1 // SPDX-License-Identifier: GPL-2.0
3 * bcache sysfs interfaces
5 * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
6 * Copyright 2012 Google, Inc.
9 #ifndef NO_BCACHEFS_SYSFS
12 #include "alloc_background.h"
13 #include "alloc_foreground.h"
15 #include "btree_cache.h"
17 #include "btree_iter.h"
18 #include "btree_key_cache.h"
19 #include "btree_update.h"
20 #include "btree_update_interior.h"
24 #include "disk_groups.h"
31 #include "rebalance.h"
36 #include <linux/blkdev.h>
37 #include <linux/sort.h>
38 #include <linux/sched/clock.h>
42 #define SYSFS_OPS(type) \
43 const struct sysfs_ops type ## _sysfs_ops = { \
44 .show = type ## _show, \
45 .store = type ## _store \
49 static ssize_t fn ## _to_text(struct printbuf *, \
50 struct kobject *, struct attribute *);\
52 static ssize_t fn ## _show(struct kobject *kobj, struct attribute *attr,\
55 struct printbuf out = PRINTBUF; \
56 ssize_t ret = fn ## _to_text(&out, kobj, attr); \
58 if (out.pos && out.buf[out.pos - 1] != '\n') \
61 if (!ret && out.allocation_failure) \
65 ret = min_t(size_t, out.pos, PAGE_SIZE - 1); \
66 memcpy(buf, out.buf, ret); \
68 printbuf_exit(&out); \
72 static ssize_t fn ## _to_text(struct printbuf *out, struct kobject *kobj,\
73 struct attribute *attr)
76 static ssize_t fn ## _store(struct kobject *kobj, struct attribute *attr,\
77 const char *buf, size_t size) \
79 #define __sysfs_attribute(_name, _mode) \
80 static struct attribute sysfs_##_name = \
81 { .name = #_name, .mode = _mode }
83 #define write_attribute(n) __sysfs_attribute(n, S_IWUSR)
84 #define read_attribute(n) __sysfs_attribute(n, S_IRUGO)
85 #define rw_attribute(n) __sysfs_attribute(n, S_IRUGO|S_IWUSR)
87 #define sysfs_printf(file, fmt, ...) \
89 if (attr == &sysfs_ ## file) \
90 pr_buf(out, fmt "\n", __VA_ARGS__); \
93 #define sysfs_print(file, var) \
95 if (attr == &sysfs_ ## file) \
99 #define sysfs_hprint(file, val) \
101 if (attr == &sysfs_ ## file) \
102 bch2_hprint(out, val); \
105 #define var_printf(_var, fmt) sysfs_printf(_var, fmt, var(_var))
106 #define var_print(_var) sysfs_print(_var, var(_var))
107 #define var_hprint(_var) sysfs_hprint(_var, var(_var))
109 #define sysfs_strtoul(file, var) \
111 if (attr == &sysfs_ ## file) \
112 return strtoul_safe(buf, var) ?: (ssize_t) size; \
115 #define sysfs_strtoul_clamp(file, var, min, max) \
117 if (attr == &sysfs_ ## file) \
118 return strtoul_safe_clamp(buf, var, min, max) \
122 #define strtoul_or_return(cp) \
125 int _r = kstrtoul(cp, 10, &_v); \
131 #define strtoul_restrict_or_return(cp, min, max) \
133 unsigned long __v = 0; \
134 int _r = strtoul_safe_restrict(cp, __v, min, max); \
140 #define strtoi_h_or_return(cp) \
143 int _r = strtoi_h(cp, &_v); \
149 #define sysfs_hatoi(file, var) \
151 if (attr == &sysfs_ ## file) \
152 return strtoi_h(buf, &var) ?: (ssize_t) size; \
155 write_attribute(trigger_gc);
156 write_attribute(trigger_discards);
157 write_attribute(prune_cache);
158 rw_attribute(btree_gc_periodic);
159 rw_attribute(gc_gens_pos);
161 read_attribute(uuid);
162 read_attribute(minor);
163 read_attribute(bucket_size);
164 read_attribute(first_bucket);
165 read_attribute(nbuckets);
166 read_attribute(durability);
167 read_attribute(iodone);
169 read_attribute(io_latency_read);
170 read_attribute(io_latency_write);
171 read_attribute(io_latency_stats_read);
172 read_attribute(io_latency_stats_write);
173 read_attribute(congested);
175 read_attribute(btree_avg_write_size);
177 read_attribute(btree_cache_size);
178 read_attribute(compression_stats);
179 read_attribute(journal_debug);
180 read_attribute(btree_updates);
181 read_attribute(btree_cache);
182 read_attribute(btree_key_cache);
183 read_attribute(btree_transactions);
184 read_attribute(stripes_heap);
185 read_attribute(open_buckets);
187 read_attribute(internal_uuid);
189 read_attribute(has_data);
190 read_attribute(alloc_debug);
192 read_attribute(read_realloc_races);
193 read_attribute(extent_migrate_done);
194 read_attribute(extent_migrate_raced);
195 read_attribute(bucket_alloc_fail);
197 #define x(t, n, ...) read_attribute(t);
198 BCH_PERSISTENT_COUNTERS()
201 rw_attribute(discard);
204 rw_attribute(copy_gc_enabled);
205 read_attribute(copy_gc_wait);
207 rw_attribute(rebalance_enabled);
208 sysfs_pd_controller_attribute(rebalance);
209 read_attribute(rebalance_work);
210 rw_attribute(promote_whole_extents);
212 read_attribute(new_stripes);
214 read_attribute(io_timers_read);
215 read_attribute(io_timers_write);
217 read_attribute(data_jobs);
219 #ifdef CONFIG_BCACHEFS_TESTS
220 write_attribute(perf_test);
221 #endif /* CONFIG_BCACHEFS_TESTS */
224 static struct attribute sysfs_time_stat_##_name = \
225 { .name = #_name, .mode = S_IRUGO };
229 static struct attribute sysfs_state_rw = {
234 static size_t bch2_btree_cache_size(struct bch_fs *c)
239 mutex_lock(&c->btree_cache.lock);
240 list_for_each_entry(b, &c->btree_cache.live, list)
241 ret += btree_bytes(c);
243 mutex_unlock(&c->btree_cache.lock);
247 static size_t bch2_btree_avg_write_size(struct bch_fs *c)
249 u64 nr = atomic64_read(&c->btree_writes_nr);
250 u64 sectors = atomic64_read(&c->btree_writes_sectors);
252 return nr ? div64_u64(sectors, nr) : 0;
255 static long data_progress_to_text(struct printbuf *out, struct bch_fs *c)
258 struct bch_move_stats *stats;
260 mutex_lock(&c->data_progress_lock);
261 list_for_each_entry(stats, &c->data_progress_list, list) {
262 pr_buf(out, "%s: data type %s btree_id %s position: ",
264 bch2_data_types[stats->data_type],
265 bch2_btree_ids[stats->btree_id]);
266 bch2_bpos_to_text(out, stats->pos);
267 pr_buf(out, "%s", "\n");
270 mutex_unlock(&c->data_progress_lock);
274 static int bch2_compression_stats_to_text(struct printbuf *out, struct bch_fs *c)
276 struct btree_trans trans;
277 struct btree_iter iter;
280 u64 nr_uncompressed_extents = 0,
281 nr_compressed_extents = 0,
282 nr_incompressible_extents = 0,
283 uncompressed_sectors = 0,
284 incompressible_sectors = 0,
285 compressed_sectors_compressed = 0,
286 compressed_sectors_uncompressed = 0;
289 if (!test_bit(BCH_FS_STARTED, &c->flags))
292 bch2_trans_init(&trans, c, 0, 0);
294 for (id = 0; id < BTREE_ID_NR; id++) {
295 if (!((1U << id) & BTREE_ID_HAS_PTRS))
298 for_each_btree_key(&trans, iter, id, POS_MIN,
299 BTREE_ITER_ALL_SNAPSHOTS, k, ret) {
300 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
301 const union bch_extent_entry *entry;
302 struct extent_ptr_decoded p;
303 bool compressed = false, uncompressed = false, incompressible = false;
305 bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
306 switch (p.crc.compression_type) {
307 case BCH_COMPRESSION_TYPE_none:
309 uncompressed_sectors += k.k->size;
311 case BCH_COMPRESSION_TYPE_incompressible:
312 incompressible = true;
313 incompressible_sectors += k.k->size;
316 compressed_sectors_compressed +=
317 p.crc.compressed_size;
318 compressed_sectors_uncompressed +=
319 p.crc.uncompressed_size;
326 nr_incompressible_extents++;
327 else if (uncompressed)
328 nr_uncompressed_extents++;
330 nr_compressed_extents++;
332 bch2_trans_iter_exit(&trans, &iter);
335 bch2_trans_exit(&trans);
340 pr_buf(out, "uncompressed:\n");
341 pr_buf(out, " nr extents: %llu\n", nr_uncompressed_extents);
342 pr_buf(out, " size: ");
343 bch2_hprint(out, uncompressed_sectors << 9);
346 pr_buf(out, "compressed:\n");
347 pr_buf(out, " nr extents: %llu\n", nr_compressed_extents);
348 pr_buf(out, " compressed size: ");
349 bch2_hprint(out, compressed_sectors_compressed << 9);
351 pr_buf(out, " uncompressed size: ");
352 bch2_hprint(out, compressed_sectors_uncompressed << 9);
355 pr_buf(out, "incompressible:\n");
356 pr_buf(out, " nr extents: %llu\n", nr_incompressible_extents);
357 pr_buf(out, " size: ");
358 bch2_hprint(out, incompressible_sectors << 9);
363 static void bch2_gc_gens_pos_to_text(struct printbuf *out, struct bch_fs *c)
365 pr_buf(out, "%s: ", bch2_btree_ids[c->gc_gens_btree]);
366 bch2_bpos_to_text(out, c->gc_gens_pos);
372 struct bch_fs *c = container_of(kobj, struct bch_fs, kobj);
374 sysfs_print(minor, c->minor);
375 sysfs_printf(internal_uuid, "%pU", c->sb.uuid.b);
377 sysfs_hprint(btree_cache_size, bch2_btree_cache_size(c));
378 sysfs_hprint(btree_avg_write_size, bch2_btree_avg_write_size(c));
380 sysfs_print(read_realloc_races,
381 atomic_long_read(&c->read_realloc_races));
382 sysfs_print(extent_migrate_done,
383 atomic_long_read(&c->extent_migrate_done));
384 sysfs_print(extent_migrate_raced,
385 atomic_long_read(&c->extent_migrate_raced));
386 sysfs_print(bucket_alloc_fail,
387 atomic_long_read(&c->bucket_alloc_fail));
389 sysfs_printf(btree_gc_periodic, "%u", (int) c->btree_gc_periodic);
391 if (attr == &sysfs_gc_gens_pos)
392 bch2_gc_gens_pos_to_text(out, c);
394 sysfs_printf(copy_gc_enabled, "%i", c->copy_gc_enabled);
396 sysfs_printf(rebalance_enabled, "%i", c->rebalance.enabled);
397 sysfs_pd_controller_show(rebalance, &c->rebalance.pd); /* XXX */
398 sysfs_hprint(copy_gc_wait,
399 max(0LL, c->copygc_wait -
400 atomic64_read(&c->io_clock[WRITE].now)) << 9);
402 if (attr == &sysfs_rebalance_work)
403 bch2_rebalance_work_to_text(out, c);
405 sysfs_print(promote_whole_extents, c->promote_whole_extents);
409 if (attr == &sysfs_journal_debug)
410 bch2_journal_debug_to_text(out, &c->journal);
412 if (attr == &sysfs_btree_updates)
413 bch2_btree_updates_to_text(out, c);
415 if (attr == &sysfs_btree_cache)
416 bch2_btree_cache_to_text(out, c);
418 if (attr == &sysfs_btree_key_cache)
419 bch2_btree_key_cache_to_text(out, &c->btree_key_cache);
421 if (attr == &sysfs_btree_transactions)
422 bch2_btree_trans_to_text(out, c);
424 if (attr == &sysfs_stripes_heap)
425 bch2_stripes_heap_to_text(out, c);
427 if (attr == &sysfs_open_buckets)
428 bch2_open_buckets_to_text(out, c);
430 if (attr == &sysfs_compression_stats)
431 bch2_compression_stats_to_text(out, c);
433 if (attr == &sysfs_new_stripes)
434 bch2_new_stripes_to_text(out, c);
436 if (attr == &sysfs_io_timers_read)
437 bch2_io_timers_to_text(out, &c->io_clock[READ]);
439 if (attr == &sysfs_io_timers_write)
440 bch2_io_timers_to_text(out, &c->io_clock[WRITE]);
442 if (attr == &sysfs_data_jobs)
443 data_progress_to_text(out, c);
450 struct bch_fs *c = container_of(kobj, struct bch_fs, kobj);
452 if (attr == &sysfs_btree_gc_periodic) {
453 ssize_t ret = strtoul_safe(buf, c->btree_gc_periodic)
456 wake_up_process(c->gc_thread);
460 if (attr == &sysfs_copy_gc_enabled) {
461 ssize_t ret = strtoul_safe(buf, c->copy_gc_enabled)
464 if (c->copygc_thread)
465 wake_up_process(c->copygc_thread);
469 if (attr == &sysfs_rebalance_enabled) {
470 ssize_t ret = strtoul_safe(buf, c->rebalance.enabled)
477 sysfs_pd_controller_store(rebalance, &c->rebalance.pd);
479 sysfs_strtoul(promote_whole_extents, c->promote_whole_extents);
483 if (!test_bit(BCH_FS_STARTED, &c->flags))
488 if (!test_bit(BCH_FS_RW, &c->flags))
491 if (attr == &sysfs_prune_cache) {
492 struct shrink_control sc;
494 sc.gfp_mask = GFP_KERNEL;
495 sc.nr_to_scan = strtoul_or_return(buf);
496 c->btree_cache.shrink.scan_objects(&c->btree_cache.shrink, &sc);
499 if (attr == &sysfs_trigger_gc) {
501 * Full gc is currently incompatible with btree key cache:
504 down_read(&c->state_lock);
505 bch2_gc(c, false, false);
506 up_read(&c->state_lock);
512 if (attr == &sysfs_trigger_discards)
515 #ifdef CONFIG_BCACHEFS_TESTS
516 if (attr == &sysfs_perf_test) {
517 char *tmp = kstrdup(buf, GFP_KERNEL), *p = tmp;
518 char *test = strsep(&p, " \t\n");
519 char *nr_str = strsep(&p, " \t\n");
520 char *threads_str = strsep(&p, " \t\n");
526 !(ret = kstrtouint(threads_str, 10, &threads)) &&
527 !(ret = bch2_strtoull_h(nr_str, &nr)))
528 ret = bch2_btree_perf_test(c, test, nr, threads);
539 struct attribute *bch2_fs_files[] = {
541 &sysfs_btree_cache_size,
542 &sysfs_btree_avg_write_size,
544 &sysfs_promote_whole_extents,
546 &sysfs_compression_stats,
548 #ifdef CONFIG_BCACHEFS_TESTS
556 SHOW(bch2_fs_counters)
558 struct bch_fs *c = container_of(kobj, struct bch_fs, counters_kobj);
560 u64 counter_since_mount = 0;
562 out->tabstops[0] = 32;
564 if (attr == &sysfs_##t) { \
565 counter = percpu_u64_get(&c->counters[BCH_COUNTER_##t]);\
566 counter_since_mount = counter - c->counters_on_mount[BCH_COUNTER_##t];\
567 pr_buf(out, "since mount:"); \
569 bch2_hprint(out, counter_since_mount << 9); \
572 pr_buf(out, "since filesystem creation:"); \
574 bch2_hprint(out, counter << 9); \
577 BCH_PERSISTENT_COUNTERS()
582 STORE(bch2_fs_counters) {
586 SYSFS_OPS(bch2_fs_counters);
588 struct attribute *bch2_fs_counters_files[] = {
591 BCH_PERSISTENT_COUNTERS()
595 /* internal dir - just a wrapper */
597 SHOW(bch2_fs_internal)
599 struct bch_fs *c = container_of(kobj, struct bch_fs, internal);
600 return bch2_fs_to_text(out, &c->kobj, attr);
603 STORE(bch2_fs_internal)
605 struct bch_fs *c = container_of(kobj, struct bch_fs, internal);
606 return bch2_fs_store(&c->kobj, attr, buf, size);
608 SYSFS_OPS(bch2_fs_internal);
610 struct attribute *bch2_fs_internal_files[] = {
611 &sysfs_journal_debug,
612 &sysfs_btree_updates,
614 &sysfs_btree_key_cache,
615 &sysfs_btree_transactions,
619 &sysfs_io_timers_read,
620 &sysfs_io_timers_write,
623 &sysfs_trigger_discards,
626 &sysfs_read_realloc_races,
627 &sysfs_extent_migrate_done,
628 &sysfs_extent_migrate_raced,
629 &sysfs_bucket_alloc_fail,
633 &sysfs_copy_gc_enabled,
636 &sysfs_rebalance_enabled,
637 &sysfs_rebalance_work,
638 sysfs_pd_controller_files(rebalance),
642 &sysfs_internal_uuid,
648 SHOW(bch2_fs_opts_dir)
650 struct bch_fs *c = container_of(kobj, struct bch_fs, opts_dir);
651 const struct bch_option *opt = container_of(attr, struct bch_option, attr);
652 int id = opt - bch2_opt_table;
653 u64 v = bch2_opt_get_by_id(&c->opts, id);
655 bch2_opt_to_text(out, c, c->disk_sb.sb, opt, v, OPT_SHOW_FULL_LIST);
661 STORE(bch2_fs_opts_dir)
663 struct bch_fs *c = container_of(kobj, struct bch_fs, opts_dir);
664 const struct bch_option *opt = container_of(attr, struct bch_option, attr);
665 int ret, id = opt - bch2_opt_table;
670 * We don't need to take c->writes for correctness, but it eliminates an
671 * unsightly error message in the dmesg log when we're RO:
673 if (unlikely(!percpu_ref_tryget(&c->writes)))
676 tmp = kstrdup(buf, GFP_KERNEL);
682 ret = bch2_opt_parse(c, opt, strim(tmp), &v, NULL);
688 ret = bch2_opt_check_may_set(c, id, v);
692 bch2_opt_set_sb(c, opt, v);
693 bch2_opt_set_by_id(&c->opts, id, v);
695 if ((id == Opt_background_target ||
696 id == Opt_background_compression) && v) {
697 bch2_rebalance_add_work(c, S64_MAX);
703 percpu_ref_put(&c->writes);
706 SYSFS_OPS(bch2_fs_opts_dir);
708 struct attribute *bch2_fs_opts_dir_files[] = { NULL };
710 int bch2_opts_create_sysfs_files(struct kobject *kobj)
712 const struct bch_option *i;
715 for (i = bch2_opt_table;
716 i < bch2_opt_table + bch2_opts_nr;
718 if (!(i->flags & OPT_FS))
721 ret = sysfs_create_file(kobj, &i->attr);
731 SHOW(bch2_fs_time_stats)
733 struct bch_fs *c = container_of(kobj, struct bch_fs, time_stats);
736 if (attr == &sysfs_time_stat_##name) \
737 bch2_time_stats_to_text(out, &c->times[BCH_TIME_##name]);
744 STORE(bch2_fs_time_stats)
748 SYSFS_OPS(bch2_fs_time_stats);
750 struct attribute *bch2_fs_time_stats_files[] = {
752 &sysfs_time_stat_##name,
758 static void dev_alloc_debug_to_text(struct printbuf *out, struct bch_dev *ca)
760 struct bch_fs *c = ca->fs;
761 struct bch_dev_usage stats = bch2_dev_usage_read(ca);
762 unsigned i, nr[BCH_DATA_NR];
764 memset(nr, 0, sizeof(nr));
766 for (i = 0; i < ARRAY_SIZE(c->open_buckets); i++)
767 nr[c->open_buckets[i].data_type]++;
770 "\t\t\t buckets\t sectors fragmented\n"
771 "capacity\t%16llu\n",
772 ca->mi.nbuckets - ca->mi.first_bucket);
774 for (i = 0; i < BCH_DATA_NR; i++)
775 pr_buf(out, "%-16s%16llu%16llu%16llu\n",
776 bch2_data_types[i], stats.d[i].buckets,
777 stats.d[i].sectors, stats.d[i].fragmented);
782 "freelist_wait\t\t%s\n"
783 "open buckets allocated\t%u\n"
784 "open buckets this dev\t%u\n"
785 "open buckets total\t%u\n"
786 "open_buckets_wait\t%s\n"
787 "open_buckets_btree\t%u\n"
788 "open_buckets_user\t%u\n"
789 "btree reserve cache\t%u\n",
791 c->freelist_wait.list.first ? "waiting" : "empty",
792 OPEN_BUCKETS_COUNT - c->open_buckets_nr_free,
795 c->open_buckets_wait.list.first ? "waiting" : "empty",
798 c->btree_reserve_cache_nr);
801 static const char * const bch2_rw[] = {
807 static void dev_iodone_to_text(struct printbuf *out, struct bch_dev *ca)
811 for (rw = 0; rw < 2; rw++) {
812 pr_buf(out, "%s:\n", bch2_rw[rw]);
814 for (i = 1; i < BCH_DATA_NR; i++)
815 pr_buf(out, "%-12s:%12llu\n",
817 percpu_u64_get(&ca->io_done->sectors[rw][i]) << 9);
823 struct bch_dev *ca = container_of(kobj, struct bch_dev, kobj);
824 struct bch_fs *c = ca->fs;
826 sysfs_printf(uuid, "%pU\n", ca->uuid.b);
828 sysfs_print(bucket_size, bucket_bytes(ca));
829 sysfs_print(first_bucket, ca->mi.first_bucket);
830 sysfs_print(nbuckets, ca->mi.nbuckets);
831 sysfs_print(durability, ca->mi.durability);
832 sysfs_print(discard, ca->mi.discard);
834 if (attr == &sysfs_label) {
836 mutex_lock(&c->sb_lock);
837 bch2_disk_path_to_text(out, c->disk_sb.sb,
839 mutex_unlock(&c->sb_lock);
845 if (attr == &sysfs_has_data) {
846 bch2_flags_to_text(out, bch2_data_types,
847 bch2_dev_has_data(c, ca));
851 if (attr == &sysfs_state_rw) {
852 bch2_string_opt_to_text(out, bch2_member_states,
857 if (attr == &sysfs_iodone)
858 dev_iodone_to_text(out, ca);
860 sysfs_print(io_latency_read, atomic64_read(&ca->cur_latency[READ]));
861 sysfs_print(io_latency_write, atomic64_read(&ca->cur_latency[WRITE]));
863 if (attr == &sysfs_io_latency_stats_read)
864 bch2_time_stats_to_text(out, &ca->io_latency[READ]);
866 if (attr == &sysfs_io_latency_stats_write)
867 bch2_time_stats_to_text(out, &ca->io_latency[WRITE]);
869 sysfs_printf(congested, "%u%%",
870 clamp(atomic_read(&ca->congested), 0, CONGESTED_MAX)
871 * 100 / CONGESTED_MAX);
873 if (attr == &sysfs_alloc_debug)
874 dev_alloc_debug_to_text(out, ca);
881 struct bch_dev *ca = container_of(kobj, struct bch_dev, kobj);
882 struct bch_fs *c = ca->fs;
883 struct bch_member *mi;
885 if (attr == &sysfs_discard) {
886 bool v = strtoul_or_return(buf);
888 mutex_lock(&c->sb_lock);
889 mi = &bch2_sb_get_members(c->disk_sb.sb)->members[ca->dev_idx];
891 if (v != BCH_MEMBER_DISCARD(mi)) {
892 SET_BCH_MEMBER_DISCARD(mi, v);
895 mutex_unlock(&c->sb_lock);
898 if (attr == &sysfs_label) {
902 tmp = kstrdup(buf, GFP_KERNEL);
906 ret = bch2_dev_group_set(c, ca, strim(tmp));
916 struct attribute *bch2_dev_files[] = {
931 &sysfs_io_latency_read,
932 &sysfs_io_latency_write,
933 &sysfs_io_latency_stats_read,
934 &sysfs_io_latency_stats_write,
942 #endif /* _BCACHEFS_SYSFS_H_ */