1 // SPDX-License-Identifier: GPL-2.0
2
3 #include "bcachefs.h"
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"
9 #include "buckets.h"
10 #include "clock.h"
11 #include "compress.h"
12 #include "disk_groups.h"
13 #include "errcode.h"
14 #include "error.h"
15 #include "inode.h"
16 #include "io_write.h"
17 #include "move.h"
18 #include "rebalance.h"
19 #include "subvolume.h"
20 #include "super-io.h"
21 #include "trace.h"
22
23 #include <linux/freezer.h>
24 #include <linux/kthread.h>
25 #include <linux/sched/cputime.h>
26
27 #define REBALANCE_WORK_SCAN_OFFSET (U64_MAX - 1)
28
29 static const char * const bch2_rebalance_state_strs[] = {
30 #define x(t) #t,
31 BCH_REBALANCE_STATES()
32 NULL
33 #undef x
34 };
35
__bch2_set_rebalance_needs_scan(struct btree_trans * trans,u64 inum)36 static int __bch2_set_rebalance_needs_scan(struct btree_trans *trans, u64 inum)
37 {
38 struct btree_iter iter;
39 struct bkey_s_c k;
40 struct bkey_i_cookie *cookie;
41 u64 v;
42 int ret;
43
44 bch2_trans_iter_init(trans, &iter, BTREE_ID_rebalance_work,
45 SPOS(inum, REBALANCE_WORK_SCAN_OFFSET, U32_MAX),
46 BTREE_ITER_intent);
47 k = bch2_btree_iter_peek_slot(&iter);
48 ret = bkey_err(k);
49 if (ret)
50 goto err;
51
52 v = k.k->type == KEY_TYPE_cookie
53 ? le64_to_cpu(bkey_s_c_to_cookie(k).v->cookie)
54 : 0;
55
56 cookie = bch2_trans_kmalloc(trans, sizeof(*cookie));
57 ret = PTR_ERR_OR_ZERO(cookie);
58 if (ret)
59 goto err;
60
61 bkey_cookie_init(&cookie->k_i);
62 cookie->k.p = iter.pos;
63 cookie->v.cookie = cpu_to_le64(v + 1);
64
65 ret = bch2_trans_update(trans, &iter, &cookie->k_i, 0);
66 err:
67 bch2_trans_iter_exit(trans, &iter);
68 return ret;
69 }
70
bch2_set_rebalance_needs_scan(struct bch_fs * c,u64 inum)71 int bch2_set_rebalance_needs_scan(struct bch_fs *c, u64 inum)
72 {
73 int ret = bch2_trans_commit_do(c, NULL, NULL,
74 BCH_TRANS_COMMIT_no_enospc|
75 BCH_TRANS_COMMIT_lazy_rw,
76 __bch2_set_rebalance_needs_scan(trans, inum));
77 rebalance_wakeup(c);
78 return ret;
79 }
80
bch2_set_fs_needs_rebalance(struct bch_fs * c)81 int bch2_set_fs_needs_rebalance(struct bch_fs *c)
82 {
83 return bch2_set_rebalance_needs_scan(c, 0);
84 }
85
bch2_clear_rebalance_needs_scan(struct btree_trans * trans,u64 inum,u64 cookie)86 static int bch2_clear_rebalance_needs_scan(struct btree_trans *trans, u64 inum, u64 cookie)
87 {
88 struct btree_iter iter;
89 struct bkey_s_c k;
90 u64 v;
91 int ret;
92
93 bch2_trans_iter_init(trans, &iter, BTREE_ID_rebalance_work,
94 SPOS(inum, REBALANCE_WORK_SCAN_OFFSET, U32_MAX),
95 BTREE_ITER_intent);
96 k = bch2_btree_iter_peek_slot(&iter);
97 ret = bkey_err(k);
98 if (ret)
99 goto err;
100
101 v = k.k->type == KEY_TYPE_cookie
102 ? le64_to_cpu(bkey_s_c_to_cookie(k).v->cookie)
103 : 0;
104
105 if (v == cookie)
106 ret = bch2_btree_delete_at(trans, &iter, 0);
107 err:
108 bch2_trans_iter_exit(trans, &iter);
109 return ret;
110 }
111
next_rebalance_entry(struct btree_trans * trans,struct btree_iter * work_iter)112 static struct bkey_s_c next_rebalance_entry(struct btree_trans *trans,
113 struct btree_iter *work_iter)
114 {
115 return !kthread_should_stop()
116 ? bch2_btree_iter_peek(work_iter)
117 : bkey_s_c_null;
118 }
119
bch2_bkey_clear_needs_rebalance(struct btree_trans * trans,struct btree_iter * iter,struct bkey_s_c k)120 static int bch2_bkey_clear_needs_rebalance(struct btree_trans *trans,
121 struct btree_iter *iter,
122 struct bkey_s_c k)
123 {
124 struct bkey_i *n = bch2_bkey_make_mut(trans, iter, &k, 0);
125 int ret = PTR_ERR_OR_ZERO(n);
126 if (ret)
127 return ret;
128
129 extent_entry_drop(bkey_i_to_s(n),
130 (void *) bch2_bkey_rebalance_opts(bkey_i_to_s_c(n)));
131 return bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc);
132 }
133
next_rebalance_extent(struct btree_trans * trans,struct bpos work_pos,struct btree_iter * extent_iter,struct data_update_opts * data_opts)134 static struct bkey_s_c next_rebalance_extent(struct btree_trans *trans,
135 struct bpos work_pos,
136 struct btree_iter *extent_iter,
137 struct data_update_opts *data_opts)
138 {
139 struct bch_fs *c = trans->c;
140 struct bkey_s_c k;
141
142 bch2_trans_iter_exit(trans, extent_iter);
143 bch2_trans_iter_init(trans, extent_iter,
144 work_pos.inode ? BTREE_ID_extents : BTREE_ID_reflink,
145 work_pos,
146 BTREE_ITER_all_snapshots);
147 k = bch2_btree_iter_peek_slot(extent_iter);
148 if (bkey_err(k))
149 return k;
150
151 const struct bch_extent_rebalance *r = k.k ? bch2_bkey_rebalance_opts(k) : NULL;
152 if (!r) {
153 /* raced due to btree write buffer, nothing to do */
154 return bkey_s_c_null;
155 }
156
157 memset(data_opts, 0, sizeof(*data_opts));
158
159 data_opts->rewrite_ptrs =
160 bch2_bkey_ptrs_need_rebalance(c, k, r->target, r->compression);
161 data_opts->target = r->target;
162 data_opts->write_flags |= BCH_WRITE_ONLY_SPECIFIED_DEVS;
163
164 if (!data_opts->rewrite_ptrs) {
165 /*
166 * device we would want to write to offline? devices in target
167 * changed?
168 *
169 * We'll now need a full scan before this extent is picked up
170 * again:
171 */
172 int ret = bch2_bkey_clear_needs_rebalance(trans, extent_iter, k);
173 if (ret)
174 return bkey_s_c_err(ret);
175 return bkey_s_c_null;
176 }
177
178 if (trace_rebalance_extent_enabled()) {
179 struct printbuf buf = PRINTBUF;
180
181 prt_str(&buf, "target=");
182 bch2_target_to_text(&buf, c, r->target);
183 prt_str(&buf, " compression=");
184 bch2_compression_opt_to_text(&buf, r->compression);
185 prt_str(&buf, " ");
186 bch2_bkey_val_to_text(&buf, c, k);
187
188 trace_rebalance_extent(c, buf.buf);
189 printbuf_exit(&buf);
190 }
191
192 return k;
193 }
194
195 noinline_for_stack
do_rebalance_extent(struct moving_context * ctxt,struct bpos work_pos,struct btree_iter * extent_iter)196 static int do_rebalance_extent(struct moving_context *ctxt,
197 struct bpos work_pos,
198 struct btree_iter *extent_iter)
199 {
200 struct btree_trans *trans = ctxt->trans;
201 struct bch_fs *c = trans->c;
202 struct bch_fs_rebalance *r = &trans->c->rebalance;
203 struct data_update_opts data_opts;
204 struct bch_io_opts io_opts;
205 struct bkey_s_c k;
206 struct bkey_buf sk;
207 int ret;
208
209 ctxt->stats = &r->work_stats;
210 r->state = BCH_REBALANCE_working;
211
212 bch2_bkey_buf_init(&sk);
213
214 ret = bkey_err(k = next_rebalance_extent(trans, work_pos,
215 extent_iter, &data_opts));
216 if (ret || !k.k)
217 goto out;
218
219 ret = bch2_move_get_io_opts_one(trans, &io_opts, k);
220 if (ret)
221 goto out;
222
223 atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
224
225 /*
226 * The iterator gets unlocked by __bch2_read_extent - need to
227 * save a copy of @k elsewhere:
228 */
229 bch2_bkey_buf_reassemble(&sk, c, k);
230 k = bkey_i_to_s_c(sk.k);
231
232 ret = bch2_move_extent(ctxt, NULL, extent_iter, k, io_opts, data_opts);
233 if (ret) {
234 if (bch2_err_matches(ret, ENOMEM)) {
235 /* memory allocation failure, wait for some IO to finish */
236 bch2_move_ctxt_wait_for_io(ctxt);
237 ret = -BCH_ERR_transaction_restart_nested;
238 }
239
240 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
241 goto out;
242
243 /* skip it and continue, XXX signal failure */
244 ret = 0;
245 }
246 out:
247 bch2_bkey_buf_exit(&sk, c);
248 return ret;
249 }
250
rebalance_pred(struct bch_fs * c,void * arg,struct bkey_s_c k,struct bch_io_opts * io_opts,struct data_update_opts * data_opts)251 static bool rebalance_pred(struct bch_fs *c, void *arg,
252 struct bkey_s_c k,
253 struct bch_io_opts *io_opts,
254 struct data_update_opts *data_opts)
255 {
256 unsigned target, compression;
257
258 if (k.k->p.inode) {
259 target = io_opts->background_target;
260 compression = background_compression(*io_opts);
261 } else {
262 const struct bch_extent_rebalance *r = bch2_bkey_rebalance_opts(k);
263
264 target = r ? r->target : io_opts->background_target;
265 compression = r ? r->compression : background_compression(*io_opts);
266 }
267
268 data_opts->rewrite_ptrs = bch2_bkey_ptrs_need_rebalance(c, k, target, compression);
269 data_opts->target = target;
270 data_opts->write_flags |= BCH_WRITE_ONLY_SPECIFIED_DEVS;
271 return data_opts->rewrite_ptrs != 0;
272 }
273
do_rebalance_scan(struct moving_context * ctxt,u64 inum,u64 cookie)274 static int do_rebalance_scan(struct moving_context *ctxt, u64 inum, u64 cookie)
275 {
276 struct btree_trans *trans = ctxt->trans;
277 struct bch_fs_rebalance *r = &trans->c->rebalance;
278 int ret;
279
280 bch2_move_stats_init(&r->scan_stats, "rebalance_scan");
281 ctxt->stats = &r->scan_stats;
282
283 if (!inum) {
284 r->scan_start = BBPOS_MIN;
285 r->scan_end = BBPOS_MAX;
286 } else {
287 r->scan_start = BBPOS(BTREE_ID_extents, POS(inum, 0));
288 r->scan_end = BBPOS(BTREE_ID_extents, POS(inum, U64_MAX));
289 }
290
291 r->state = BCH_REBALANCE_scanning;
292
293 ret = __bch2_move_data(ctxt, r->scan_start, r->scan_end, rebalance_pred, NULL) ?:
294 commit_do(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
295 bch2_clear_rebalance_needs_scan(trans, inum, cookie));
296
297 bch2_move_stats_exit(&r->scan_stats, trans->c);
298 return ret;
299 }
300
rebalance_wait(struct bch_fs * c)301 static void rebalance_wait(struct bch_fs *c)
302 {
303 struct bch_fs_rebalance *r = &c->rebalance;
304 struct io_clock *clock = &c->io_clock[WRITE];
305 u64 now = atomic64_read(&clock->now);
306 u64 min_member_capacity = bch2_min_rw_member_capacity(c);
307
308 if (min_member_capacity == U64_MAX)
309 min_member_capacity = 128 * 2048;
310
311 r->wait_iotime_end = now + (min_member_capacity >> 6);
312
313 if (r->state != BCH_REBALANCE_waiting) {
314 r->wait_iotime_start = now;
315 r->wait_wallclock_start = ktime_get_real_ns();
316 r->state = BCH_REBALANCE_waiting;
317 }
318
319 bch2_kthread_io_clock_wait(clock, r->wait_iotime_end, MAX_SCHEDULE_TIMEOUT);
320 }
321
do_rebalance(struct moving_context * ctxt)322 static int do_rebalance(struct moving_context *ctxt)
323 {
324 struct btree_trans *trans = ctxt->trans;
325 struct bch_fs *c = trans->c;
326 struct bch_fs_rebalance *r = &c->rebalance;
327 struct btree_iter rebalance_work_iter, extent_iter = { NULL };
328 struct bkey_s_c k;
329 int ret = 0;
330
331 bch2_trans_begin(trans);
332
333 bch2_move_stats_init(&r->work_stats, "rebalance_work");
334 bch2_move_stats_init(&r->scan_stats, "rebalance_scan");
335
336 bch2_trans_iter_init(trans, &rebalance_work_iter,
337 BTREE_ID_rebalance_work, POS_MIN,
338 BTREE_ITER_all_snapshots);
339
340 while (!bch2_move_ratelimit(ctxt)) {
341 if (!r->enabled) {
342 bch2_moving_ctxt_flush_all(ctxt);
343 kthread_wait_freezable(r->enabled ||
344 kthread_should_stop());
345 }
346
347 if (kthread_should_stop())
348 break;
349
350 bch2_trans_begin(trans);
351
352 ret = bkey_err(k = next_rebalance_entry(trans, &rebalance_work_iter));
353 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
354 continue;
355 if (ret || !k.k)
356 break;
357
358 ret = k.k->type == KEY_TYPE_cookie
359 ? do_rebalance_scan(ctxt, k.k->p.inode,
360 le64_to_cpu(bkey_s_c_to_cookie(k).v->cookie))
361 : do_rebalance_extent(ctxt, k.k->p, &extent_iter);
362
363 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
364 continue;
365 if (ret)
366 break;
367
368 bch2_btree_iter_advance(&rebalance_work_iter);
369 }
370
371 bch2_trans_iter_exit(trans, &extent_iter);
372 bch2_trans_iter_exit(trans, &rebalance_work_iter);
373 bch2_move_stats_exit(&r->scan_stats, c);
374
375 if (!ret &&
376 !kthread_should_stop() &&
377 !atomic64_read(&r->work_stats.sectors_seen) &&
378 !atomic64_read(&r->scan_stats.sectors_seen)) {
379 bch2_moving_ctxt_flush_all(ctxt);
380 bch2_trans_unlock_long(trans);
381 rebalance_wait(c);
382 }
383
384 if (!bch2_err_matches(ret, EROFS))
385 bch_err_fn(c, ret);
386 return ret;
387 }
388
bch2_rebalance_thread(void * arg)389 static int bch2_rebalance_thread(void *arg)
390 {
391 struct bch_fs *c = arg;
392 struct bch_fs_rebalance *r = &c->rebalance;
393 struct moving_context ctxt;
394
395 set_freezable();
396
397 bch2_moving_ctxt_init(&ctxt, c, NULL, &r->work_stats,
398 writepoint_ptr(&c->rebalance_write_point),
399 true);
400
401 while (!kthread_should_stop() && !do_rebalance(&ctxt))
402 ;
403
404 bch2_moving_ctxt_exit(&ctxt);
405
406 return 0;
407 }
408
bch2_rebalance_status_to_text(struct printbuf * out,struct bch_fs * c)409 void bch2_rebalance_status_to_text(struct printbuf *out, struct bch_fs *c)
410 {
411 struct bch_fs_rebalance *r = &c->rebalance;
412
413 prt_str(out, bch2_rebalance_state_strs[r->state]);
414 prt_newline(out);
415 printbuf_indent_add(out, 2);
416
417 switch (r->state) {
418 case BCH_REBALANCE_waiting: {
419 u64 now = atomic64_read(&c->io_clock[WRITE].now);
420
421 prt_str(out, "io wait duration: ");
422 bch2_prt_human_readable_s64(out, (r->wait_iotime_end - r->wait_iotime_start) << 9);
423 prt_newline(out);
424
425 prt_str(out, "io wait remaining: ");
426 bch2_prt_human_readable_s64(out, (r->wait_iotime_end - now) << 9);
427 prt_newline(out);
428
429 prt_str(out, "duration waited: ");
430 bch2_pr_time_units(out, ktime_get_real_ns() - r->wait_wallclock_start);
431 prt_newline(out);
432 break;
433 }
434 case BCH_REBALANCE_working:
435 bch2_move_stats_to_text(out, &r->work_stats);
436 break;
437 case BCH_REBALANCE_scanning:
438 bch2_move_stats_to_text(out, &r->scan_stats);
439 break;
440 }
441 prt_newline(out);
442 printbuf_indent_sub(out, 2);
443 }
444
bch2_rebalance_stop(struct bch_fs * c)445 void bch2_rebalance_stop(struct bch_fs *c)
446 {
447 struct task_struct *p;
448
449 c->rebalance.pd.rate.rate = UINT_MAX;
450 bch2_ratelimit_reset(&c->rebalance.pd.rate);
451
452 p = rcu_dereference_protected(c->rebalance.thread, 1);
453 c->rebalance.thread = NULL;
454
455 if (p) {
456 /* for sychronizing with rebalance_wakeup() */
457 synchronize_rcu();
458
459 kthread_stop(p);
460 put_task_struct(p);
461 }
462 }
463
bch2_rebalance_start(struct bch_fs * c)464 int bch2_rebalance_start(struct bch_fs *c)
465 {
466 struct task_struct *p;
467 int ret;
468
469 if (c->rebalance.thread)
470 return 0;
471
472 if (c->opts.nochanges)
473 return 0;
474
475 p = kthread_create(bch2_rebalance_thread, c, "bch-rebalance/%s", c->name);
476 ret = PTR_ERR_OR_ZERO(p);
477 bch_err_msg(c, ret, "creating rebalance thread");
478 if (ret)
479 return ret;
480
481 get_task_struct(p);
482 rcu_assign_pointer(c->rebalance.thread, p);
483 wake_up_process(p);
484 return 0;
485 }
486
bch2_fs_rebalance_init(struct bch_fs * c)487 void bch2_fs_rebalance_init(struct bch_fs *c)
488 {
489 bch2_pd_controller_init(&c->rebalance.pd);
490 }
491