1 // SPDX-License-Identifier: GPL-2.0
4 #include "alloc_background.h"
5 #include "alloc_foreground.h"
6 #include "btree_iter.h"
7 #include "btree_update.h"
8 #include "btree_write_buffer.h"
12 #include "disk_groups.h"
17 #include "rebalance.h"
18 #include "subvolume.h"
22 #include <linux/freezer.h>
23 #include <linux/kthread.h>
24 #include <linux/sched/cputime.h>
26 #define REBALANCE_WORK_SCAN_OFFSET (U64_MAX - 1)
28 static const char * const bch2_rebalance_state_strs[] = {
30 BCH_REBALANCE_STATES()
35 static int __bch2_set_rebalance_needs_scan(struct btree_trans *trans, u64 inum)
37 struct btree_iter iter;
39 struct bkey_i_cookie *cookie;
43 bch2_trans_iter_init(trans, &iter, BTREE_ID_rebalance_work,
44 SPOS(inum, REBALANCE_WORK_SCAN_OFFSET, U32_MAX),
46 k = bch2_btree_iter_peek_slot(&iter);
51 v = k.k->type == KEY_TYPE_cookie
52 ? le64_to_cpu(bkey_s_c_to_cookie(k).v->cookie)
55 cookie = bch2_trans_kmalloc(trans, sizeof(*cookie));
56 ret = PTR_ERR_OR_ZERO(cookie);
60 bkey_cookie_init(&cookie->k_i);
61 cookie->k.p = iter.pos;
62 cookie->v.cookie = cpu_to_le64(v + 1);
64 ret = bch2_trans_update(trans, &iter, &cookie->k_i, 0);
66 bch2_trans_iter_exit(trans, &iter);
70 int bch2_set_rebalance_needs_scan(struct bch_fs *c, u64 inum)
72 int ret = bch2_trans_do(c, NULL, NULL, BCH_TRANS_COMMIT_no_enospc|BCH_TRANS_COMMIT_lazy_rw,
73 __bch2_set_rebalance_needs_scan(trans, inum));
78 int bch2_set_fs_needs_rebalance(struct bch_fs *c)
80 return bch2_set_rebalance_needs_scan(c, 0);
83 static int bch2_clear_rebalance_needs_scan(struct btree_trans *trans, u64 inum, u64 cookie)
85 struct btree_iter iter;
90 bch2_trans_iter_init(trans, &iter, BTREE_ID_rebalance_work,
91 SPOS(inum, REBALANCE_WORK_SCAN_OFFSET, U32_MAX),
93 k = bch2_btree_iter_peek_slot(&iter);
98 v = k.k->type == KEY_TYPE_cookie
99 ? le64_to_cpu(bkey_s_c_to_cookie(k).v->cookie)
103 ret = bch2_btree_delete_at(trans, &iter, 0);
105 bch2_trans_iter_exit(trans, &iter);
109 static struct bkey_s_c next_rebalance_entry(struct btree_trans *trans,
110 struct btree_iter *work_iter)
112 return !kthread_should_stop()
113 ? bch2_btree_iter_peek(work_iter)
117 static int bch2_bkey_clear_needs_rebalance(struct btree_trans *trans,
118 struct btree_iter *iter,
121 struct bkey_i *n = bch2_bkey_make_mut(trans, iter, &k, 0);
122 int ret = PTR_ERR_OR_ZERO(n);
126 extent_entry_drop(bkey_i_to_s(n),
127 (void *) bch2_bkey_rebalance_opts(bkey_i_to_s_c(n)));
128 return bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc);
131 static struct bkey_s_c next_rebalance_extent(struct btree_trans *trans,
132 struct bpos work_pos,
133 struct btree_iter *extent_iter,
134 struct data_update_opts *data_opts)
136 struct bch_fs *c = trans->c;
139 bch2_trans_iter_exit(trans, extent_iter);
140 bch2_trans_iter_init(trans, extent_iter,
141 work_pos.inode ? BTREE_ID_extents : BTREE_ID_reflink,
143 BTREE_ITER_ALL_SNAPSHOTS);
144 k = bch2_btree_iter_peek_slot(extent_iter);
148 const struct bch_extent_rebalance *r = k.k ? bch2_bkey_rebalance_opts(k) : NULL;
150 /* raced due to btree write buffer, nothing to do */
151 return bkey_s_c_null;
154 memset(data_opts, 0, sizeof(*data_opts));
156 data_opts->rewrite_ptrs =
157 bch2_bkey_ptrs_need_rebalance(c, k, r->target, r->compression);
158 data_opts->target = r->target;
160 if (!data_opts->rewrite_ptrs) {
162 * device we would want to write to offline? devices in target
165 * We'll now need a full scan before this extent is picked up
168 int ret = bch2_bkey_clear_needs_rebalance(trans, extent_iter, k);
170 return bkey_s_c_err(ret);
171 return bkey_s_c_null;
178 static int do_rebalance_extent(struct moving_context *ctxt,
179 struct bpos work_pos,
180 struct btree_iter *extent_iter)
182 struct btree_trans *trans = ctxt->trans;
183 struct bch_fs *c = trans->c;
184 struct bch_fs_rebalance *r = &trans->c->rebalance;
185 struct data_update_opts data_opts;
186 struct bch_io_opts io_opts;
191 ctxt->stats = &r->work_stats;
192 r->state = BCH_REBALANCE_working;
194 bch2_bkey_buf_init(&sk);
196 ret = bkey_err(k = next_rebalance_extent(trans, work_pos,
197 extent_iter, &data_opts));
201 ret = bch2_move_get_io_opts_one(trans, &io_opts, k);
205 atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
208 * The iterator gets unlocked by __bch2_read_extent - need to
209 * save a copy of @k elsewhere:
211 bch2_bkey_buf_reassemble(&sk, c, k);
212 k = bkey_i_to_s_c(sk.k);
214 ret = bch2_move_extent(ctxt, NULL, extent_iter, k, io_opts, data_opts);
216 if (bch2_err_matches(ret, ENOMEM)) {
217 /* memory allocation failure, wait for some IO to finish */
218 bch2_move_ctxt_wait_for_io(ctxt);
219 ret = -BCH_ERR_transaction_restart_nested;
222 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
225 /* skip it and continue, XXX signal failure */
229 bch2_bkey_buf_exit(&sk, c);
233 static bool rebalance_pred(struct bch_fs *c, void *arg,
235 struct bch_io_opts *io_opts,
236 struct data_update_opts *data_opts)
238 unsigned target, compression;
241 target = io_opts->background_target;
242 compression = io_opts->background_compression ?: io_opts->compression;
244 const struct bch_extent_rebalance *r = bch2_bkey_rebalance_opts(k);
246 target = r ? r->target : io_opts->background_target;
247 compression = r ? r->compression :
248 (io_opts->background_compression ?: io_opts->compression);
251 data_opts->rewrite_ptrs = bch2_bkey_ptrs_need_rebalance(c, k, target, compression);
252 data_opts->target = target;
253 return data_opts->rewrite_ptrs != 0;
256 static int do_rebalance_scan(struct moving_context *ctxt, u64 inum, u64 cookie)
258 struct btree_trans *trans = ctxt->trans;
259 struct bch_fs_rebalance *r = &trans->c->rebalance;
262 bch2_move_stats_init(&r->scan_stats, "rebalance_scan");
263 ctxt->stats = &r->scan_stats;
266 r->scan_start = BBPOS_MIN;
267 r->scan_end = BBPOS_MAX;
269 r->scan_start = BBPOS(BTREE_ID_extents, POS(inum, 0));
270 r->scan_end = BBPOS(BTREE_ID_extents, POS(inum, U64_MAX));
273 r->state = BCH_REBALANCE_scanning;
275 ret = __bch2_move_data(ctxt, r->scan_start, r->scan_end, rebalance_pred, NULL) ?:
276 commit_do(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
277 bch2_clear_rebalance_needs_scan(trans, inum, cookie));
279 bch2_move_stats_exit(&r->scan_stats, trans->c);
283 static void rebalance_wait(struct bch_fs *c)
285 struct bch_fs_rebalance *r = &c->rebalance;
286 struct io_clock *clock = &c->io_clock[WRITE];
287 u64 now = atomic64_read(&clock->now);
288 u64 min_member_capacity = bch2_min_rw_member_capacity(c);
290 if (min_member_capacity == U64_MAX)
291 min_member_capacity = 128 * 2048;
293 r->wait_iotime_end = now + (min_member_capacity >> 6);
295 if (r->state != BCH_REBALANCE_waiting) {
296 r->wait_iotime_start = now;
297 r->wait_wallclock_start = ktime_get_real_ns();
298 r->state = BCH_REBALANCE_waiting;
301 bch2_kthread_io_clock_wait(clock, r->wait_iotime_end, MAX_SCHEDULE_TIMEOUT);
304 static int do_rebalance(struct moving_context *ctxt)
306 struct btree_trans *trans = ctxt->trans;
307 struct bch_fs *c = trans->c;
308 struct bch_fs_rebalance *r = &c->rebalance;
309 struct btree_iter rebalance_work_iter, extent_iter = { NULL };
313 bch2_move_stats_init(&r->work_stats, "rebalance_work");
314 bch2_move_stats_init(&r->scan_stats, "rebalance_scan");
316 bch2_trans_iter_init(trans, &rebalance_work_iter,
317 BTREE_ID_rebalance_work, POS_MIN,
318 BTREE_ITER_ALL_SNAPSHOTS);
320 while (!bch2_move_ratelimit(ctxt) &&
321 !kthread_wait_freezable(r->enabled)) {
322 bch2_trans_begin(trans);
324 ret = bkey_err(k = next_rebalance_entry(trans, &rebalance_work_iter));
325 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
330 ret = k.k->type == KEY_TYPE_cookie
331 ? do_rebalance_scan(ctxt, k.k->p.inode,
332 le64_to_cpu(bkey_s_c_to_cookie(k).v->cookie))
333 : do_rebalance_extent(ctxt, k.k->p, &extent_iter);
335 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
340 bch2_btree_iter_advance(&rebalance_work_iter);
343 bch2_trans_iter_exit(trans, &extent_iter);
344 bch2_trans_iter_exit(trans, &rebalance_work_iter);
345 bch2_move_stats_exit(&r->scan_stats, c);
348 !kthread_should_stop() &&
349 !atomic64_read(&r->work_stats.sectors_seen) &&
350 !atomic64_read(&r->scan_stats.sectors_seen)) {
351 bch2_trans_unlock_long(trans);
355 if (!bch2_err_matches(ret, EROFS))
360 static int bch2_rebalance_thread(void *arg)
362 struct bch_fs *c = arg;
363 struct bch_fs_rebalance *r = &c->rebalance;
364 struct moving_context ctxt;
369 bch2_moving_ctxt_init(&ctxt, c, NULL, &r->work_stats,
370 writepoint_ptr(&c->rebalance_write_point),
373 while (!kthread_should_stop() &&
374 !(ret = do_rebalance(&ctxt)))
377 bch2_moving_ctxt_exit(&ctxt);
382 void bch2_rebalance_status_to_text(struct printbuf *out, struct bch_fs *c)
384 struct bch_fs_rebalance *r = &c->rebalance;
386 prt_str(out, bch2_rebalance_state_strs[r->state]);
388 printbuf_indent_add(out, 2);
391 case BCH_REBALANCE_waiting: {
392 u64 now = atomic64_read(&c->io_clock[WRITE].now);
394 prt_str(out, "io wait duration: ");
395 bch2_prt_human_readable_s64(out, r->wait_iotime_end - r->wait_iotime_start);
398 prt_str(out, "io wait remaining: ");
399 bch2_prt_human_readable_s64(out, r->wait_iotime_end - now);
402 prt_str(out, "duration waited: ");
403 bch2_pr_time_units(out, ktime_get_real_ns() - r->wait_wallclock_start);
407 case BCH_REBALANCE_working:
408 bch2_move_stats_to_text(out, &r->work_stats);
410 case BCH_REBALANCE_scanning:
411 bch2_move_stats_to_text(out, &r->scan_stats);
415 printbuf_indent_sub(out, 2);
418 void bch2_rebalance_stop(struct bch_fs *c)
420 struct task_struct *p;
422 c->rebalance.pd.rate.rate = UINT_MAX;
423 bch2_ratelimit_reset(&c->rebalance.pd.rate);
425 p = rcu_dereference_protected(c->rebalance.thread, 1);
426 c->rebalance.thread = NULL;
429 /* for sychronizing with rebalance_wakeup() */
437 int bch2_rebalance_start(struct bch_fs *c)
439 struct task_struct *p;
442 if (c->rebalance.thread)
445 if (c->opts.nochanges)
448 p = kthread_create(bch2_rebalance_thread, c, "bch-rebalance/%s", c->name);
449 ret = PTR_ERR_OR_ZERO(p);
451 bch_err_msg(c, ret, "creating rebalance thread");
456 rcu_assign_pointer(c->rebalance.thread, p);
461 void bch2_fs_rebalance_init(struct bch_fs *c)
463 bch2_pd_controller_init(&c->rebalance.pd);