1 // SPDX-License-Identifier: GPL-2.0
4 #include "alloc_foreground.h"
5 #include "btree_iter.h"
8 #include "disk_groups.h"
12 #include "rebalance.h"
15 #include <linux/freezer.h>
16 #include <linux/kthread.h>
17 #include <linux/sched/cputime.h>
18 #include <trace/events/bcachefs.h>
21 * Check if an extent should be moved:
22 * returns -1 if it should not be moved, or
23 * device of pointer that should be moved, if known, or INT_MAX if unknown
25 static bool rebalance_pred(struct bch_fs *c, void *arg,
27 struct bch_io_opts *io_opts,
28 struct data_update_opts *data_opts)
30 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
33 data_opts->rewrite_ptrs = 0;
34 data_opts->target = io_opts->background_target;
35 data_opts->extra_replicas = 0;
36 data_opts->btree_insert_flags = 0;
38 if (io_opts->background_compression &&
39 !bch2_bkey_is_incompressible(k)) {
40 const union bch_extent_entry *entry;
41 struct extent_ptr_decoded p;
44 bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
46 p.crc.compression_type !=
47 bch2_compression_opt_to_type[io_opts->background_compression])
48 data_opts->rewrite_ptrs |= 1U << i;
53 if (io_opts->background_target) {
54 const struct bch_extent_ptr *ptr;
57 bkey_for_each_ptr(ptrs, ptr) {
59 !bch2_dev_in_target(c, ptr->dev, io_opts->background_target))
60 data_opts->rewrite_ptrs |= 1U << i;
65 return data_opts->rewrite_ptrs != 0;
68 void bch2_rebalance_add_key(struct bch_fs *c,
70 struct bch_io_opts *io_opts)
72 struct data_update_opts update_opts = { 0 };
73 struct bkey_ptrs_c ptrs;
74 const struct bch_extent_ptr *ptr;
77 if (!rebalance_pred(c, NULL, k, io_opts, &update_opts))
81 ptrs = bch2_bkey_ptrs_c(k);
82 bkey_for_each_ptr(ptrs, ptr) {
83 if ((1U << i) && update_opts.rewrite_ptrs)
84 if (atomic64_add_return(k.k->size,
85 &bch_dev_bkey_exists(c, ptr->dev)->rebalance_work) ==
92 void bch2_rebalance_add_work(struct bch_fs *c, u64 sectors)
94 if (atomic64_add_return(sectors, &c->rebalance.work_unknown_dev) ==
99 struct rebalance_work {
100 int dev_most_full_idx;
101 unsigned dev_most_full_percent;
102 u64 dev_most_full_work;
103 u64 dev_most_full_capacity;
107 static void rebalance_work_accumulate(struct rebalance_work *w,
108 u64 dev_work, u64 unknown_dev, u64 capacity, int idx)
110 unsigned percent_full;
111 u64 work = dev_work + unknown_dev;
113 if (work < dev_work || work < unknown_dev)
115 work = min(work, capacity);
117 percent_full = div64_u64(work * 100, capacity);
119 if (percent_full >= w->dev_most_full_percent) {
120 w->dev_most_full_idx = idx;
121 w->dev_most_full_percent = percent_full;
122 w->dev_most_full_work = work;
123 w->dev_most_full_capacity = capacity;
126 if (w->total_work + dev_work >= w->total_work &&
127 w->total_work + dev_work >= dev_work)
128 w->total_work += dev_work;
131 static struct rebalance_work rebalance_work(struct bch_fs *c)
134 struct rebalance_work ret = { .dev_most_full_idx = -1 };
135 u64 unknown_dev = atomic64_read(&c->rebalance.work_unknown_dev);
138 for_each_online_member(ca, c, i)
139 rebalance_work_accumulate(&ret,
140 atomic64_read(&ca->rebalance_work),
142 bucket_to_sector(ca, ca->mi.nbuckets -
143 ca->mi.first_bucket),
146 rebalance_work_accumulate(&ret,
147 unknown_dev, 0, c->capacity, -1);
152 static void rebalance_work_reset(struct bch_fs *c)
157 for_each_online_member(ca, c, i)
158 atomic64_set(&ca->rebalance_work, 0);
160 atomic64_set(&c->rebalance.work_unknown_dev, 0);
163 static unsigned long curr_cputime(void)
167 task_cputime_adjusted(current, &utime, &stime);
168 return nsecs_to_jiffies(utime + stime);
171 static int bch2_rebalance_thread(void *arg)
173 struct bch_fs *c = arg;
174 struct bch_fs_rebalance *r = &c->rebalance;
175 struct io_clock *clock = &c->io_clock[WRITE];
176 struct rebalance_work w, p;
177 struct bch_move_stats move_stats;
178 unsigned long start, prev_start;
179 unsigned long prev_run_time, prev_run_cputime;
180 unsigned long cputime, prev_cputime;
186 io_start = atomic64_read(&clock->now);
187 p = rebalance_work(c);
188 prev_start = jiffies;
189 prev_cputime = curr_cputime();
191 bch_move_stats_init(&move_stats, "rebalance");
192 while (!kthread_wait_freezable(r->enabled)) {
196 cputime = curr_cputime();
198 prev_run_time = start - prev_start;
199 prev_run_cputime = cputime - prev_cputime;
201 w = rebalance_work(c);
202 BUG_ON(!w.dev_most_full_capacity);
205 r->state = REBALANCE_WAITING;
206 kthread_wait_freezable(rebalance_work(c).total_work);
211 * If there isn't much work to do, throttle cpu usage:
213 throttle = prev_run_cputime * 100 /
214 max(1U, w.dev_most_full_percent) -
217 if (w.dev_most_full_percent < 20 && throttle > 0) {
218 r->throttled_until_iotime = io_start +
219 div_u64(w.dev_most_full_capacity *
220 (20 - w.dev_most_full_percent),
223 if (atomic64_read(&clock->now) + clock->max_slop <
224 r->throttled_until_iotime) {
225 r->throttled_until_cputime = start + throttle;
226 r->state = REBALANCE_THROTTLED;
228 bch2_kthread_io_clock_wait(clock,
229 r->throttled_until_iotime,
235 /* minimum 1 mb/sec: */
239 max(p.dev_most_full_percent, 1U) /
240 max(w.dev_most_full_percent, 1U));
242 io_start = atomic64_read(&clock->now);
245 prev_cputime = cputime;
247 r->state = REBALANCE_RUNNING;
248 memset(&move_stats, 0, sizeof(move_stats));
249 rebalance_work_reset(c);
253 BTREE_ID_NR, POS_MAX,
254 /* ratelimiting disabled for now */
255 NULL, /* &r->pd.rate, */
257 writepoint_ptr(&c->rebalance_write_point),
259 rebalance_pred, NULL);
265 void bch2_rebalance_work_to_text(struct printbuf *out, struct bch_fs *c)
267 struct bch_fs_rebalance *r = &c->rebalance;
268 struct rebalance_work w = rebalance_work(c);
270 out->tabstops[0] = 20;
272 prt_printf(out, "fullest_dev (%i):", w.dev_most_full_idx);
275 prt_human_readable_u64(out, w.dev_most_full_work << 9);
276 prt_printf(out, "/");
277 prt_human_readable_u64(out, w.dev_most_full_capacity << 9);
280 prt_printf(out, "total work:");
283 prt_human_readable_u64(out, w.total_work << 9);
284 prt_printf(out, "/");
285 prt_human_readable_u64(out, c->capacity << 9);
288 prt_printf(out, "rate:");
290 prt_printf(out, "%u", r->pd.rate.rate);
294 case REBALANCE_WAITING:
295 prt_printf(out, "waiting");
297 case REBALANCE_THROTTLED:
298 prt_printf(out, "throttled for %lu sec or ",
299 (r->throttled_until_cputime - jiffies) / HZ);
300 prt_human_readable_u64(out,
301 (r->throttled_until_iotime -
302 atomic64_read(&c->io_clock[WRITE].now)) << 9);
303 prt_printf(out, " io");
305 case REBALANCE_RUNNING:
306 prt_printf(out, "running");
312 void bch2_rebalance_stop(struct bch_fs *c)
314 struct task_struct *p;
316 c->rebalance.pd.rate.rate = UINT_MAX;
317 bch2_ratelimit_reset(&c->rebalance.pd.rate);
319 p = rcu_dereference_protected(c->rebalance.thread, 1);
320 c->rebalance.thread = NULL;
323 /* for sychronizing with rebalance_wakeup() */
331 int bch2_rebalance_start(struct bch_fs *c)
333 struct task_struct *p;
335 if (c->rebalance.thread)
338 if (c->opts.nochanges)
341 p = kthread_create(bch2_rebalance_thread, c, "bch-rebalance/%s", c->name);
343 bch_err(c, "error creating rebalance thread: %li", PTR_ERR(p));
348 rcu_assign_pointer(c->rebalance.thread, p);
353 void bch2_fs_rebalance_init(struct bch_fs *c)
355 bch2_pd_controller_init(&c->rebalance.pd);
357 atomic64_set(&c->rebalance.work_unknown_dev, S64_MAX);