1 // SPDX-License-Identifier: GPL-2.0
3 * Moving/copying garbage collector
5 * Copyright 2012 Google, Inc.
9 #include "alloc_foreground.h"
10 #include "btree_iter.h"
11 #include "btree_update.h"
14 #include "disk_groups.h"
17 #include "eytzinger.h"
24 #include <trace/events/bcachefs.h>
25 #include <linux/freezer.h>
26 #include <linux/kthread.h>
27 #include <linux/math64.h>
28 #include <linux/sched/task.h>
29 #include <linux/sort.h>
30 #include <linux/wait.h>
33 * We can't use the entire copygc reserve in one iteration of copygc: we may
34 * need the buckets we're freeing up to go back into the copygc reserve to make
35 * forward progress, but if the copygc reserve is full they'll be available for
36 * any allocation - and it's possible that in a given iteration, we free up most
37 * of the buckets we're going to free before we allocate most of the buckets
38 * we're going to allocate.
40 * If we only use half of the reserve per iteration, then in steady state we'll
41 * always have room in the reserve for the buckets we're going to need in the
44 #define COPYGC_BUCKETS_PER_ITER(ca) \
45 ((ca)->free[RESERVE_MOVINGGC].size / 2)
47 static int bucket_offset_cmp(const void *_l, const void *_r, size_t size)
49 const struct copygc_heap_entry *l = _l;
50 const struct copygc_heap_entry *r = _r;
52 return cmp_int(l->dev, r->dev) ?:
53 cmp_int(l->offset, r->offset);
56 static enum data_cmd copygc_pred(struct bch_fs *c, void *arg,
58 struct bch_io_opts *io_opts,
59 struct data_opts *data_opts)
61 copygc_heap *h = &c->copygc_heap;
62 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
63 const union bch_extent_entry *entry;
64 struct extent_ptr_decoded p;
66 bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
67 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
68 struct copygc_heap_entry search = {
70 .offset = p.ptr.offset,
73 ssize_t i = eytzinger0_find_le(h->data, h->used,
75 bucket_offset_cmp, &search);
77 /* eytzinger search verify code: */
80 for (k = 0; k < h->used; k++)
81 if (h->data[k].offset <= ptr->offset &&
82 (j < 0 || h->data[k].offset > h->data[j].offset))
88 p.ptr.offset < h->data[i].offset + ca->mi.bucket_size &&
89 p.ptr.gen == h->data[i].gen) {
90 data_opts->target = io_opts->background_target;
91 data_opts->nr_replicas = 1;
92 data_opts->btree_insert_flags = BTREE_INSERT_USE_RESERVE;
93 data_opts->rewrite_dev = p.ptr.dev;
96 struct stripe *m = genradix_ptr(&c->stripes[0], p.ec.idx);
98 data_opts->nr_replicas += m->nr_redundant;
108 static bool have_copygc_reserve(struct bch_dev *ca)
112 spin_lock(&ca->fs->freelist_lock);
113 ret = fifo_full(&ca->free[RESERVE_MOVINGGC]) ||
114 ca->allocator_state != ALLOCATOR_RUNNING;
115 spin_unlock(&ca->fs->freelist_lock);
120 static inline int fragmentation_cmp(copygc_heap *heap,
121 struct copygc_heap_entry l,
122 struct copygc_heap_entry r)
124 return cmp_int(l.fragmentation, r.fragmentation);
127 static int bch2_copygc(struct bch_fs *c)
129 copygc_heap *h = &c->copygc_heap;
130 struct copygc_heap_entry e, *i;
131 struct bucket_array *buckets;
132 struct bch_move_stats move_stats;
133 u64 sectors_to_move = 0, sectors_not_moved = 0;
134 u64 sectors_reserved = 0;
135 u64 buckets_to_move, buckets_not_moved = 0;
138 size_t b, heap_size = 0;
141 memset(&move_stats, 0, sizeof(move_stats));
143 * Find buckets with lowest sector counts, skipping completely
144 * empty buckets, by building a maxheap sorted by sector count,
145 * and repeatedly replacing the maximum element until all
146 * buckets have been visited.
150 for_each_rw_member(ca, c, dev_idx)
151 heap_size += ca->mi.nbuckets >> 7;
153 if (h->size < heap_size) {
154 free_heap(&c->copygc_heap);
155 if (!init_heap(&c->copygc_heap, heap_size, GFP_KERNEL)) {
156 bch_err(c, "error allocating copygc heap");
161 for_each_rw_member(ca, c, dev_idx) {
162 closure_wait_event(&c->freelist_wait, have_copygc_reserve(ca));
164 spin_lock(&ca->fs->freelist_lock);
165 sectors_reserved += fifo_used(&ca->free[RESERVE_MOVINGGC]) * ca->mi.bucket_size;
166 spin_unlock(&ca->fs->freelist_lock);
168 down_read(&ca->bucket_lock);
169 buckets = bucket_array(ca);
171 for (b = buckets->first_bucket; b < buckets->nbuckets; b++) {
172 struct bucket *g = buckets->b + b;
173 struct bucket_mark m = READ_ONCE(g->mark);
174 struct copygc_heap_entry e;
176 if (m.owned_by_allocator ||
177 m.data_type != BCH_DATA_user ||
178 !bucket_sectors_used(m) ||
179 bucket_sectors_used(m) >= ca->mi.bucket_size)
182 WARN_ON(m.stripe && !g->ec_redundancy);
184 e = (struct copygc_heap_entry) {
187 .replicas = 1 + g->ec_redundancy,
188 .fragmentation = bucket_sectors_used(m) * (1U << 15)
189 / ca->mi.bucket_size,
190 .sectors = bucket_sectors_used(m),
191 .offset = bucket_to_sector(ca, b),
193 heap_add_or_replace(h, e, -fragmentation_cmp, NULL);
195 up_read(&ca->bucket_lock);
198 if (!sectors_reserved) {
199 bch2_fs_fatal_error(c, "stuck, ran out of copygc reserve!");
203 for (i = h->data; i < h->data + h->used; i++)
204 sectors_to_move += i->sectors * i->replicas;
206 while (sectors_to_move > sectors_reserved) {
207 BUG_ON(!heap_pop(h, e, -fragmentation_cmp, NULL));
208 sectors_to_move -= e.sectors * e.replicas;
211 buckets_to_move = h->used;
213 if (!buckets_to_move)
216 eytzinger0_sort(h->data, h->used,
218 bucket_offset_cmp, NULL);
220 ret = bch2_move_data(c, &c->copygc_pd.rate,
221 writepoint_ptr(&c->copygc_write_point),
226 for_each_rw_member(ca, c, dev_idx) {
227 down_read(&ca->bucket_lock);
228 buckets = bucket_array(ca);
229 for (i = h->data; i < h->data + h->used; i++) {
230 struct bucket_mark m;
233 if (i->dev != dev_idx)
236 b = sector_to_bucket(ca, i->offset);
237 m = READ_ONCE(buckets->b[b].mark);
239 if (i->gen == m.gen &&
240 bucket_sectors_used(m)) {
241 sectors_not_moved += bucket_sectors_used(m);
245 up_read(&ca->bucket_lock);
248 if (sectors_not_moved && !ret)
249 bch_warn_ratelimited(c,
250 "copygc finished but %llu/%llu sectors, %llu/%llu buckets not moved (move stats: moved %llu sectors, raced %llu keys, %llu sectors)",
251 sectors_not_moved, sectors_to_move,
252 buckets_not_moved, buckets_to_move,
253 atomic64_read(&move_stats.sectors_moved),
254 atomic64_read(&move_stats.keys_raced),
255 atomic64_read(&move_stats.sectors_raced));
258 atomic64_read(&move_stats.sectors_moved), sectors_not_moved,
259 buckets_to_move, buckets_not_moved);
264 * Copygc runs when the amount of fragmented data is above some arbitrary
267 * The threshold at the limit - when the device is full - is the amount of space
268 * we reserved in bch2_recalc_capacity; we can't have more than that amount of
269 * disk space stranded due to fragmentation and store everything we have
272 * But we don't want to be running copygc unnecessarily when the device still
273 * has plenty of free space - rather, we want copygc to smoothly run every so
274 * often and continually reduce the amount of fragmented space as the device
275 * fills up. So, we increase the threshold by half the current free space.
277 unsigned long bch2_copygc_wait_amount(struct bch_fs *c)
281 u64 fragmented_allowed = c->copygc_threshold;
284 for_each_rw_member(ca, c, dev_idx) {
285 struct bch_dev_usage usage = bch2_dev_usage_read(ca);
287 fragmented_allowed += ((__dev_buckets_available(ca, usage) *
288 ca->mi.bucket_size) >> 1);
289 fragmented += usage.sectors_fragmented;
292 return max_t(s64, 0, fragmented_allowed - fragmented);
295 static int bch2_copygc_thread(void *arg)
297 struct bch_fs *c = arg;
298 struct io_clock *clock = &c->io_clock[WRITE];
299 unsigned long last, wait;
303 while (!kthread_should_stop()) {
304 if (kthread_wait_freezable(c->copy_gc_enabled))
307 last = atomic_long_read(&clock->now);
308 wait = bch2_copygc_wait_amount(c);
310 if (wait > clock->max_slop) {
311 bch2_kthread_io_clock_wait(clock, last + wait,
312 MAX_SCHEDULE_TIMEOUT);
323 void bch2_copygc_stop(struct bch_fs *c)
325 c->copygc_pd.rate.rate = UINT_MAX;
326 bch2_ratelimit_reset(&c->copygc_pd.rate);
328 if (c->copygc_thread) {
329 kthread_stop(c->copygc_thread);
330 put_task_struct(c->copygc_thread);
332 c->copygc_thread = NULL;
335 int bch2_copygc_start(struct bch_fs *c)
337 struct task_struct *t;
339 if (c->copygc_thread)
342 if (c->opts.nochanges)
345 if (bch2_fs_init_fault("copygc_start"))
348 t = kthread_create(bch2_copygc_thread, c, "bch_copygc");
354 c->copygc_thread = t;
355 wake_up_process(c->copygc_thread);
360 void bch2_fs_copygc_init(struct bch_fs *c)
362 bch2_pd_controller_init(&c->copygc_pd);
363 c->copygc_pd.d_term = 0;