420 	 *
421 	 * Ideally we'd start copygc/rebalance earlier instead of waiting for
422 	 * all of recovery/fsck to complete:
423 	 */
424 	ret = bch2_copygc_start(c);
425 	if (ret) {
426 		bch_err(c, "error starting copygc thread");
427 		return ret;
428 	}
429 
430 	ret = bch2_rebalance_start(c);
431 	if (ret) {
432 		bch_err(c, "error starting rebalance thread");
433 		return ret;
434 	}
435 
436 	return 0;
437 }
438 
439 static int __bch2_fs_read_write(struct bch_fs *c, bool early)
440 {
441 	int ret;
442 
443 	if (test_bit(BCH_FS_initial_gc_unfixed, &c->flags)) {
444 		bch_err(c, "cannot go rw, unfixed btree errors");
445 		return -BCH_ERR_erofs_unfixed_errors;
446 	}
447 
448 	if (test_bit(BCH_FS_rw, &c->flags))
449 		return 0;
450 
451 	bch_info(c, "going read-write");
452 
453 	ret = bch2_sb_members_v2_init(c);
454 	if (ret)
455 		goto err;
456 
457 	ret = bch2_fs_mark_dirty(c);
458 	if (ret)
459 		goto err;
460 
461 	clear_bit(BCH_FS_clean_shutdown, &c->flags);
462 
463 	/*
464 	 * First journal write must be a flush write: after a clean shutdown we
465 	 * don't read the journal, so the first journal write may end up
466 	 * overwriting whatever was there previously, and there must always be
467 	 * at least one non-flush write in the journal or recovery will fail:
468 	 */
469 	set_bit(JOURNAL_need_flush_write, &c->journal.flags);
470 	set_bit(JOURNAL_running, &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_journal_reclaim_start(&c->journal);
489 	if (ret)
490 		goto err;
491 
492 	if (!early) {
493 		ret = bch2_fs_read_write_late(c);
494 		if (ret)
495 			goto err;
496 	}
497 
498 	bch2_do_discards(c);
499 	bch2_do_invalidates(c);
500 	bch2_do_stripe_deletes(c);
501 	bch2_do_pending_node_rewrites(c);
502 	return 0;
503 err:
504 	if (test_bit(BCH_FS_rw, &c->flags))
505 		bch2_fs_read_only(c);
506 	else
507 		__bch2_fs_read_only(c);
508 	return ret;
509 }
510 
511 int bch2_fs_read_write(struct bch_fs *c)
512 {
513 	if (c->opts.recovery_pass_last &&
514 	    c->opts.recovery_pass_last < BCH_RECOVERY_PASS_journal_replay)
515 		return -BCH_ERR_erofs_norecovery;
516 
517 	if (c->opts.nochanges)
518 		return -BCH_ERR_erofs_nochanges;
519 
520 	return __bch2_fs_read_write(c, false);
521 }
522 
523 int bch2_fs_read_write_early(struct bch_fs *c)
524 {
525 	lockdep_assert_held(&c->state_lock);
526 
527 	return __bch2_fs_read_write(c, true);
528 }
529 
530 /* Filesystem startup/shutdown: */
531 
532 static void __bch2_fs_free(struct bch_fs *c)
533 {
534 	for (unsigned i = 0; i < BCH_TIME_STAT_NR; i++)
535 		bch2_time_stats_exit(&c->times[i]);
536 
537 	bch2_find_btree_nodes_exit(&c->found_btree_nodes);
538 	bch2_free_pending_node_rewrites(c);
539 	bch2_fs_accounting_exit(c);
540 	bch2_fs_sb_errors_exit(c);
541 	bch2_fs_counters_exit(c);
542 	bch2_fs_snapshots_exit(c);
543 	bch2_fs_quota_exit(c);
544 	bch2_fs_fs_io_direct_exit(c);
545 	bch2_fs_fs_io_buffered_exit(c);
546 	bch2_fs_fsio_exit(c);
547 	bch2_fs_vfs_exit(c);
548 	bch2_fs_ec_exit(c);
549 	bch2_fs_encryption_exit(c);
550 	bch2_fs_nocow_locking_exit(c);
551 	bch2_fs_io_write_exit(c);
552 	bch2_fs_io_read_exit(c);
553 	bch2_fs_buckets_waiting_for_journal_exit(c);
554 	bch2_fs_btree_interior_update_exit(c);
555 	bch2_fs_btree_key_cache_exit(&c->btree_key_cache);
556 	bch2_fs_btree_cache_exit(c);
557 	bch2_fs_btree_iter_exit(c);
558 	bch2_fs_replicas_exit(c);
559 	bch2_fs_journal_exit(&c->journal);
560 	bch2_io_clock_exit(&c->io_clock[WRITE]);
561 	bch2_io_clock_exit(&c->io_clock[READ]);
562 	bch2_fs_compress_exit(c);
563 	bch2_journal_keys_put_initial(c);
564 	bch2_find_btree_nodes_exit(&c->found_btree_nodes);
565 	BUG_ON(atomic_read(&c->journal_keys.ref));
566 	bch2_fs_btree_write_buffer_exit(c);
567 	percpu_free_rwsem(&c->mark_lock);
568 	if (c->online_reserved) {
569 		u64 v = percpu_u64_get(c->online_reserved);
570 		WARN(v, "online_reserved not 0 at shutdown: %lli", v);
571 		free_percpu(c->online_reserved);
572 	}
573 
574 	darray_exit(&c->btree_roots_extra);
575 	free_percpu(c->pcpu);
576 	free_percpu(c->usage);
577 	mempool_exit(&c->large_bkey_pool);
578 	mempool_exit(&c->btree_bounce_pool);
579 	bioset_exit(&c->btree_bio);
580 	mempool_exit(&c->fill_iter);
581 #ifndef BCH_WRITE_REF_DEBUG
582 	percpu_ref_exit(&c->writes);
583 #endif
584 	kfree(rcu_dereference_protected(c->disk_groups, 1));
585 	kfree(c->journal_seq_blacklist_table);
586 	kfree(c->unused_inode_hints);
587 
588 	if (c->write_ref_wq)
589 		destroy_workqueue(c->write_ref_wq);
590 	if (c->btree_write_submit_wq)
591 		destroy_workqueue(c->btree_write_submit_wq);
592 	if (c->btree_read_complete_wq)
593 		destroy_workqueue(c->btree_read_complete_wq);
594 	if (c->copygc_wq)
595 		destroy_workqueue(c->copygc_wq);
596 	if (c->btree_io_complete_wq)
597 		destroy_workqueue(c->btree_io_complete_wq);
598 	if (c->btree_update_wq)
599 		destroy_workqueue(c->btree_update_wq);
600 
601 	bch2_free_super(&c->disk_sb);
602 	kvfree(c);
603 	module_put(THIS_MODULE);
604 }
605 
606 static void bch2_fs_release(struct kobject *kobj)
607 {
608 	struct bch_fs *c = container_of(kobj, struct bch_fs, kobj);
609 
610 	__bch2_fs_free(c);
611 }
612 
613 void __bch2_fs_stop(struct bch_fs *c)
614 {
615 	bch_verbose(c, "shutting down");
616 
617 	set_bit(BCH_FS_stopping, &c->flags);
618 
619 	down_write(&c->state_lock);
620 	bch2_fs_read_only(c);
621 	up_write(&c->state_lock);
622 
623 	for_each_member_device(c, ca)
624 		bch2_dev_unlink(ca);
625 
626 	if (c->kobj.state_in_sysfs)
627 		kobject_del(&c->kobj);
628 
629 	bch2_fs_debug_exit(c);
630 	bch2_fs_chardev_exit(c);
631 
632 	bch2_ro_ref_put(c);
633 	wait_event(c->ro_ref_wait, !refcount_read(&c->ro_ref));
634 
635 	kobject_put(&c->counters_kobj);
636 	kobject_put(&c->time_stats);
637 	kobject_put(&c->opts_dir);
638 	kobject_put(&c->internal);
639 
640 	/* btree prefetch might have kicked off reads in the background: */
641 	bch2_btree_flush_all_reads(c);
642 
643 	for_each_member_device(c, ca)
644 		cancel_work_sync(&ca->io_error_work);
645 
646 	cancel_work_sync(&c->read_only_work);
647 }
648 
649 void bch2_fs_free(struct bch_fs *c)
650 {
651 	unsigned i;
652 
653 	mutex_lock(&bch_fs_list_lock);
654 	list_del(&c->list);
655 	mutex_unlock(&bch_fs_list_lock);
656 
657 	closure_sync(&c->cl);
658 	closure_debug_destroy(&c->cl);
659 
660 	for (i = 0; i < c->sb.nr_devices; i++) {
661 		struct bch_dev *ca = rcu_dereference_protected(c->devs[i], true);
662 
663 		if (ca) {
664 			EBUG_ON(atomic_long_read(&ca->ref) != 1);
665 			bch2_free_super(&ca->disk_sb);
666 			bch2_dev_free(ca);
667 		}
668 	}
669 
670 	bch_verbose(c, "shutdown complete");
671 
672 	kobject_put(&c->kobj);
673 }
674 
675 void bch2_fs_stop(struct bch_fs *c)
676 {
677 	__bch2_fs_stop(c);
678 	bch2_fs_free(c);
679 }
680 
681 static int bch2_fs_online(struct bch_fs *c)
682 {
683 	int ret = 0;
684 
685 	lockdep_assert_held(&bch_fs_list_lock);
686 
687 	if (__bch2_uuid_to_fs(c->sb.uuid)) {
688 		bch_err(c, "filesystem UUID already open");
689 		return -EINVAL;
690 	}
691 
692 	ret = bch2_fs_chardev_init(c);
693 	if (ret) {
694 		bch_err(c, "error creating character device");
695 		return ret;
696 	}
697 
698 	bch2_fs_debug_init(c);
699 
700 	ret = kobject_add(&c->kobj, NULL, "%pU", c->sb.user_uuid.b) ?:
701 	    kobject_add(&c->internal, &c->kobj, "internal") ?:
702 	    kobject_add(&c->opts_dir, &c->kobj, "options") ?:
703 #ifndef CONFIG_BCACHEFS_NO_LATENCY_ACCT
704 	    kobject_add(&c->time_stats, &c->kobj, "time_stats") ?:
705 #endif
706 	    kobject_add(&c->counters_kobj, &c->kobj, "counters") ?:
707 	    bch2_opts_create_sysfs_files(&c->opts_dir);
708 	if (ret) {
709 		bch_err(c, "error creating sysfs objects");
710 		return ret;
711 	}
712 
713 	down_write(&c->state_lock);
714 
715 	for_each_member_device(c, ca) {
716 		ret = bch2_dev_sysfs_online(c, ca);
717 		if (ret) {
718 			bch_err(c, "error creating sysfs objects");
719 			bch2_dev_put(ca);
720 			goto err;
721 		}
722 	}
723 
724 	BUG_ON(!list_empty(&c->list));
725 	list_add(&c->list, &bch_fs_list);
726 err:
727 	up_write(&c->state_lock);
728 	return ret;
729 }
730 
731 static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
732 {
733 	struct bch_fs *c;
734 	struct printbuf name = PRINTBUF;
735 	unsigned i, iter_size;
736 	int ret = 0;
737 
738 	c = kvmalloc(sizeof(struct bch_fs), GFP_KERNEL|__GFP_ZERO);
739 	if (!c) {
740 		c = ERR_PTR(-BCH_ERR_ENOMEM_fs_alloc);
741 		goto out;
742 	}
743 
744 	c->stdio = (void *)(unsigned long) opts.stdio;
745 
746 	__module_get(THIS_MODULE);
747 
748 	closure_init(&c->cl, NULL);
749 
750 	c->kobj.kset = bcachefs_kset;
751 	kobject_init(&c->kobj, &bch2_fs_ktype);
752 	kobject_init(&c->internal, &bch2_fs_internal_ktype);
753 	kobject_init(&c->opts_dir, &bch2_fs_opts_dir_ktype);
754 	kobject_init(&c->time_stats, &bch2_fs_time_stats_ktype);
755 	kobject_init(&c->counters_kobj, &bch2_fs_counters_ktype);
756 
757 	c->minor		= -1;
758 	c->disk_sb.fs_sb	= true;
759 
760 	init_rwsem(&c->state_lock);
761 	mutex_init(&c->sb_lock);
762 	mutex_init(&c->replicas_gc_lock);
763 	mutex_init(&c->btree_root_lock);
764 	INIT_WORK(&c->read_only_work, bch2_fs_read_only_work);
765 
766 	refcount_set(&c->ro_ref, 1);
767 	init_waitqueue_head(&c->ro_ref_wait);
768 	sema_init(&c->online_fsck_mutex, 1);
769 
770 	init_rwsem(&c->gc_lock);
771 	mutex_init(&c->gc_gens_lock);
772 	atomic_set(&c->journal_keys.ref, 1);
773 	c->journal_keys.initial_ref_held = true;
774 
775 	for (i = 0; i < BCH_TIME_STAT_NR; i++)
776 		bch2_time_stats_init(&c->times[i]);
777 
778 	bch2_fs_gc_init(c);
779 	bch2_fs_copygc_init(c);
780 	bch2_fs_btree_key_cache_init_early(&c->btree_key_cache);
781 	bch2_fs_btree_iter_init_early(c);
782 	bch2_fs_btree_interior_update_init_early(c);
783 	bch2_fs_allocator_background_init(c);
784 	bch2_fs_allocator_foreground_init(c);
785 	bch2_fs_rebalance_init(c);
786 	bch2_fs_quota_init(c);
787 	bch2_fs_ec_init_early(c);
788 	bch2_fs_move_init(c);
789 	bch2_fs_sb_errors_init_early(c);
790 
791 	INIT_LIST_HEAD(&c->list);
792 
793 	mutex_init(&c->bio_bounce_pages_lock);
794 	mutex_init(&c->snapshot_table_lock);
795 	init_rwsem(&c->snapshot_create_lock);
796 
797 	spin_lock_init(&c->btree_write_error_lock);
798 
799 	INIT_LIST_HEAD(&c->journal_iters);
800 
801 	INIT_LIST_HEAD(&c->fsck_error_msgs);
802 	mutex_init(&c->fsck_error_msgs_lock);
803 
804 	seqcount_init(&c->usage_lock);
805 
806 	sema_init(&c->io_in_flight, 128);
807 
808 	INIT_LIST_HEAD(&c->vfs_inodes_list);
809 	mutex_init(&c->vfs_inodes_lock);
810 
811 	c->copy_gc_enabled		= 1;
812 	c->rebalance.enabled		= 1;
813 
814 	c->journal.flush_write_time	= &c->times[BCH_TIME_journal_flush_write];
815 	c->journal.noflush_write_time	= &c->times[BCH_TIME_journal_noflush_write];
816 	c->journal.flush_seq_time	= &c->times[BCH_TIME_journal_flush_seq];
817 
818 	bch2_fs_btree_cache_init_early(&c->btree_cache);
819 
820 	mutex_init(&c->sectors_available_lock);
821 
822 	ret = percpu_init_rwsem(&c->mark_lock);
823 	if (ret)
824 		goto err;
825 
826 	mutex_lock(&c->sb_lock);
827 	ret = bch2_sb_to_fs(c, sb);
828 	mutex_unlock(&c->sb_lock);
829 
830 	if (ret)
831 		goto err;
832 
833 	pr_uuid(&name, c->sb.user_uuid.b);
834 	ret = name.allocation_failure ? -BCH_ERR_ENOMEM_fs_name_alloc : 0;
835 	if (ret)
836 		goto err;
837 
838 	strscpy(c->name, name.buf, sizeof(c->name));
839 	printbuf_exit(&name);
840 
841 	/* Compat: */
842 	if (le16_to_cpu(sb->version) <= bcachefs_metadata_version_inode_v2 &&
843 	    !BCH_SB_JOURNAL_FLUSH_DELAY(sb))
844 		SET_BCH_SB_JOURNAL_FLUSH_DELAY(sb, 1000);
845 
846 	if (le16_to_cpu(sb->version) <= bcachefs_metadata_version_inode_v2 &&
847 	    !BCH_SB_JOURNAL_RECLAIM_DELAY(sb))
848 		SET_BCH_SB_JOURNAL_RECLAIM_DELAY(sb, 100);
849 
850 	c->opts = bch2_opts_default;
851 	ret = bch2_opts_from_sb(&c->opts, sb);
852 	if (ret)
853 		goto err;
854 
855 	bch2_opts_apply(&c->opts, opts);
856 
857 	c->btree_key_cache_btrees |= 1U << BTREE_ID_alloc;
858 	if (c->opts.inodes_use_key_cache)
859 		c->btree_key_cache_btrees |= 1U << BTREE_ID_inodes;
860 	c->btree_key_cache_btrees |= 1U << BTREE_ID_logged_ops;
861 
862 	c->block_bits		= ilog2(block_sectors(c));
863 	c->btree_foreground_merge_threshold = BTREE_FOREGROUND_MERGE_THRESHOLD(c);
864 
865 	if (bch2_fs_init_fault("fs_alloc")) {
866 		bch_err(c, "fs_alloc fault injected");
867 		ret = -EFAULT;
868 		goto err;
869 	}
870 
871 	iter_size = sizeof(struct sort_iter) +
872 		(btree_blocks(c) + 1) * 2 *
873 		sizeof(struct sort_iter_set);
874 
875 	c->inode_shard_bits = ilog2(roundup_pow_of_two(num_possible_cpus()));
876 
877 	if (!(c->btree_update_wq = alloc_workqueue("bcachefs",
878 				WQ_HIGHPRI|WQ_FREEZABLE|WQ_MEM_RECLAIM|WQ_UNBOUND, 512)) ||
879 	    !(c->btree_io_complete_wq = alloc_workqueue("bcachefs_btree_io",
880 				WQ_HIGHPRI|WQ_FREEZABLE|WQ_MEM_RECLAIM, 1)) ||
881 	    !(c->copygc_wq = alloc_workqueue("bcachefs_copygc",
882 				WQ_HIGHPRI|WQ_FREEZABLE|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE, 1)) ||
883 	    !(c->btree_read_complete_wq = alloc_workqueue("bcachefs_btree_read_complete",
884 				WQ_HIGHPRI|WQ_FREEZABLE|WQ_MEM_RECLAIM, 512)) ||
885 	    !(c->btree_write_submit_wq = alloc_workqueue("bcachefs_btree_write_sumit",
886 				WQ_HIGHPRI|WQ_FREEZABLE|WQ_MEM_RECLAIM, 1)) ||
887 	    !(c->write_ref_wq = alloc_workqueue("bcachefs_write_ref",
888 				WQ_FREEZABLE, 0)) ||
889 #ifndef BCH_WRITE_REF_DEBUG
890 	    percpu_ref_init(&c->writes, bch2_writes_disabled,
891 			    PERCPU_REF_INIT_DEAD, GFP_KERNEL) ||
892 #endif
893 	    mempool_init_kmalloc_pool(&c->fill_iter, 1, iter_size) ||
894 	    bioset_init(&c->btree_bio, 1,
895 			max(offsetof(struct btree_read_bio, bio),
896 			    offsetof(struct btree_write_bio, wbio.bio)),
897 			BIOSET_NEED_BVECS) ||
898 	    !(c->pcpu = alloc_percpu(struct bch_fs_pcpu)) ||
899 	    !(c->usage = alloc_percpu(struct bch_fs_usage_base)) ||
900 	    !(c->online_reserved = alloc_percpu(u64)) ||
901 	    mempool_init_kvmalloc_pool(&c->btree_bounce_pool, 1,
902 				       c->opts.btree_node_size) ||
903 	    mempool_init_kmalloc_pool(&c->large_bkey_pool, 1, 2048) ||
904 	    !(c->unused_inode_hints = kcalloc(1U << c->inode_shard_bits,
905 					      sizeof(u64), GFP_KERNEL))) {
906 		ret = -BCH_ERR_ENOMEM_fs_other_alloc;
907 		goto err;
908 	}
909 
910 	ret = bch2_fs_counters_init(c) ?:
911 	    bch2_fs_sb_errors_init(c) ?:
912 	    bch2_io_clock_init(&c->io_clock[READ]) ?:
913 	    bch2_io_clock_init(&c->io_clock[WRITE]) ?:
914 	    bch2_fs_journal_init(&c->journal) ?:
915 	    bch2_fs_btree_iter_init(c) ?:
916 	    bch2_fs_btree_cache_init(c) ?:
917 	    bch2_fs_btree_key_cache_init(&c->btree_key_cache) ?:
918 	    bch2_fs_btree_interior_update_init(c) ?:
919 	    bch2_fs_buckets_waiting_for_journal_init(c) ?:
920 	    bch2_fs_btree_write_buffer_init(c) ?:
921 	    bch2_fs_subvolumes_init(c) ?:
922 	    bch2_fs_io_read_init(c) ?:
923 	    bch2_fs_io_write_init(c) ?:
924 	    bch2_fs_nocow_locking_init(c) ?:
925 	    bch2_fs_encryption_init(c) ?:
926 	    bch2_fs_compress_init(c) ?:
927 	    bch2_fs_ec_init(c) ?:
928 	    bch2_fs_vfs_init(c) ?:
929 	    bch2_fs_fsio_init(c) ?:
930 	    bch2_fs_fs_io_buffered_init(c) ?:
931 	    bch2_fs_fs_io_direct_init(c);
932 	if (ret)
933 		goto err;
934 
935 	for (i = 0; i < c->sb.nr_devices; i++) {
936 		if (!bch2_member_exists(c->disk_sb.sb, i))
937 			continue;
938 		ret = bch2_dev_alloc(c, i);
939 		if (ret)
940 			goto err;
941 	}
942 
943 	bch2_journal_entry_res_resize(&c->journal,
944 			&c->btree_root_journal_res,
945 			BTREE_ID_NR * (JSET_KEYS_U64s + BKEY_BTREE_PTR_U64s_MAX));
946 	bch2_journal_entry_res_resize(&c->journal,
947 			&c->clock_journal_res,
948 			(sizeof(struct jset_entry_clock) / sizeof(u64)) * 2);
949 
950 	mutex_lock(&bch_fs_list_lock);
951 	ret = bch2_fs_online(c);
952 	mutex_unlock(&bch_fs_list_lock);
953 
954 	if (ret)
955 		goto err;
956 out:
957 	return c;
958 err:
959 	bch2_fs_free(c);
960 	c = ERR_PTR(ret);
961 	goto out;
962 }
963 
964 noinline_for_stack
965 static void print_mount_opts(struct bch_fs *c)
966 {
967 	enum bch_opt_id i;
968 	struct printbuf p = PRINTBUF;
969 	bool first = true;
970 
971 	prt_str(&p, "starting version ");
972 	bch2_version_to_text(&p, c->sb.version);
973 
974 	if (c->opts.read_only) {
975 		prt_str(&p, " opts=");
976 		first = false;
977 		prt_printf(&p, "ro");
978 	}
979 
980 	for (i = 0; i < bch2_opts_nr; i++) {
981 		const struct bch_option *opt = &bch2_opt_table[i];
982 		u64 v = bch2_opt_get_by_id(&c->opts, i);
983 
984 		if (!(opt->flags & OPT_MOUNT))
985 			continue;
986 
987 		if (v == bch2_opt_get_by_id(&bch2_opts_default, i))
988 			continue;
989 
990 		prt_str(&p, first ? " opts=" : ",");
991 		first = false;
992 		bch2_opt_to_text(&p, c, c->disk_sb.sb, opt, v, OPT_SHOW_MOUNT_STYLE);
993 	}
994 
995 	bch_info(c, "%s", p.buf);
996 	printbuf_exit(&p);
997 }
998 
999 int bch2_fs_start(struct bch_fs *c)
1000 {
1001 	time64_t now = ktime_get_real_seconds();
1002 	int ret;
1003 
1004 	print_mount_opts(c);
1005 
1006 	down_write(&c->state_lock);
1007 
1008 	BUG_ON(test_bit(BCH_FS_started, &c->flags));
1009 
1010 	mutex_lock(&c->sb_lock);
1011 
1012 	ret = bch2_sb_members_v2_init(c);
1013 	if (ret) {
1014 		mutex_unlock(&c->sb_lock);
1015 		goto err;
1016 	}
1017 
1018 	for_each_online_member(c, ca)
1019 		bch2_members_v2_get_mut(c->disk_sb.sb, ca->dev_idx)->last_mount = cpu_to_le64(now);
1020 
1021 	struct bch_sb_field_ext *ext =
1022 		bch2_sb_field_get_minsize(&c->disk_sb, ext, sizeof(*ext) / sizeof(u64));
1023 	mutex_unlock(&c->sb_lock);
1024 
1025 	if (!ext) {
1026 		bch_err(c, "insufficient space in superblock for sb_field_ext");
1027 		ret = -BCH_ERR_ENOSPC_sb;
1028 		goto err;
1029 	}
1030 
1031 	for_each_rw_member(c, ca)
1032 		bch2_dev_allocator_add(c, ca);
1033 	bch2_recalc_capacity(c);
1034 
1035 	ret = BCH_SB_INITIALIZED(c->disk_sb.sb)
1036 		? bch2_fs_recovery(c)
1037 		: bch2_fs_initialize(c);
1038 	if (ret)
1039 		goto err;
1040 
1041 	ret = bch2_opts_check_may_set(c);
1042 	if (ret)
1043 		goto err;
1044 
1045 	if (bch2_fs_init_fault("fs_start")) {
1046 		bch_err(c, "fs_start fault injected");
1047 		ret = -EINVAL;
1048 		goto err;
1049 	}
1050 
1051 	set_bit(BCH_FS_started, &c->flags);
1052 
1053 	if (c->opts.read_only) {
1054 		bch2_fs_read_only(c);
1055 	} else {
1056 		ret = !test_bit(BCH_FS_rw, &c->flags)
1057 			? bch2_fs_read_write(c)
1058 			: bch2_fs_read_write_late(c);
1059 		if (ret)
1060 			goto err;
1061 	}
1062 
1063 	ret = 0;
1064 err:
1065 	if (ret)
1066 		bch_err_msg(c, ret, "starting filesystem");
1067 	else
1068 		bch_verbose(c, "done starting filesystem");
1069 	up_write(&c->state_lock);
1070 	return ret;
1071 }
1072 
1073 static int bch2_dev_may_add(struct bch_sb *sb, struct bch_fs *c)
1074 {
1075 	struct bch_member m = bch2_sb_member_get(sb, sb->dev_idx);
1076 
1077 	if (le16_to_cpu(sb->block_size) != block_sectors(c))
1078 		return -BCH_ERR_mismatched_block_size;
1079 
1080 	if (le16_to_cpu(m.bucket_size) <
1081 	    BCH_SB_BTREE_NODE_SIZE(c->disk_sb.sb))
1082 		return -BCH_ERR_bucket_size_too_small;
1083 
1084 	return 0;
1085 }
1086 
1087 static int bch2_dev_in_fs(struct bch_sb_handle *fs,
1088 			  struct bch_sb_handle *sb,
1089 			  struct bch_opts *opts)
1090 {
1091 	if (fs == sb)
1092 		return 0;
1093 
1094 	if (!uuid_equal(&fs->sb->uuid, &sb->sb->uuid))
1095 		return -BCH_ERR_device_not_a_member_of_filesystem;
1096 
1097 	if (!bch2_member_exists(fs->sb, sb->sb->dev_idx))
1098 		return -BCH_ERR_device_has_been_removed;
1099 
1100 	if (fs->sb->block_size != sb->sb->block_size)
1101 		return -BCH_ERR_mismatched_block_size;
1102 
1103 	if (le16_to_cpu(fs->sb->version) < bcachefs_metadata_version_member_seq ||
1104 	    le16_to_cpu(sb->sb->version) < bcachefs_metadata_version_member_seq)
1105 		return 0;
1106 
1107 	if (fs->sb->seq == sb->sb->seq &&
1108 	    fs->sb->write_time != sb->sb->write_time) {
1109 		struct printbuf buf = PRINTBUF;
1110 
1111 		prt_str(&buf, "Split brain detected between ");
1112 		prt_bdevname(&buf, sb->bdev);
1113 		prt_str(&buf, " and ");
1114 		prt_bdevname(&buf, fs->bdev);
1115 		prt_char(&buf, ':');
1116 		prt_newline(&buf);
1117 		prt_printf(&buf, "seq=%llu but write_time different, got", le64_to_cpu(sb->sb->seq));
1118 		prt_newline(&buf);
1119 
1120 		prt_bdevname(&buf, fs->bdev);
1121 		prt_char(&buf, ' ');
1122 		bch2_prt_datetime(&buf, le64_to_cpu(fs->sb->write_time));;
1123 		prt_newline(&buf);
1124 
1125 		prt_bdevname(&buf, sb->bdev);
1126 		prt_char(&buf, ' ');
1127 		bch2_prt_datetime(&buf, le64_to_cpu(sb->sb->write_time));;
1128 		prt_newline(&buf);
1129 
1130 		if (!opts->no_splitbrain_check)
1131 			prt_printf(&buf, "Not using older sb");
1132 
1133 		pr_err("%s", buf.buf);
1134 		printbuf_exit(&buf);
1135 
1136 		if (!opts->no_splitbrain_check)
1137 			return -BCH_ERR_device_splitbrain;
1138 	}
1139 
1140 	struct bch_member m = bch2_sb_member_get(fs->sb, sb->sb->dev_idx);
1141 	u64 seq_from_fs		= le64_to_cpu(m.seq);
1142 	u64 seq_from_member	= le64_to_cpu(sb->sb->seq);
1143 
1144 	if (seq_from_fs && seq_from_fs < seq_from_member) {
1145 		struct printbuf buf = PRINTBUF;
1146 
1147 		prt_str(&buf, "Split brain detected between ");
1148 		prt_bdevname(&buf, sb->bdev);
1149 		prt_str(&buf, " and ");
1150 		prt_bdevname(&buf, fs->bdev);
1151 		prt_char(&buf, ':');
1152 		prt_newline(&buf);
1153 
1154 		prt_bdevname(&buf, fs->bdev);
1155 		prt_str(&buf, " believes seq of ");
1156 		prt_bdevname(&buf, sb->bdev);
1157 		prt_printf(&buf, " to be %llu, but ", seq_from_fs);
1158 		prt_bdevname(&buf, sb->bdev);
1159 		prt_printf(&buf, " has %llu\n", seq_from_member);
1160 
1161 		if (!opts->no_splitbrain_check) {
1162 			prt_str(&buf, "Not using ");
1163 			prt_bdevname(&buf, sb->bdev);
1164 		}
1165 
1166 		pr_err("%s", buf.buf);
1167 		printbuf_exit(&buf);
1168 
1169 		if (!opts->no_splitbrain_check)
1170 			return -BCH_ERR_device_splitbrain;
1171 	}
1172 
1173 	return 0;
1174 }
1175 
1176 /* Device startup/shutdown: */
1177 
1178 static void bch2_dev_release(struct kobject *kobj)
1179 {
1180 	struct bch_dev *ca = container_of(kobj, struct bch_dev, kobj);
1181 
1182 	kfree(ca);
1183 }
1184 
1185 static void bch2_dev_free(struct bch_dev *ca)
1186 {
1187 	cancel_work_sync(&ca->io_error_work);
1188 
1189 	bch2_dev_unlink(ca);
1190 
1191 	if (ca->kobj.state_in_sysfs)
1192 		kobject_del(&ca->kobj);
1193 
1194 	bch2_free_super(&ca->disk_sb);
1195 	bch2_dev_allocator_background_exit(ca);
1196 	bch2_dev_journal_exit(ca);
1197 
1198 	free_percpu(ca->io_done);
1199 	bch2_dev_buckets_free(ca);
1200 	free_page((unsigned long) ca->sb_read_scratch);
1201 
1202 	bch2_time_stats_quantiles_exit(&ca->io_latency[WRITE]);
1203 	bch2_time_stats_quantiles_exit(&ca->io_latency[READ]);
1204 
1205 	percpu_ref_exit(&ca->io_ref);
1206 #ifndef CONFIG_BCACHEFS_DEBUG
1207 	percpu_ref_exit(&ca->ref);
1208 #endif
1209 	kobject_put(&ca->kobj);
1210 }
1211 
1212 static void __bch2_dev_offline(struct bch_fs *c, struct bch_dev *ca)
1213 {
1214 
1215 	lockdep_assert_held(&c->state_lock);
1216 
1217 	if (percpu_ref_is_zero(&ca->io_ref))
1218 		return;
1219 
1220 	__bch2_dev_read_only(c, ca);
1221 
1222 	reinit_completion(&ca->io_ref_completion);
1223 	percpu_ref_kill(&ca->io_ref);
1224 	wait_for_completion(&ca->io_ref_completion);
1225 
1226 	bch2_dev_unlink(ca);
1227 
1228 	bch2_free_super(&ca->disk_sb);
1229 	bch2_dev_journal_exit(ca);
1230 }
1231 
1232 #ifndef CONFIG_BCACHEFS_DEBUG
1233 static void bch2_dev_ref_complete(struct percpu_ref *ref)
1234 {
1235 	struct bch_dev *ca = container_of(ref, struct bch_dev, ref);
1236 
1237 	complete(&ca->ref_completion);
1238 }
1239 #endif
1240 
1241 static void bch2_dev_io_ref_complete(struct percpu_ref *ref)
1242 {
1243 	struct bch_dev *ca = container_of(ref, struct bch_dev, io_ref);
1244 
1245 	complete(&ca->io_ref_completion);
1246 }
1247 
1248 static void bch2_dev_unlink(struct bch_dev *ca)
1249 {
1250 	struct kobject *b;
1251 
1252 	/*
1253 	 * This is racy w.r.t. the underlying block device being hot-removed,
1254 	 * which removes it from sysfs.
1255 	 *
1256 	 * It'd be lovely if we had a way to handle this race, but the sysfs
1257 	 * code doesn't appear to provide a good method and block/holder.c is
1258 	 * susceptible as well:
1259 	 */
1260 	if (ca->kobj.state_in_sysfs &&
1261 	    ca->disk_sb.bdev &&
1262 	    (b = bdev_kobj(ca->disk_sb.bdev))->state_in_sysfs) {
1263 		sysfs_remove_link(b, "bcachefs");
1264 		sysfs_remove_link(&ca->kobj, "block");
1265 	}
1266 }
1267 
1268 static int bch2_dev_sysfs_online(struct bch_fs *c, struct bch_dev *ca)
1269 {
1270 	int ret;
1271 
1272 	if (!c->kobj.state_in_sysfs)
1273 		return 0;
1274 
1275 	if (!ca->kobj.state_in_sysfs) {
1276 		ret = kobject_add(&ca->kobj, &c->kobj,
1277 				  "dev-%u", ca->dev_idx);
1278 		if (ret)
1279 			return ret;
1280 	}
1281 
1282 	if (ca->disk_sb.bdev) {
1283 		struct kobject *block = bdev_kobj(ca->disk_sb.bdev);
1284 
1285 		ret = sysfs_create_link(block, &ca->kobj, "bcachefs");
1286 		if (ret)
1287 			return ret;
1288 
1289 		ret = sysfs_create_link(&ca->kobj, block, "block");
1290 		if (ret)
1291 			return ret;
1292 	}
1293 
1294 	return 0;
1295 }
1296 
1297 static struct bch_dev *__bch2_dev_alloc(struct bch_fs *c,
1298 					struct bch_member *member)
1299 {
1300 	struct bch_dev *ca;
1301 	unsigned i;
1302 
1303 	ca = kzalloc(sizeof(*ca), GFP_KERNEL);
1304 	if (!ca)
1305 		return NULL;
1306 
1307 	kobject_init(&ca->kobj, &bch2_dev_ktype);
1308 	init_completion(&ca->ref_completion);
1309 	init_completion(&ca->io_ref_completion);
1310 
1311 	init_rwsem(&ca->bucket_lock);
1312 
1313 	INIT_WORK(&ca->io_error_work, bch2_io_error_work);
1314 
1315 	bch2_time_stats_quantiles_init(&ca->io_latency[READ]);
1316 	bch2_time_stats_quantiles_init(&ca->io_latency[WRITE]);
1317 
1318 	ca->mi = bch2_mi_to_cpu(member);
1319 
1320 	for (i = 0; i < ARRAY_SIZE(member->errors); i++)
1321 		atomic64_set(&ca->errors[i], le64_to_cpu(member->errors[i]));
1322 
1323 	ca->uuid = member->uuid;
1324 
1325 	ca->nr_btree_reserve = DIV_ROUND_UP(BTREE_NODE_RESERVE,
1326 			     ca->mi.bucket_size / btree_sectors(c));
1327 
1328 #ifndef CONFIG_BCACHEFS_DEBUG
1329 	if (percpu_ref_init(&ca->ref, bch2_dev_ref_complete, 0, GFP_KERNEL))
1330 		goto err;
1331 #else
1332 	atomic_long_set(&ca->ref, 1);
1333 #endif
1334 
1335 	bch2_dev_allocator_background_init(ca);
1336 
1337 	if (percpu_ref_init(&ca->io_ref, bch2_dev_io_ref_complete,
1338 			    PERCPU_REF_INIT_DEAD, GFP_KERNEL) ||
1339 	    !(ca->sb_read_scratch = (void *) __get_free_page(GFP_KERNEL)) ||
1340 	    bch2_dev_buckets_alloc(c, ca) ||
1341 	    !(ca->io_done	= alloc_percpu(*ca->io_done)))
1342 		goto err;
1343 
1344 	return ca;
1345 err:
1346 	bch2_dev_free(ca);
1347 	return NULL;
1348 }
1349 
1350 static void bch2_dev_attach(struct bch_fs *c, struct bch_dev *ca,
1351 			    unsigned dev_idx)
1352 {
1353 	ca->dev_idx = dev_idx;
1354 	__set_bit(ca->dev_idx, ca->self.d);
1355 	scnprintf(ca->name, sizeof(ca->name), "dev-%u", dev_idx);
1356 
1357 	ca->fs = c;
1358 	rcu_assign_pointer(c->devs[ca->dev_idx], ca);
1359 
1360 	if (bch2_dev_sysfs_online(c, ca))
1361 		pr_warn("error creating sysfs objects");
1362 }
1363 
1364 static int bch2_dev_alloc(struct bch_fs *c, unsigned dev_idx)
1365 {
1366 	struct bch_member member = bch2_sb_member_get(c->disk_sb.sb, dev_idx);
1367 	struct bch_dev *ca = NULL;
1368 	int ret = 0;
1369 
1370 	if (bch2_fs_init_fault("dev_alloc"))
1371 		goto err;
1372 
1373 	ca = __bch2_dev_alloc(c, &member);
1374 	if (!ca)
1375 		goto err;
1376 
1377 	ca->fs = c;
1378 
1379 	bch2_dev_attach(c, ca, dev_idx);
1380 	return ret;
1381 err:
1382 	if (ca)
1383 		bch2_dev_free(ca);
1384 	return -BCH_ERR_ENOMEM_dev_alloc;
1385 }
1386 
1387 static int __bch2_dev_attach_bdev(struct bch_dev *ca, struct bch_sb_handle *sb)
1388 {
1389 	unsigned ret;
1390 
1391 	if (bch2_dev_is_online(ca)) {
1392 		bch_err(ca, "already have device online in slot %u",
1393 			sb->sb->dev_idx);
1394 		return -BCH_ERR_device_already_online;
1395 	}
1396 
1397 	if (get_capacity(sb->bdev->bd_disk) <
1398 	    ca->mi.bucket_size * ca->mi.nbuckets) {
1399 		bch_err(ca, "cannot online: device too small");
1400 		return -BCH_ERR_device_size_too_small;
1401 	}
1402 
1403 	BUG_ON(!percpu_ref_is_zero(&ca->io_ref));
1404 
1405 	ret = bch2_dev_journal_init(ca, sb->sb);
1406 	if (ret)
1407 		return ret;
1408 
1409 	/* Commit: */
1410 	ca->disk_sb = *sb;
1411 	memset(sb, 0, sizeof(*sb));
1412 
1413 	ca->dev = ca->disk_sb.bdev->bd_dev;
1414 
1415 	percpu_ref_reinit(&ca->io_ref);
1416 
1417 	return 0;
1418 }
1419 
1420 static int bch2_dev_attach_bdev(struct bch_fs *c, struct bch_sb_handle *sb)
1421 {
1422 	struct bch_dev *ca;
1423 	int ret;
1424 
1425 	lockdep_assert_held(&c->state_lock);
1426 
1427 	if (le64_to_cpu(sb->sb->seq) >
1428 	    le64_to_cpu(c->disk_sb.sb->seq))
1429 		bch2_sb_to_fs(c, sb->sb);
1430 
1431 	BUG_ON(!bch2_dev_exists(c, sb->sb->dev_idx));
1432 
1433 	ca = bch2_dev_locked(c, sb->sb->dev_idx);
1434 
1435 	ret = __bch2_dev_attach_bdev(ca, sb);
1436 	if (ret)
1437 		return ret;
1438 
1439 	bch2_dev_sysfs_online(c, ca);
1440 
1441 	struct printbuf name = PRINTBUF;
1442 	prt_bdevname(&name, ca->disk_sb.bdev);
1443 
1444 	if (c->sb.nr_devices == 1)
1445 		strscpy(c->name, name.buf, sizeof(c->name));
1446 	strscpy(ca->name, name.buf, sizeof(ca->name));
1447 
1448 	printbuf_exit(&name);
1449 
1450 	rebalance_wakeup(c);
1451 	return 0;
1452 }
1453 
1454 /* Device management: */
1455 
1456 /*
1457  * Note: this function is also used by the error paths - when a particular
1458  * device sees an error, we call it to determine whether we can just set the
1459  * device RO, or - if this function returns false - we'll set the whole
1460  * filesystem RO:
1461  *
1462  * XXX: maybe we should be more explicit about whether we're changing state
1463  * because we got an error or what have you?
1464  */
1465 bool bch2_dev_state_allowed(struct bch_fs *c, struct bch_dev *ca,
1466 			    enum bch_member_state new_state, int flags)
1467 {
1468 	struct bch_devs_mask new_online_devs;
1469 	int nr_rw = 0, required;
1470 
1471 	lockdep_assert_held(&c->state_lock);
1472 
1473 	switch (new_state) {
1474 	case BCH_MEMBER_STATE_rw:
1475 		return true;
1476 	case BCH_MEMBER_STATE_ro:
1477 		if (ca->mi.state != BCH_MEMBER_STATE_rw)
1478 			return true;
1479 
1480 		/* do we have enough devices to write to?  */
1481 		for_each_member_device(c, ca2)
1482 			if (ca2 != ca)
1483 				nr_rw += ca2->mi.state == BCH_MEMBER_STATE_rw;
1484 
1485 		required = max(!(flags & BCH_FORCE_IF_METADATA_DEGRADED)
1486 			       ? c->opts.metadata_replicas
1487 			       : metadata_replicas_required(c),
1488 			       !(flags & BCH_FORCE_IF_DATA_DEGRADED)
1489 			       ? c->opts.data_replicas
1490 			       : data_replicas_required(c));
1491 
1492 		return nr_rw >= required;
1493 	case BCH_MEMBER_STATE_failed:
1494 	case BCH_MEMBER_STATE_spare:
1495 		if (ca->mi.state != BCH_MEMBER_STATE_rw &&
1496 		    ca->mi.state != BCH_MEMBER_STATE_ro)
1497 			return true;
1498 
1499 		/* do we have enough devices to read from?  */
1500 		new_online_devs = bch2_online_devs(c);
1501 		__clear_bit(ca->dev_idx, new_online_devs.d);
1502 
1503 		return bch2_have_enough_devs(c, new_online_devs, flags, false);
1504 	default:
1505 		BUG();
1506 	}
1507 }
1508 
1509 static bool bch2_fs_may_start(struct bch_fs *c)
1510 {
1511 	struct bch_dev *ca;
1512 	unsigned i, flags = 0;
1513 
1514 	if (c->opts.very_degraded)
1515 		flags |= BCH_FORCE_IF_DEGRADED|BCH_FORCE_IF_LOST;
1516 
1517 	if (c->opts.degraded)
1518 		flags |= BCH_FORCE_IF_DEGRADED;
1519 
1520 	if (!c->opts.degraded &&
1521 	    !c->opts.very_degraded) {
1522 		mutex_lock(&c->sb_lock);
1523 
1524 		for (i = 0; i < c->disk_sb.sb->nr_devices; i++) {
1525 			if (!bch2_member_exists(c->disk_sb.sb, i))
1526 				continue;
1527 
1528 			ca = bch2_dev_locked(c, i);
1529 
1530 			if (!bch2_dev_is_online(ca) &&
1531 			    (ca->mi.state == BCH_MEMBER_STATE_rw ||
1532 			     ca->mi.state == BCH_MEMBER_STATE_ro)) {
1533 				mutex_unlock(&c->sb_lock);
1534 				return false;
1535 			}
1536 		}
1537 		mutex_unlock(&c->sb_lock);
1538 	}
1539 
1540 	return bch2_have_enough_devs(c, bch2_online_devs(c), flags, true);
1541 }
1542 
1543 static void __bch2_dev_read_only(struct bch_fs *c, struct bch_dev *ca)
1544 {
1545 	/*
1546 	 * The allocator thread itself allocates btree nodes, so stop it first:
1547 	 */
1548 	bch2_dev_allocator_remove(c, ca);
1549 	bch2_recalc_capacity(c);
1550 	bch2_dev_journal_stop(&c->journal, ca);
1551 }
1552 
1553 static void __bch2_dev_read_write(struct bch_fs *c, struct bch_dev *ca)
1554 {
1555 	lockdep_assert_held(&c->state_lock);
1556 
1557 	BUG_ON(ca->mi.state != BCH_MEMBER_STATE_rw);
1558 
1559 	bch2_dev_allocator_add(c, ca);
1560 	bch2_recalc_capacity(c);
1561 	bch2_dev_do_discards(ca);
1562 }
1563 
1564 int __bch2_dev_set_state(struct bch_fs *c, struct bch_dev *ca,
1565 			 enum bch_member_state new_state, int flags)
1566 {
1567 	struct bch_member *m;
1568 	int ret = 0;
1569 
1570 	if (ca->mi.state == new_state)
1571 		return 0;
1572 
1573 	if (!bch2_dev_state_allowed(c, ca, new_state, flags))
1574 		return -BCH_ERR_device_state_not_allowed;
1575 
1576 	if (new_state != BCH_MEMBER_STATE_rw)
1577 		__bch2_dev_read_only(c, ca);
1578 
1579 	bch_notice(ca, "%s", bch2_member_states[new_state]);
1580 
1581 	mutex_lock(&c->sb_lock);
1582 	m = bch2_members_v2_get_mut(c->disk_sb.sb, ca->dev_idx);
1583 	SET_BCH_MEMBER_STATE(m, new_state);
1584 	bch2_write_super(c);
1585 	mutex_unlock(&c->sb_lock);
1586 
1587 	if (new_state == BCH_MEMBER_STATE_rw)
1588 		__bch2_dev_read_write(c, ca);
1589 
1590 	rebalance_wakeup(c);
1591 
1592 	return ret;
1593 }
1594 
1595 int bch2_dev_set_state(struct bch_fs *c, struct bch_dev *ca,
1596 		       enum bch_member_state new_state, int flags)
1597 {
1598 	int ret;
1599 
1600 	down_write(&c->state_lock);
1601 	ret = __bch2_dev_set_state(c, ca, new_state, flags);
1602 	up_write(&c->state_lock);
1603 
1604 	return ret;
1605 }
1606 
1607 /* Device add/removal: */
1608 
1609 int bch2_dev_remove(struct bch_fs *c, struct bch_dev *ca, int flags)
1610 {
1611 	struct bch_member *m;
1612 	unsigned dev_idx = ca->dev_idx, data;
1613 	int ret;
1614 
1615 	down_write(&c->state_lock);
1616 
1617 	/*
1618 	 * We consume a reference to ca->ref, regardless of whether we succeed
1619 	 * or fail:
1620 	 */
1621 	bch2_dev_put(ca);
1622 
1623 	if (!bch2_dev_state_allowed(c, ca, BCH_MEMBER_STATE_failed, flags)) {
1624 		bch_err(ca, "Cannot remove without losing data");
1625 		ret = -BCH_ERR_device_state_not_allowed;
1626 		goto err;
1627 	}
1628 
1629 	__bch2_dev_read_only(c, ca);
1630 
1631 	ret = bch2_dev_data_drop(c, ca->dev_idx, flags);
1632 	bch_err_msg(ca, ret, "bch2_dev_data_drop()");
1633 	if (ret)
1634 		goto err;
1635 
1636 	ret = bch2_dev_remove_alloc(c, ca);
1637 	bch_err_msg(ca, ret, "bch2_dev_remove_alloc()");
1638 	if (ret)
1639 		goto err;
1640 
1641 	/*
1642 	 * We need to flush the entire journal to get rid of keys that reference
1643 	 * the device being removed before removing the superblock entry
1644 	 */
1645 	bch2_journal_flush_all_pins(&c->journal);
1646 
1647 	/*
1648 	 * this is really just needed for the bch2_replicas_gc_(start|end)
1649 	 * calls, and could be cleaned up:
1650 	 */
1651 	ret = bch2_journal_flush_device_pins(&c->journal, ca->dev_idx);
1652 	bch_err_msg(ca, ret, "bch2_journal_flush_device_pins()");
1653 	if (ret)
1654 		goto err;
1655 
1656 	ret = bch2_journal_flush(&c->journal);
1657 	bch_err_msg(ca, ret, "bch2_journal_flush()");
1658 	if (ret)
1659 		goto err;
1660 
1661 	ret = bch2_replicas_gc2(c);
1662 	bch_err_msg(ca, ret, "bch2_replicas_gc2()");
1663 	if (ret)
1664 		goto err;
1665 
1666 	data = bch2_dev_has_data(c, ca);
1667 	if (data) {
1668 		struct printbuf data_has = PRINTBUF;
1669 
1670 		prt_bitflags(&data_has, __bch2_data_types, data);
1671 		bch_err(ca, "Remove failed, still has data (%s)", data_has.buf);
1672 		printbuf_exit(&data_has);
1673 		ret = -EBUSY;
1674 		goto err;
1675 	}
1676 
1677 	__bch2_dev_offline(c, ca);
1678 
1679 	mutex_lock(&c->sb_lock);
1680 	rcu_assign_pointer(c->devs[ca->dev_idx], NULL);
1681 	mutex_unlock(&c->sb_lock);
1682 
1683 #ifndef CONFIG_BCACHEFS_DEBUG
1684 	percpu_ref_kill(&ca->ref);
1685 #else
1686 	ca->dying = true;
1687 	bch2_dev_put(ca);
1688 #endif
1689 	wait_for_completion(&ca->ref_completion);
1690 
1691 	bch2_dev_free(ca);
1692 
1693 	/*
1694 	 * Free this device's slot in the bch_member array - all pointers to
1695 	 * this device must be gone:
1696 	 */
1697 	mutex_lock(&c->sb_lock);
1698 	m = bch2_members_v2_get_mut(c->disk_sb.sb, dev_idx);
1699 	memset(&m->uuid, 0, sizeof(m->uuid));
1700 
1701 	bch2_write_super(c);
1702 
1703 	mutex_unlock(&c->sb_lock);
1704 	up_write(&c->state_lock);
1705 	return 0;
1706 err:
1707 	if (ca->mi.state == BCH_MEMBER_STATE_rw &&
1708 	    !percpu_ref_is_zero(&ca->io_ref))
1709 		__bch2_dev_read_write(c, ca);
1710 	up_write(&c->state_lock);
1711 	return ret;
1712 }
1713 
1714 /* Add new device to running filesystem: */
1715 int bch2_dev_add(struct bch_fs *c, const char *path)
1716 {
1717 	struct bch_opts opts = bch2_opts_empty();
1718 	struct bch_sb_handle sb;
1719 	struct bch_dev *ca = NULL;
1720 	struct printbuf errbuf = PRINTBUF;
1721 	struct printbuf label = PRINTBUF;
1722 	int ret;
1723 
1724 	ret = bch2_read_super(path, &opts, &sb);
1725 	bch_err_msg(c, ret, "reading super");
1726 	if (ret)
1727 		goto err;
1728 
1729 	struct bch_member dev_mi = bch2_sb_member_get(sb.sb, sb.sb->dev_idx);
1730 
1731 	if (BCH_MEMBER_GROUP(&dev_mi)) {
1732 		bch2_disk_path_to_text_sb(&label, sb.sb, BCH_MEMBER_GROUP(&dev_mi) - 1);
1733 		if (label.allocation_failure) {
1734 			ret = -ENOMEM;
1735 			goto err;
1736 		}
1737 	}
1738 
1739 	ret = bch2_dev_may_add(sb.sb, c);
1740 	if (ret)
1741 		goto err;
1742 
1743 	ca = __bch2_dev_alloc(c, &dev_mi);
1744 	if (!ca) {
1745 		ret = -ENOMEM;
1746 		goto err;
1747 	}
1748 
1749 	ret = __bch2_dev_attach_bdev(ca, &sb);
1750 	if (ret)
1751 		goto err;
1752 
1753 	ret = bch2_dev_journal_alloc(ca, true);
1754 	bch_err_msg(c, ret, "allocating journal");
1755 	if (ret)
1756 		goto err;
1757 
1758 	down_write(&c->state_lock);
1759 	mutex_lock(&c->sb_lock);
1760 
1761 	ret = bch2_sb_from_fs(c, ca);
1762 	bch_err_msg(c, ret, "setting up new superblock");
1763 	if (ret)
1764 		goto err_unlock;
1765 
1766 	if (dynamic_fault("bcachefs:add:no_slot"))
1767 		goto err_unlock;
1768 
1769 	ret = bch2_sb_member_alloc(c);
1770 	if (ret < 0) {
1771 		bch_err_msg(c, ret, "setting up new superblock");
1772 		goto err_unlock;
1773 	}
1774 	unsigned dev_idx = ret;
1775 
1776 	/* success: */
1777 
1778 	dev_mi.last_mount = cpu_to_le64(ktime_get_real_seconds());
1779 	*bch2_members_v2_get_mut(c->disk_sb.sb, dev_idx) = dev_mi;
1780 
1781 	ca->disk_sb.sb->dev_idx	= dev_idx;
1782 	bch2_dev_attach(c, ca, dev_idx);
1783 
1784 	if (BCH_MEMBER_GROUP(&dev_mi)) {
1785 		ret = __bch2_dev_group_set(c, ca, label.buf);
1786 		bch_err_msg(c, ret, "creating new label");
1787 		if (ret)
1788 			goto err_unlock;
1789 	}
1790 
1791 	bch2_write_super(c);
1792 	mutex_unlock(&c->sb_lock);
1793 
1794 	ret = bch2_dev_usage_init(ca, false);
1795 	if (ret)
1796 		goto err_late;
1797 
1798 	ret = bch2_trans_mark_dev_sb(c, ca, BTREE_TRIGGER_transactional);
1799 	bch_err_msg(ca, ret, "marking new superblock");
1800 	if (ret)
1801 		goto err_late;
1802 
1803 	ret = bch2_fs_freespace_init(c);
1804 	bch_err_msg(ca, ret, "initializing free space");
1805 	if (ret)
1806 		goto err_late;
1807 
1808 	ca->new_fs_bucket_idx = 0;
1809 
1810 	if (ca->mi.state == BCH_MEMBER_STATE_rw)
1811 		__bch2_dev_read_write(c, ca);
1812 
1813 	up_write(&c->state_lock);
1814 	return 0;
1815 
1816 err_unlock:
1817 	mutex_unlock(&c->sb_lock);
1818 	up_write(&c->state_lock);
1819 err:
1820 	if (ca)
1821 		bch2_dev_free(ca);
1822 	bch2_free_super(&sb);
1823 	printbuf_exit(&label);
1824 	printbuf_exit(&errbuf);
1825 	bch_err_fn(c, ret);
1826 	return ret;
1827 err_late:
1828 	up_write(&c->state_lock);
1829 	ca = NULL;
1830 	goto err;
1831 }
1832 
1833 /* Hot add existing device to running filesystem: */
1834 int bch2_dev_online(struct bch_fs *c, const char *path)
1835 {
1836 	struct bch_opts opts = bch2_opts_empty();
1837 	struct bch_sb_handle sb = { NULL };
1838 	struct bch_dev *ca;
1839 	unsigned dev_idx;
1840 	int ret;
1841 
1842 	down_write(&c->state_lock);
1843 
1844 	ret = bch2_read_super(path, &opts, &sb);
1845 	if (ret) {
1846 		up_write(&c->state_lock);
1847 		return ret;
1848 	}
1849 
1850 	dev_idx = sb.sb->dev_idx;
1851 
1852 	ret = bch2_dev_in_fs(&c->disk_sb, &sb, &c->opts);
1853 	bch_err_msg(c, ret, "bringing %s online", path);
1854 	if (ret)
1855 		goto err;
1856 
1857 	ret = bch2_dev_attach_bdev(c, &sb);
1858 	if (ret)
1859 		goto err;
1860 
1861 	ca = bch2_dev_locked(c, dev_idx);
1862 
1863 	ret = bch2_trans_mark_dev_sb(c, ca, BTREE_TRIGGER_transactional);
1864 	bch_err_msg(c, ret, "bringing %s online: error from bch2_trans_mark_dev_sb", path);
1865 	if (ret)
1866 		goto err;
1867 
1868 	if (ca->mi.state == BCH_MEMBER_STATE_rw)
1869 		__bch2_dev_read_write(c, ca);
1870 
1871 	if (!ca->mi.freespace_initialized) {
1872 		ret = bch2_dev_freespace_init(c, ca, 0, ca->mi.nbuckets);
1873 		bch_err_msg(ca, ret, "initializing free space");
1874 		if (ret)
1875 			goto err;
1876 	}
1877 
1878 	if (!ca->journal.nr) {
1879 		ret = bch2_dev_journal_alloc(ca, false);
1880 		bch_err_msg(ca, ret, "allocating journal");
1881 		if (ret)
1882 			goto err;
1883 	}
1884 
1885 	mutex_lock(&c->sb_lock);
1886 	bch2_members_v2_get_mut(c->disk_sb.sb, ca->dev_idx)->last_mount =
1887 		cpu_to_le64(ktime_get_real_seconds());
1888 	bch2_write_super(c);
1889 	mutex_unlock(&c->sb_lock);
1890 
1891 	up_write(&c->state_lock);
1892 	return 0;
1893 err:
1894 	up_write(&c->state_lock);
1895 	bch2_free_super(&sb);
1896 	return ret;
1897 }
1898 
1899 int bch2_dev_offline(struct bch_fs *c, struct bch_dev *ca, int flags)
1900 {
1901 	down_write(&c->state_lock);
1902 
1903 	if (!bch2_dev_is_online(ca)) {
1904 		bch_err(ca, "Already offline");
1905 		up_write(&c->state_lock);
1906 		return 0;
1907 	}
1908 
1909 	if (!bch2_dev_state_allowed(c, ca, BCH_MEMBER_STATE_failed, flags)) {
1910 		bch_err(ca, "Cannot offline required disk");
1911 		up_write(&c->state_lock);
1912 		return -BCH_ERR_device_state_not_allowed;
1913 	}
1914 
1915 	__bch2_dev_offline(c, ca);
1916 
1917 	up_write(&c->state_lock);
1918 	return 0;
1919 }
1920 
1921 int bch2_dev_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
1922 {
1923 	struct bch_member *m;
1924 	u64 old_nbuckets;
1925 	int ret = 0;
1926 
1927 	down_write(&c->state_lock);
1928 	old_nbuckets = ca->mi.nbuckets;
1929 
1930 	if (nbuckets < ca->mi.nbuckets) {
1931 		bch_err(ca, "Cannot shrink yet");
1932 		ret = -EINVAL;
1933 		goto err;
1934 	}
1935 
1936 	if (nbuckets > BCH_MEMBER_NBUCKETS_MAX) {
1937 		bch_err(ca, "New device size too big (%llu greater than max %u)",
1938 			nbuckets, BCH_MEMBER_NBUCKETS_MAX);
1939 		ret = -BCH_ERR_device_size_too_big;
1940 		goto err;
1941 	}
1942 
1943 	if (bch2_dev_is_online(ca) &&
1944 	    get_capacity(ca->disk_sb.bdev->bd_disk) <
1945 	    ca->mi.bucket_size * nbuckets) {
1946 		bch_err(ca, "New size larger than device");
1947 		ret = -BCH_ERR_device_size_too_small;
1948 		goto err;
1949 	}
1950 
1951 	ret = bch2_dev_buckets_resize(c, ca, nbuckets);
1952 	bch_err_msg(ca, ret, "resizing buckets");
1953 	if (ret)
1954 		goto err;
1955 
1956 	ret = bch2_trans_mark_dev_sb(c, ca, BTREE_TRIGGER_transactional);
1957 	if (ret)
1958 		goto err;
1959 
1960 	mutex_lock(&c->sb_lock);
1961 	m = bch2_members_v2_get_mut(c->disk_sb.sb, ca->dev_idx);
1962 	m->nbuckets = cpu_to_le64(nbuckets);
1963 
1964 	bch2_write_super(c);
1965 	mutex_unlock(&c->sb_lock);
1966 
1967 	if (ca->mi.freespace_initialized) {
1968 		struct disk_accounting_pos acc = {
1969 			.type = BCH_DISK_ACCOUNTING_dev_data_type,
1970 			.dev_data_type.dev = ca->dev_idx,
1971 			.dev_data_type.data_type = BCH_DATA_free,
1972 		};
1973 		u64 v[3] = { nbuckets - old_nbuckets, 0, 0 };
1974 
1975 		ret   = bch2_trans_commit_do(ca->fs, NULL, NULL, 0,
1976 				bch2_disk_accounting_mod(trans, &acc, v, ARRAY_SIZE(v), false)) ?:
1977 			bch2_dev_freespace_init(c, ca, old_nbuckets, nbuckets);
1978 		if (ret)
1979 			goto err;
1980 	}
1981 
1982 	bch2_recalc_capacity(c);
1983 err:
1984 	up_write(&c->state_lock);
1985 	return ret;
1986 }
1987 
1988 /* return with ref on ca->ref: */
1989 struct bch_dev *bch2_dev_lookup(struct bch_fs *c, const char *name)
1990 {
1991 	if (!strncmp(name, "/dev/", strlen("/dev/")))
1992 		name += strlen("/dev/");
1993 
1994 	for_each_member_device(c, ca)
1995 		if (!strcmp(name, ca->name))
1996 			return ca;
1997 	return ERR_PTR(-BCH_ERR_ENOENT_dev_not_found);
1998 }
1999 
2000 /* Filesystem open: */
2001 
2002 static inline int sb_cmp(struct bch_sb *l, struct bch_sb *r)
2003 {
2004 	return  cmp_int(le64_to_cpu(l->seq), le64_to_cpu(r->seq)) ?:
2005 		cmp_int(le64_to_cpu(l->write_time), le64_to_cpu(r->write_time));
2006 }
2007 
2008 struct bch_fs *bch2_fs_open(char * const *devices, unsigned nr_devices,
2009 			    struct bch_opts opts)
2010 {
2011 	DARRAY(struct bch_sb_handle) sbs = { 0 };
2012 	struct bch_fs *c = NULL;
2013 	struct bch_sb_handle *best = NULL;
2014 	struct printbuf errbuf = PRINTBUF;
2015 	int ret = 0;
2016 
2017 	if (!try_module_get(THIS_MODULE))
2018 		return ERR_PTR(-ENODEV);
2019 
2020 	if (!nr_devices) {
2021 		ret = -EINVAL;
2022 		goto err;
2023 	}
2024 
2025 	ret = darray_make_room(&sbs, nr_devices);
2026 	if (ret)
2027 		goto err;
2028 
2029 	for (unsigned i = 0; i < nr_devices; i++) {
2030 		struct bch_sb_handle sb = { NULL };
2031 
2032 		ret = bch2_read_super(devices[i], &opts, &sb);
2033 		if (ret)
2034 			goto err;
2035 
2036 		BUG_ON(darray_push(&sbs, sb));
2037 	}
2038 
2039 	if (opts.nochanges && !opts.read_only) {
2040 		ret = -BCH_ERR_erofs_nochanges;
2041 		goto err_print;
2042 	}
2043 
2044 	darray_for_each(sbs, sb)
2045 		if (!best || sb_cmp(sb->sb, best->sb) > 0)
2046 			best = sb;
2047 
2048 	darray_for_each_reverse(sbs, sb) {
2049 		ret = bch2_dev_in_fs(best, sb, &opts);
2050 
2051 		if (ret == -BCH_ERR_device_has_been_removed ||
2052 		    ret == -BCH_ERR_device_splitbrain) {
2053 			bch2_free_super(sb);
2054 			darray_remove_item(&sbs, sb);
2055 			best -= best > sb;
2056 			ret = 0;
2057 			continue;
2058 		}
2059 
2060 		if (ret)
2061 			goto err_print;
2062 	}
2063 
2064 	c = bch2_fs_alloc(best->sb, opts);
2065 	ret = PTR_ERR_OR_ZERO(c);
2066 	if (ret)
2067 		goto err;
2068 
2069 	down_write(&c->state_lock);
2070 	darray_for_each(sbs, sb) {
2071 		ret = bch2_dev_attach_bdev(c, sb);
2072 		if (ret) {
2073 			up_write(&c->state_lock);
2074 			goto err;
2075 		}
2076 	}
2077 	up_write(&c->state_lock);
2078 
2079 	if (!bch2_fs_may_start(c)) {
2080 		ret = -BCH_ERR_insufficient_devices_to_start;
2081 		goto err_print;
2082 	}
2083 
2084 	if (!c->opts.nostart) {
2085 		ret = bch2_fs_start(c);
2086 		if (ret)
2087 			goto err;
2088 	}
2089 out:
2090 	darray_for_each(sbs, sb)
2091 		bch2_free_super(sb);
2092 	darray_exit(&sbs);
2093 	printbuf_exit(&errbuf);
2094 	module_put(THIS_MODULE);
2095 	return c;
2096 err_print:
2097 	pr_err("bch_fs_open err opening %s: %s",
2098 	       devices[0], bch2_err_str(ret));
2099 err:
2100 	if (!IS_ERR_OR_NULL(c))
2101 		bch2_fs_stop(c);
2102 	c = ERR_PTR(ret);
2103 	goto out;
2104 }
2105 
2106 /* Global interfaces/init */
2107 
2108 static void bcachefs_exit(void)
2109 {
2110 	bch2_debug_exit();
2111 	bch2_vfs_exit();
2112 	bch2_chardev_exit();
2113 	bch2_btree_key_cache_exit();
2114 	if (bcachefs_kset)
2115 		kset_unregister(bcachefs_kset);
2116 }
2117 
2118 static int __init bcachefs_init(void)
2119 {
2120 	bch2_bkey_pack_test();
2121 
2122 	if (!(bcachefs_kset = kset_create_and_add("bcachefs", NULL, fs_kobj)) ||
2123 	    bch2_btree_key_cache_init() ||
2124 	    bch2_chardev_init() ||
2125 	    bch2_vfs_init() ||
2126 	    bch2_debug_init())
2127 		goto err;
2128 
2129 	return 0;
2130 err:
2131 	bcachefs_exit();
2132 	return -ENOMEM;
2133 }
2134 
2135 #define BCH_DEBUG_PARAM(name, description)			\
2136 	bool bch2_##name;					\
2137 	module_param_named(name, bch2_##name, bool, 0644);	\
2138 	MODULE_PARM_DESC(name, description);
2139 BCH_DEBUG_PARAMS()
2140 #undef BCH_DEBUG_PARAM
2141 
2142 __maybe_unused
2143 static unsigned bch2_metadata_version = bcachefs_metadata_version_current;
2144 module_param_named(version, bch2_metadata_version, uint, 0400);
2145 
2146 module_exit(bcachefs_exit);
2147 module_init(bcachefs_init);
2148