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 int __copygc_pred(struct bch_fs *c, struct bkey_s_c k)
58 copygc_heap *h = &c->copygc_heap;
59 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
60 const struct bch_extent_ptr *ptr;
62 bkey_for_each_ptr(ptrs, ptr) {
63 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
64 struct copygc_heap_entry search = {
69 ssize_t i = eytzinger0_find_le(h->data, h->used,
71 bucket_offset_cmp, &search);
73 /* eytzinger search verify code: */
76 for (k = 0; k < h->used; k++)
77 if (h->data[k].offset <= ptr->offset &&
78 (j < 0 || h->data[k].offset > h->data[j].offset))
84 ptr->offset < h->data[i].offset + ca->mi.bucket_size &&
85 ptr->gen == h->data[i].gen)
92 static enum data_cmd copygc_pred(struct bch_fs *c, void *arg,
94 struct bch_io_opts *io_opts,
95 struct data_opts *data_opts)
97 int dev_idx = __copygc_pred(c, k);
101 data_opts->target = io_opts->background_target;
102 data_opts->btree_insert_flags = BTREE_INSERT_USE_RESERVE;
103 data_opts->rewrite_dev = dev_idx;
107 static bool have_copygc_reserve(struct bch_dev *ca)
111 spin_lock(&ca->fs->freelist_lock);
112 ret = fifo_full(&ca->free[RESERVE_MOVINGGC]) ||
113 ca->allocator_state != ALLOCATOR_RUNNING;
114 spin_unlock(&ca->fs->freelist_lock);
119 static inline int fragmentation_cmp(copygc_heap *heap,
120 struct copygc_heap_entry l,
121 struct copygc_heap_entry r)
123 return cmp_int(l.fragmentation, r.fragmentation);
126 static int bch2_copygc(struct bch_fs *c)
128 copygc_heap *h = &c->copygc_heap;
129 struct copygc_heap_entry e, *i;
130 struct bucket_array *buckets;
131 struct bch_move_stats move_stats;
132 u64 sectors_to_move = 0, sectors_not_moved = 0;
133 u64 sectors_reserved = 0;
134 u64 buckets_to_move, buckets_not_moved = 0;
137 size_t b, heap_size = 0;
140 memset(&move_stats, 0, sizeof(move_stats));
142 * Find buckets with lowest sector counts, skipping completely
143 * empty buckets, by building a maxheap sorted by sector count,
144 * and repeatedly replacing the maximum element until all
145 * buckets have been visited.
149 for_each_rw_member(ca, c, dev_idx)
150 heap_size += ca->mi.nbuckets >> 7;
152 if (h->size < heap_size) {
153 free_heap(&c->copygc_heap);
154 if (!init_heap(&c->copygc_heap, heap_size, GFP_KERNEL)) {
155 bch_err(c, "error allocating copygc heap");
160 for_each_rw_member(ca, c, dev_idx) {
161 closure_wait_event(&c->freelist_wait, have_copygc_reserve(ca));
163 spin_lock(&ca->fs->freelist_lock);
164 sectors_reserved += fifo_used(&ca->free[RESERVE_MOVINGGC]) * ca->mi.bucket_size;
165 spin_unlock(&ca->fs->freelist_lock);
167 down_read(&ca->bucket_lock);
168 buckets = bucket_array(ca);
170 for (b = buckets->first_bucket; b < buckets->nbuckets; b++) {
171 struct bucket_mark m = READ_ONCE(buckets->b[b].mark);
172 struct copygc_heap_entry e;
174 if (m.owned_by_allocator ||
175 m.data_type != BCH_DATA_user ||
176 !bucket_sectors_used(m) ||
177 bucket_sectors_used(m) >= ca->mi.bucket_size)
180 e = (struct copygc_heap_entry) {
183 .fragmentation = bucket_sectors_used(m) * (1U << 15)
184 / ca->mi.bucket_size,
185 .sectors = bucket_sectors_used(m),
186 .offset = bucket_to_sector(ca, b),
188 heap_add_or_replace(h, e, -fragmentation_cmp, NULL);
190 up_read(&ca->bucket_lock);
193 if (!sectors_reserved) {
194 bch2_fs_fatal_error(c, "stuck, ran out of copygc reserve!");
198 for (i = h->data; i < h->data + h->used; i++)
199 sectors_to_move += i->sectors;
201 while (sectors_to_move > sectors_reserved) {
202 BUG_ON(!heap_pop(h, e, -fragmentation_cmp, NULL));
203 sectors_to_move -= e.sectors;
206 buckets_to_move = h->used;
208 if (!buckets_to_move)
211 eytzinger0_sort(h->data, h->used,
213 bucket_offset_cmp, NULL);
215 ret = bch2_move_data(c, &c->copygc_pd.rate,
216 writepoint_ptr(&c->copygc_write_point),
221 for_each_rw_member(ca, c, dev_idx) {
222 down_read(&ca->bucket_lock);
223 buckets = bucket_array(ca);
224 for (i = h->data; i < h->data + h->used; i++) {
225 struct bucket_mark m;
228 if (i->dev != dev_idx)
231 b = sector_to_bucket(ca, i->offset);
232 m = READ_ONCE(buckets->b[b].mark);
234 if (i->gen == m.gen &&
235 bucket_sectors_used(m)) {
236 sectors_not_moved += bucket_sectors_used(m);
240 up_read(&ca->bucket_lock);
243 if (sectors_not_moved && !ret)
244 bch_warn_ratelimited(c,
245 "copygc finished but %llu/%llu sectors, %llu/%llu buckets not moved (move stats: moved %llu sectors, raced %llu keys, %llu sectors)",
246 sectors_not_moved, sectors_to_move,
247 buckets_not_moved, buckets_to_move,
248 atomic64_read(&move_stats.sectors_moved),
249 atomic64_read(&move_stats.keys_raced),
250 atomic64_read(&move_stats.sectors_raced));
253 atomic64_read(&move_stats.sectors_moved), sectors_not_moved,
254 buckets_to_move, buckets_not_moved);
259 * Copygc runs when the amount of fragmented data is above some arbitrary
262 * The threshold at the limit - when the device is full - is the amount of space
263 * we reserved in bch2_recalc_capacity; we can't have more than that amount of
264 * disk space stranded due to fragmentation and store everything we have
267 * But we don't want to be running copygc unnecessarily when the device still
268 * has plenty of free space - rather, we want copygc to smoothly run every so
269 * often and continually reduce the amount of fragmented space as the device
270 * fills up. So, we increase the threshold by half the current free space.
272 unsigned long bch2_copygc_wait_amount(struct bch_fs *c)
276 u64 fragmented_allowed = c->copygc_threshold;
279 for_each_rw_member(ca, c, dev_idx) {
280 struct bch_dev_usage usage = bch2_dev_usage_read(ca);
282 fragmented_allowed += ((__dev_buckets_available(ca, usage) *
283 ca->mi.bucket_size) >> 1);
284 fragmented += usage.sectors_fragmented;
287 return max_t(s64, 0, fragmented_allowed - fragmented);
290 static int bch2_copygc_thread(void *arg)
292 struct bch_fs *c = arg;
293 struct io_clock *clock = &c->io_clock[WRITE];
294 unsigned long last, wait;
298 while (!kthread_should_stop()) {
299 if (kthread_wait_freezable(c->copy_gc_enabled))
302 last = atomic_long_read(&clock->now);
303 wait = bch2_copygc_wait_amount(c);
305 if (wait > clock->max_slop) {
306 bch2_kthread_io_clock_wait(clock, last + wait,
307 MAX_SCHEDULE_TIMEOUT);
318 void bch2_copygc_stop(struct bch_fs *c)
320 c->copygc_pd.rate.rate = UINT_MAX;
321 bch2_ratelimit_reset(&c->copygc_pd.rate);
323 if (c->copygc_thread) {
324 kthread_stop(c->copygc_thread);
325 put_task_struct(c->copygc_thread);
327 c->copygc_thread = NULL;
330 int bch2_copygc_start(struct bch_fs *c)
332 struct task_struct *t;
334 if (c->copygc_thread)
337 if (c->opts.nochanges)
340 if (bch2_fs_init_fault("copygc_start"))
343 t = kthread_create(bch2_copygc_thread, c, "bch_copygc");
349 c->copygc_thread = t;
350 wake_up_process(c->copygc_thread);
355 void bch2_fs_copygc_init(struct bch_fs *c)
357 bch2_pd_controller_init(&c->copygc_pd);
358 c->copygc_pd.d_term = 0;