xref: /linux/fs/bcachefs/move.c (revision 7d7f794482b74e39d8c0cd830333eb40fc0234d4)
1 // SPDX-License-Identifier: GPL-2.0
2 
3 #include "bcachefs.h"
4 #include "alloc_background.h"
5 #include "alloc_foreground.h"
6 #include "backpointers.h"
7 #include "bkey_buf.h"
8 #include "btree_gc.h"
9 #include "btree_update.h"
10 #include "btree_update_interior.h"
11 #include "btree_write_buffer.h"
12 #include "disk_groups.h"
13 #include "ec.h"
14 #include "errcode.h"
15 #include "error.h"
16 #include "inode.h"
17 #include "io_read.h"
18 #include "io_write.h"
19 #include "journal_reclaim.h"
20 #include "keylist.h"
21 #include "move.h"
22 #include "replicas.h"
23 #include "snapshot.h"
24 #include "super-io.h"
25 #include "trace.h"
26 
27 #include <linux/ioprio.h>
28 #include <linux/kthread.h>
29 
30 static void trace_move_extent2(struct bch_fs *c, struct bkey_s_c k)
31 {
32 	if (trace_move_extent_enabled()) {
33 		struct printbuf buf = PRINTBUF;
34 
35 		bch2_bkey_val_to_text(&buf, c, k);
36 		trace_move_extent(c, buf.buf);
37 		printbuf_exit(&buf);
38 	}
39 }
40 
41 static void trace_move_extent_read2(struct bch_fs *c, struct bkey_s_c k)
42 {
43 	if (trace_move_extent_read_enabled()) {
44 		struct printbuf buf = PRINTBUF;
45 
46 		bch2_bkey_val_to_text(&buf, c, k);
47 		trace_move_extent_read(c, buf.buf);
48 		printbuf_exit(&buf);
49 	}
50 }
51 
52 static void trace_move_extent_alloc_mem_fail2(struct bch_fs *c, struct bkey_s_c k)
53 {
54 	if (trace_move_extent_alloc_mem_fail_enabled()) {
55 		struct printbuf buf = PRINTBUF;
56 
57 		bch2_bkey_val_to_text(&buf, c, k);
58 		trace_move_extent_alloc_mem_fail(c, buf.buf);
59 		printbuf_exit(&buf);
60 	}
61 }
62 
63 struct moving_io {
64 	struct list_head		read_list;
65 	struct list_head		io_list;
66 	struct move_bucket_in_flight	*b;
67 	struct closure			cl;
68 	bool				read_completed;
69 
70 	unsigned			read_sectors;
71 	unsigned			write_sectors;
72 
73 	struct bch_read_bio		rbio;
74 
75 	struct data_update		write;
76 	/* Must be last since it is variable size */
77 	struct bio_vec			bi_inline_vecs[0];
78 };
79 
80 static void move_free(struct moving_io *io)
81 {
82 	struct moving_context *ctxt = io->write.ctxt;
83 
84 	if (io->b)
85 		atomic_dec(&io->b->count);
86 
87 	bch2_data_update_exit(&io->write);
88 
89 	mutex_lock(&ctxt->lock);
90 	list_del(&io->io_list);
91 	wake_up(&ctxt->wait);
92 	mutex_unlock(&ctxt->lock);
93 
94 	kfree(io);
95 }
96 
97 static void move_write_done(struct bch_write_op *op)
98 {
99 	struct moving_io *io = container_of(op, struct moving_io, write.op);
100 	struct moving_context *ctxt = io->write.ctxt;
101 
102 	if (io->write.op.error)
103 		ctxt->write_error = true;
104 
105 	atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
106 	atomic_dec(&io->write.ctxt->write_ios);
107 	move_free(io);
108 	closure_put(&ctxt->cl);
109 }
110 
111 static void move_write(struct moving_io *io)
112 {
113 	if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
114 		move_free(io);
115 		return;
116 	}
117 
118 	closure_get(&io->write.ctxt->cl);
119 	atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
120 	atomic_inc(&io->write.ctxt->write_ios);
121 
122 	bch2_data_update_read_done(&io->write, io->rbio.pick.crc);
123 }
124 
125 struct moving_io *bch2_moving_ctxt_next_pending_write(struct moving_context *ctxt)
126 {
127 	struct moving_io *io =
128 		list_first_entry_or_null(&ctxt->reads, struct moving_io, read_list);
129 
130 	return io && io->read_completed ? io : NULL;
131 }
132 
133 static void move_read_endio(struct bio *bio)
134 {
135 	struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
136 	struct moving_context *ctxt = io->write.ctxt;
137 
138 	atomic_sub(io->read_sectors, &ctxt->read_sectors);
139 	atomic_dec(&ctxt->read_ios);
140 	io->read_completed = true;
141 
142 	wake_up(&ctxt->wait);
143 	closure_put(&ctxt->cl);
144 }
145 
146 void bch2_moving_ctxt_do_pending_writes(struct moving_context *ctxt)
147 {
148 	struct moving_io *io;
149 
150 	while ((io = bch2_moving_ctxt_next_pending_write(ctxt))) {
151 		bch2_trans_unlock_long(ctxt->trans);
152 		list_del(&io->read_list);
153 		move_write(io);
154 	}
155 }
156 
157 void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt)
158 {
159 	unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
160 
161 	move_ctxt_wait_event(ctxt,
162 		!atomic_read(&ctxt->write_sectors) ||
163 		atomic_read(&ctxt->write_sectors) != sectors_pending);
164 }
165 
166 void bch2_moving_ctxt_exit(struct moving_context *ctxt)
167 {
168 	struct bch_fs *c = ctxt->trans->c;
169 
170 	move_ctxt_wait_event(ctxt, list_empty(&ctxt->reads));
171 	closure_sync(&ctxt->cl);
172 
173 	EBUG_ON(atomic_read(&ctxt->write_sectors));
174 	EBUG_ON(atomic_read(&ctxt->write_ios));
175 	EBUG_ON(atomic_read(&ctxt->read_sectors));
176 	EBUG_ON(atomic_read(&ctxt->read_ios));
177 
178 	mutex_lock(&c->moving_context_lock);
179 	list_del(&ctxt->list);
180 	mutex_unlock(&c->moving_context_lock);
181 
182 	bch2_trans_put(ctxt->trans);
183 	memset(ctxt, 0, sizeof(*ctxt));
184 }
185 
186 void bch2_moving_ctxt_init(struct moving_context *ctxt,
187 			   struct bch_fs *c,
188 			   struct bch_ratelimit *rate,
189 			   struct bch_move_stats *stats,
190 			   struct write_point_specifier wp,
191 			   bool wait_on_copygc)
192 {
193 	memset(ctxt, 0, sizeof(*ctxt));
194 
195 	ctxt->trans	= bch2_trans_get(c);
196 	ctxt->fn	= (void *) _RET_IP_;
197 	ctxt->rate	= rate;
198 	ctxt->stats	= stats;
199 	ctxt->wp	= wp;
200 	ctxt->wait_on_copygc = wait_on_copygc;
201 
202 	closure_init_stack(&ctxt->cl);
203 
204 	mutex_init(&ctxt->lock);
205 	INIT_LIST_HEAD(&ctxt->reads);
206 	INIT_LIST_HEAD(&ctxt->ios);
207 	init_waitqueue_head(&ctxt->wait);
208 
209 	mutex_lock(&c->moving_context_lock);
210 	list_add(&ctxt->list, &c->moving_context_list);
211 	mutex_unlock(&c->moving_context_lock);
212 }
213 
214 void bch2_move_stats_exit(struct bch_move_stats *stats, struct bch_fs *c)
215 {
216 	trace_move_data(c, stats);
217 }
218 
219 void bch2_move_stats_init(struct bch_move_stats *stats, char *name)
220 {
221 	memset(stats, 0, sizeof(*stats));
222 	stats->data_type = BCH_DATA_user;
223 	scnprintf(stats->name, sizeof(stats->name), "%s", name);
224 }
225 
226 static int bch2_extent_drop_ptrs(struct btree_trans *trans,
227 				 struct btree_iter *iter,
228 				 struct bkey_s_c k,
229 				 struct data_update_opts data_opts)
230 {
231 	struct bch_fs *c = trans->c;
232 	struct bkey_i *n;
233 	int ret;
234 
235 	n = bch2_bkey_make_mut_noupdate(trans, k);
236 	ret = PTR_ERR_OR_ZERO(n);
237 	if (ret)
238 		return ret;
239 
240 	while (data_opts.kill_ptrs) {
241 		unsigned i = 0, drop = __fls(data_opts.kill_ptrs);
242 		struct bch_extent_ptr *ptr;
243 
244 		bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, i++ == drop);
245 		data_opts.kill_ptrs ^= 1U << drop;
246 	}
247 
248 	/*
249 	 * If the new extent no longer has any pointers, bch2_extent_normalize()
250 	 * will do the appropriate thing with it (turning it into a
251 	 * KEY_TYPE_error key, or just a discard if it was a cached extent)
252 	 */
253 	bch2_extent_normalize(c, bkey_i_to_s(n));
254 
255 	/*
256 	 * Since we're not inserting through an extent iterator
257 	 * (BTREE_ITER_ALL_SNAPSHOTS iterators aren't extent iterators),
258 	 * we aren't using the extent overwrite path to delete, we're
259 	 * just using the normal key deletion path:
260 	 */
261 	if (bkey_deleted(&n->k))
262 		n->k.size = 0;
263 
264 	return bch2_trans_relock(trans) ?:
265 		bch2_trans_update(trans, iter, n, BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
266 		bch2_trans_commit(trans, NULL, NULL, BTREE_INSERT_NOFAIL);
267 }
268 
269 int bch2_move_extent(struct moving_context *ctxt,
270 		     struct move_bucket_in_flight *bucket_in_flight,
271 		     struct btree_iter *iter,
272 		     struct bkey_s_c k,
273 		     struct bch_io_opts io_opts,
274 		     struct data_update_opts data_opts)
275 {
276 	struct btree_trans *trans = ctxt->trans;
277 	struct bch_fs *c = trans->c;
278 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
279 	struct moving_io *io;
280 	const union bch_extent_entry *entry;
281 	struct extent_ptr_decoded p;
282 	unsigned sectors = k.k->size, pages;
283 	int ret = -ENOMEM;
284 
285 	if (ctxt->stats)
286 		ctxt->stats->pos = BBPOS(iter->btree_id, iter->pos);
287 	trace_move_extent2(c, k);
288 
289 	bch2_data_update_opts_normalize(k, &data_opts);
290 
291 	if (!data_opts.rewrite_ptrs &&
292 	    !data_opts.extra_replicas) {
293 		if (data_opts.kill_ptrs)
294 			return bch2_extent_drop_ptrs(trans, iter, k, data_opts);
295 		return 0;
296 	}
297 
298 	/*
299 	 * Before memory allocations & taking nocow locks in
300 	 * bch2_data_update_init():
301 	 */
302 	bch2_trans_unlock(trans);
303 
304 	/* write path might have to decompress data: */
305 	bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
306 		sectors = max_t(unsigned, sectors, p.crc.uncompressed_size);
307 
308 	pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
309 	io = kzalloc(sizeof(struct moving_io) +
310 		     sizeof(struct bio_vec) * pages, GFP_KERNEL);
311 	if (!io)
312 		goto err;
313 
314 	INIT_LIST_HEAD(&io->io_list);
315 	io->write.ctxt		= ctxt;
316 	io->read_sectors	= k.k->size;
317 	io->write_sectors	= k.k->size;
318 
319 	bio_init(&io->write.op.wbio.bio, NULL, io->bi_inline_vecs, pages, 0);
320 	bio_set_prio(&io->write.op.wbio.bio,
321 		     IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
322 
323 	if (bch2_bio_alloc_pages(&io->write.op.wbio.bio, sectors << 9,
324 				 GFP_KERNEL))
325 		goto err_free;
326 
327 	io->rbio.c		= c;
328 	io->rbio.opts		= io_opts;
329 	bio_init(&io->rbio.bio, NULL, io->bi_inline_vecs, pages, 0);
330 	io->rbio.bio.bi_vcnt = pages;
331 	bio_set_prio(&io->rbio.bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
332 	io->rbio.bio.bi_iter.bi_size = sectors << 9;
333 
334 	io->rbio.bio.bi_opf		= REQ_OP_READ;
335 	io->rbio.bio.bi_iter.bi_sector	= bkey_start_offset(k.k);
336 	io->rbio.bio.bi_end_io		= move_read_endio;
337 
338 	ret = bch2_data_update_init(trans, ctxt, &io->write, ctxt->wp,
339 				    io_opts, data_opts, iter->btree_id, k);
340 	if (ret && ret != -BCH_ERR_unwritten_extent_update)
341 		goto err_free_pages;
342 
343 	if (ret == -BCH_ERR_unwritten_extent_update) {
344 		bch2_update_unwritten_extent(trans, &io->write);
345 		move_free(io);
346 		return 0;
347 	}
348 
349 	BUG_ON(ret);
350 
351 	io->write.op.end_io = move_write_done;
352 
353 	if (ctxt->rate)
354 		bch2_ratelimit_increment(ctxt->rate, k.k->size);
355 
356 	if (ctxt->stats) {
357 		atomic64_inc(&ctxt->stats->keys_moved);
358 		atomic64_add(k.k->size, &ctxt->stats->sectors_moved);
359 	}
360 
361 	if (bucket_in_flight) {
362 		io->b = bucket_in_flight;
363 		atomic_inc(&io->b->count);
364 	}
365 
366 	this_cpu_add(c->counters[BCH_COUNTER_io_move], k.k->size);
367 	this_cpu_add(c->counters[BCH_COUNTER_move_extent_read], k.k->size);
368 	trace_move_extent_read2(c, k);
369 
370 	mutex_lock(&ctxt->lock);
371 	atomic_add(io->read_sectors, &ctxt->read_sectors);
372 	atomic_inc(&ctxt->read_ios);
373 
374 	list_add_tail(&io->read_list, &ctxt->reads);
375 	list_add_tail(&io->io_list, &ctxt->ios);
376 	mutex_unlock(&ctxt->lock);
377 
378 	/*
379 	 * dropped by move_read_endio() - guards against use after free of
380 	 * ctxt when doing wakeup
381 	 */
382 	closure_get(&ctxt->cl);
383 	bch2_read_extent(trans, &io->rbio,
384 			 bkey_start_pos(k.k),
385 			 iter->btree_id, k, 0,
386 			 BCH_READ_NODECODE|
387 			 BCH_READ_LAST_FRAGMENT);
388 	return 0;
389 err_free_pages:
390 	bio_free_pages(&io->write.op.wbio.bio);
391 err_free:
392 	kfree(io);
393 err:
394 	this_cpu_inc(c->counters[BCH_COUNTER_move_extent_alloc_mem_fail]);
395 	trace_move_extent_alloc_mem_fail2(c, k);
396 	return ret;
397 }
398 
399 struct bch_io_opts *bch2_move_get_io_opts(struct btree_trans *trans,
400 			  struct per_snapshot_io_opts *io_opts,
401 			  struct bkey_s_c extent_k)
402 {
403 	struct bch_fs *c = trans->c;
404 	u32 restart_count = trans->restart_count;
405 	int ret = 0;
406 
407 	if (io_opts->cur_inum != extent_k.k->p.inode) {
408 		struct btree_iter iter;
409 		struct bkey_s_c k;
410 
411 		io_opts->d.nr = 0;
412 
413 		for_each_btree_key(trans, iter, BTREE_ID_inodes, POS(0, extent_k.k->p.inode),
414 				   BTREE_ITER_ALL_SNAPSHOTS, k, ret) {
415 			if (k.k->p.offset != extent_k.k->p.inode)
416 				break;
417 
418 			if (!bkey_is_inode(k.k))
419 				continue;
420 
421 			struct bch_inode_unpacked inode;
422 			BUG_ON(bch2_inode_unpack(k, &inode));
423 
424 			struct snapshot_io_opts_entry e = { .snapshot = k.k->p.snapshot };
425 			bch2_inode_opts_get(&e.io_opts, trans->c, &inode);
426 
427 			ret = darray_push(&io_opts->d, e);
428 			if (ret)
429 				break;
430 		}
431 		bch2_trans_iter_exit(trans, &iter);
432 		io_opts->cur_inum = extent_k.k->p.inode;
433 	}
434 
435 	ret = ret ?: trans_was_restarted(trans, restart_count);
436 	if (ret)
437 		return ERR_PTR(ret);
438 
439 	if (extent_k.k->p.snapshot) {
440 		struct snapshot_io_opts_entry *i;
441 		darray_for_each(io_opts->d, i)
442 			if (bch2_snapshot_is_ancestor(c, extent_k.k->p.snapshot, i->snapshot))
443 				return &i->io_opts;
444 	}
445 
446 	return &io_opts->fs_io_opts;
447 }
448 
449 int bch2_move_get_io_opts_one(struct btree_trans *trans,
450 			      struct bch_io_opts *io_opts,
451 			      struct bkey_s_c extent_k)
452 {
453 	struct btree_iter iter;
454 	struct bkey_s_c k;
455 	int ret;
456 
457 	/* reflink btree? */
458 	if (!extent_k.k->p.inode) {
459 		*io_opts = bch2_opts_to_inode_opts(trans->c->opts);
460 		return 0;
461 	}
462 
463 	k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_inodes,
464 			       SPOS(0, extent_k.k->p.inode, extent_k.k->p.snapshot),
465 			       BTREE_ITER_CACHED);
466 	ret = bkey_err(k);
467 	if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
468 		return ret;
469 
470 	if (!ret && bkey_is_inode(k.k)) {
471 		struct bch_inode_unpacked inode;
472 		bch2_inode_unpack(k, &inode);
473 		bch2_inode_opts_get(io_opts, trans->c, &inode);
474 	} else {
475 		*io_opts = bch2_opts_to_inode_opts(trans->c->opts);
476 	}
477 
478 	bch2_trans_iter_exit(trans, &iter);
479 	return 0;
480 }
481 
482 int bch2_move_ratelimit(struct moving_context *ctxt)
483 {
484 	struct bch_fs *c = ctxt->trans->c;
485 	u64 delay;
486 
487 	if (ctxt->wait_on_copygc && !c->copygc_running) {
488 		bch2_trans_unlock_long(ctxt->trans);
489 		wait_event_killable(c->copygc_running_wq,
490 				    !c->copygc_running ||
491 				    kthread_should_stop());
492 	}
493 
494 	do {
495 		delay = ctxt->rate ? bch2_ratelimit_delay(ctxt->rate) : 0;
496 
497 
498 		if (delay) {
499 			if (delay > HZ / 10)
500 				bch2_trans_unlock_long(ctxt->trans);
501 			else
502 				bch2_trans_unlock(ctxt->trans);
503 			set_current_state(TASK_INTERRUPTIBLE);
504 		}
505 
506 		if ((current->flags & PF_KTHREAD) && kthread_should_stop()) {
507 			__set_current_state(TASK_RUNNING);
508 			return 1;
509 		}
510 
511 		if (delay)
512 			schedule_timeout(delay);
513 
514 		if (unlikely(freezing(current))) {
515 			move_ctxt_wait_event(ctxt, list_empty(&ctxt->reads));
516 			try_to_freeze();
517 		}
518 	} while (delay);
519 
520 	/*
521 	 * XXX: these limits really ought to be per device, SSDs and hard drives
522 	 * will want different limits
523 	 */
524 	move_ctxt_wait_event(ctxt,
525 		atomic_read(&ctxt->write_sectors) < c->opts.move_bytes_in_flight >> 9 &&
526 		atomic_read(&ctxt->read_sectors) < c->opts.move_bytes_in_flight >> 9 &&
527 		atomic_read(&ctxt->write_ios) < c->opts.move_ios_in_flight &&
528 		atomic_read(&ctxt->read_ios) < c->opts.move_ios_in_flight);
529 
530 	return 0;
531 }
532 
533 static int bch2_move_data_btree(struct moving_context *ctxt,
534 				struct bpos start,
535 				struct bpos end,
536 				move_pred_fn pred, void *arg,
537 				enum btree_id btree_id)
538 {
539 	struct btree_trans *trans = ctxt->trans;
540 	struct bch_fs *c = trans->c;
541 	struct per_snapshot_io_opts snapshot_io_opts;
542 	struct bch_io_opts *io_opts;
543 	struct bkey_buf sk;
544 	struct btree_iter iter;
545 	struct bkey_s_c k;
546 	struct data_update_opts data_opts;
547 	int ret = 0, ret2;
548 
549 	per_snapshot_io_opts_init(&snapshot_io_opts, c);
550 	bch2_bkey_buf_init(&sk);
551 
552 	if (ctxt->stats) {
553 		ctxt->stats->data_type	= BCH_DATA_user;
554 		ctxt->stats->pos	= BBPOS(btree_id, start);
555 	}
556 
557 	bch2_trans_iter_init(trans, &iter, btree_id, start,
558 			     BTREE_ITER_PREFETCH|
559 			     BTREE_ITER_ALL_SNAPSHOTS);
560 
561 	if (ctxt->rate)
562 		bch2_ratelimit_reset(ctxt->rate);
563 
564 	while (!bch2_move_ratelimit(ctxt)) {
565 		bch2_trans_begin(trans);
566 
567 		k = bch2_btree_iter_peek(&iter);
568 		if (!k.k)
569 			break;
570 
571 		ret = bkey_err(k);
572 		if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
573 			continue;
574 		if (ret)
575 			break;
576 
577 		if (bkey_ge(bkey_start_pos(k.k), end))
578 			break;
579 
580 		if (ctxt->stats)
581 			ctxt->stats->pos = BBPOS(iter.btree_id, iter.pos);
582 
583 		if (!bkey_extent_is_direct_data(k.k))
584 			goto next_nondata;
585 
586 		io_opts = bch2_move_get_io_opts(trans, &snapshot_io_opts, k);
587 		ret = PTR_ERR_OR_ZERO(io_opts);
588 		if (ret)
589 			continue;
590 
591 		memset(&data_opts, 0, sizeof(data_opts));
592 		if (!pred(c, arg, k, io_opts, &data_opts))
593 			goto next;
594 
595 		/*
596 		 * The iterator gets unlocked by __bch2_read_extent - need to
597 		 * save a copy of @k elsewhere:
598 		 */
599 		bch2_bkey_buf_reassemble(&sk, c, k);
600 		k = bkey_i_to_s_c(sk.k);
601 
602 		ret2 = bch2_move_extent(ctxt, NULL, &iter, k, *io_opts, data_opts);
603 		if (ret2) {
604 			if (bch2_err_matches(ret2, BCH_ERR_transaction_restart))
605 				continue;
606 
607 			if (ret2 == -ENOMEM) {
608 				/* memory allocation failure, wait for some IO to finish */
609 				bch2_move_ctxt_wait_for_io(ctxt);
610 				continue;
611 			}
612 
613 			/* XXX signal failure */
614 			goto next;
615 		}
616 next:
617 		if (ctxt->stats)
618 			atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
619 next_nondata:
620 		bch2_btree_iter_advance(&iter);
621 	}
622 
623 	bch2_trans_iter_exit(trans, &iter);
624 	bch2_bkey_buf_exit(&sk, c);
625 	per_snapshot_io_opts_exit(&snapshot_io_opts);
626 
627 	return ret;
628 }
629 
630 int __bch2_move_data(struct moving_context *ctxt,
631 		     struct bbpos start,
632 		     struct bbpos end,
633 		     move_pred_fn pred, void *arg)
634 {
635 	struct bch_fs *c = ctxt->trans->c;
636 	enum btree_id id;
637 	int ret = 0;
638 
639 	for (id = start.btree;
640 	     id <= min_t(unsigned, end.btree, btree_id_nr_alive(c) - 1);
641 	     id++) {
642 		ctxt->stats->pos = BBPOS(id, POS_MIN);
643 
644 		if (!btree_type_has_ptrs(id) ||
645 		    !bch2_btree_id_root(c, id)->b)
646 			continue;
647 
648 		ret = bch2_move_data_btree(ctxt,
649 				       id == start.btree ? start.pos : POS_MIN,
650 				       id == end.btree   ? end.pos   : POS_MAX,
651 				       pred, arg, id);
652 		if (ret)
653 			break;
654 	}
655 
656 	return ret;
657 }
658 
659 int bch2_move_data(struct bch_fs *c,
660 		   struct bbpos start,
661 		   struct bbpos end,
662 		   struct bch_ratelimit *rate,
663 		   struct bch_move_stats *stats,
664 		   struct write_point_specifier wp,
665 		   bool wait_on_copygc,
666 		   move_pred_fn pred, void *arg)
667 {
668 
669 	struct moving_context ctxt;
670 	int ret;
671 
672 	bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
673 	ret = __bch2_move_data(&ctxt, start, end, pred, arg);
674 	bch2_moving_ctxt_exit(&ctxt);
675 
676 	return ret;
677 }
678 
679 int __bch2_evacuate_bucket(struct moving_context *ctxt,
680 			   struct move_bucket_in_flight *bucket_in_flight,
681 			   struct bpos bucket, int gen,
682 			   struct data_update_opts _data_opts)
683 {
684 	struct btree_trans *trans = ctxt->trans;
685 	struct bch_fs *c = trans->c;
686 	struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
687 	struct btree_iter iter;
688 	struct bkey_buf sk;
689 	struct bch_backpointer bp;
690 	struct bch_alloc_v4 a_convert;
691 	const struct bch_alloc_v4 *a;
692 	struct bkey_s_c k;
693 	struct data_update_opts data_opts;
694 	unsigned dirty_sectors, bucket_size;
695 	u64 fragmentation;
696 	struct bpos bp_pos = POS_MIN;
697 	int ret = 0;
698 
699 	trace_bucket_evacuate(c, &bucket);
700 
701 	bch2_bkey_buf_init(&sk);
702 
703 	/*
704 	 * We're not run in a context that handles transaction restarts:
705 	 */
706 	bch2_trans_begin(trans);
707 
708 	bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
709 			     bucket, BTREE_ITER_CACHED);
710 	ret = lockrestart_do(trans,
711 			bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
712 	bch2_trans_iter_exit(trans, &iter);
713 
714 	if (ret) {
715 		bch_err_msg(c, ret, "looking up alloc key");
716 		goto err;
717 	}
718 
719 	a = bch2_alloc_to_v4(k, &a_convert);
720 	dirty_sectors = a->dirty_sectors;
721 	bucket_size = bch_dev_bkey_exists(c, bucket.inode)->mi.bucket_size;
722 	fragmentation = a->fragmentation_lru;
723 
724 	ret = bch2_btree_write_buffer_flush(trans);
725 	if (ret) {
726 		bch_err_msg(c, ret, "flushing btree write buffer");
727 		goto err;
728 	}
729 
730 	while (!(ret = bch2_move_ratelimit(ctxt))) {
731 		bch2_trans_begin(trans);
732 
733 		ret = bch2_get_next_backpointer(trans, bucket, gen,
734 						&bp_pos, &bp,
735 						BTREE_ITER_CACHED);
736 		if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
737 			continue;
738 		if (ret)
739 			goto err;
740 		if (bkey_eq(bp_pos, POS_MAX))
741 			break;
742 
743 		if (!bp.level) {
744 			const struct bch_extent_ptr *ptr;
745 			unsigned i = 0;
746 
747 			k = bch2_backpointer_get_key(trans, &iter, bp_pos, bp, 0);
748 			ret = bkey_err(k);
749 			if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
750 				continue;
751 			if (ret)
752 				goto err;
753 			if (!k.k)
754 				goto next;
755 
756 			bch2_bkey_buf_reassemble(&sk, c, k);
757 			k = bkey_i_to_s_c(sk.k);
758 
759 			ret = bch2_move_get_io_opts_one(trans, &io_opts, k);
760 			if (ret) {
761 				bch2_trans_iter_exit(trans, &iter);
762 				continue;
763 			}
764 
765 			data_opts = _data_opts;
766 			data_opts.target	= io_opts.background_target;
767 			data_opts.rewrite_ptrs = 0;
768 
769 			bkey_for_each_ptr(bch2_bkey_ptrs_c(k), ptr) {
770 				if (ptr->dev == bucket.inode) {
771 					data_opts.rewrite_ptrs |= 1U << i;
772 					if (ptr->cached) {
773 						bch2_trans_iter_exit(trans, &iter);
774 						goto next;
775 					}
776 				}
777 				i++;
778 			}
779 
780 			ret = bch2_move_extent(ctxt, bucket_in_flight,
781 					       &iter, k, io_opts, data_opts);
782 			bch2_trans_iter_exit(trans, &iter);
783 
784 			if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
785 				continue;
786 			if (ret == -ENOMEM) {
787 				/* memory allocation failure, wait for some IO to finish */
788 				bch2_move_ctxt_wait_for_io(ctxt);
789 				continue;
790 			}
791 			if (ret)
792 				goto err;
793 
794 			if (ctxt->stats)
795 				atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
796 		} else {
797 			struct btree *b;
798 
799 			b = bch2_backpointer_get_node(trans, &iter, bp_pos, bp);
800 			ret = PTR_ERR_OR_ZERO(b);
801 			if (ret == -BCH_ERR_backpointer_to_overwritten_btree_node)
802 				continue;
803 			if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
804 				continue;
805 			if (ret)
806 				goto err;
807 			if (!b)
808 				goto next;
809 
810 			ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
811 			bch2_trans_iter_exit(trans, &iter);
812 
813 			if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
814 				continue;
815 			if (ret)
816 				goto err;
817 
818 			if (ctxt->rate)
819 				bch2_ratelimit_increment(ctxt->rate,
820 							 c->opts.btree_node_size >> 9);
821 			if (ctxt->stats) {
822 				atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_seen);
823 				atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_moved);
824 			}
825 		}
826 next:
827 		bp_pos = bpos_nosnap_successor(bp_pos);
828 	}
829 
830 	trace_evacuate_bucket(c, &bucket, dirty_sectors, bucket_size, fragmentation, ret);
831 err:
832 	bch2_bkey_buf_exit(&sk, c);
833 	return ret;
834 }
835 
836 int bch2_evacuate_bucket(struct bch_fs *c,
837 			 struct bpos bucket, int gen,
838 			 struct data_update_opts data_opts,
839 			 struct bch_ratelimit *rate,
840 			 struct bch_move_stats *stats,
841 			 struct write_point_specifier wp,
842 			 bool wait_on_copygc)
843 {
844 	struct moving_context ctxt;
845 	int ret;
846 
847 	bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
848 	ret = __bch2_evacuate_bucket(&ctxt, NULL, bucket, gen, data_opts);
849 	bch2_moving_ctxt_exit(&ctxt);
850 
851 	return ret;
852 }
853 
854 typedef bool (*move_btree_pred)(struct bch_fs *, void *,
855 				struct btree *, struct bch_io_opts *,
856 				struct data_update_opts *);
857 
858 static int bch2_move_btree(struct bch_fs *c,
859 			   enum btree_id start_btree_id, struct bpos start_pos,
860 			   enum btree_id end_btree_id,   struct bpos end_pos,
861 			   move_btree_pred pred, void *arg,
862 			   struct bch_move_stats *stats)
863 {
864 	bool kthread = (current->flags & PF_KTHREAD) != 0;
865 	struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
866 	struct moving_context ctxt;
867 	struct btree_trans *trans;
868 	struct btree_iter iter;
869 	struct btree *b;
870 	enum btree_id id;
871 	struct data_update_opts data_opts;
872 	int ret = 0;
873 
874 	bch2_moving_ctxt_init(&ctxt, c, NULL, stats,
875 			      writepoint_ptr(&c->btree_write_point),
876 			      true);
877 	trans = ctxt.trans;
878 
879 	stats->data_type = BCH_DATA_btree;
880 
881 	for (id = start_btree_id;
882 	     id <= min_t(unsigned, end_btree_id, btree_id_nr_alive(c) - 1);
883 	     id++) {
884 		stats->pos = BBPOS(id, POS_MIN);
885 
886 		if (!bch2_btree_id_root(c, id)->b)
887 			continue;
888 
889 		bch2_trans_node_iter_init(trans, &iter, id, POS_MIN, 0, 0,
890 					  BTREE_ITER_PREFETCH);
891 retry:
892 		ret = 0;
893 		while (bch2_trans_begin(trans),
894 		       (b = bch2_btree_iter_peek_node(&iter)) &&
895 		       !(ret = PTR_ERR_OR_ZERO(b))) {
896 			if (kthread && kthread_should_stop())
897 				break;
898 
899 			if ((cmp_int(id, end_btree_id) ?:
900 			     bpos_cmp(b->key.k.p, end_pos)) > 0)
901 				break;
902 
903 			stats->pos = BBPOS(iter.btree_id, iter.pos);
904 
905 			if (!pred(c, arg, b, &io_opts, &data_opts))
906 				goto next;
907 
908 			ret = bch2_btree_node_rewrite(trans, &iter, b, 0) ?: ret;
909 			if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
910 				continue;
911 			if (ret)
912 				break;
913 next:
914 			bch2_btree_iter_next_node(&iter);
915 		}
916 		if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
917 			goto retry;
918 
919 		bch2_trans_iter_exit(trans, &iter);
920 
921 		if (kthread && kthread_should_stop())
922 			break;
923 	}
924 
925 	bch_err_fn(c, ret);
926 	bch2_moving_ctxt_exit(&ctxt);
927 	bch2_btree_interior_updates_flush(c);
928 
929 	return ret;
930 }
931 
932 static bool rereplicate_pred(struct bch_fs *c, void *arg,
933 			     struct bkey_s_c k,
934 			     struct bch_io_opts *io_opts,
935 			     struct data_update_opts *data_opts)
936 {
937 	unsigned nr_good = bch2_bkey_durability(c, k);
938 	unsigned replicas = bkey_is_btree_ptr(k.k)
939 		? c->opts.metadata_replicas
940 		: io_opts->data_replicas;
941 
942 	if (!nr_good || nr_good >= replicas)
943 		return false;
944 
945 	data_opts->target		= 0;
946 	data_opts->extra_replicas	= replicas - nr_good;
947 	data_opts->btree_insert_flags	= 0;
948 	return true;
949 }
950 
951 static bool migrate_pred(struct bch_fs *c, void *arg,
952 			 struct bkey_s_c k,
953 			 struct bch_io_opts *io_opts,
954 			 struct data_update_opts *data_opts)
955 {
956 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
957 	const struct bch_extent_ptr *ptr;
958 	struct bch_ioctl_data *op = arg;
959 	unsigned i = 0;
960 
961 	data_opts->rewrite_ptrs		= 0;
962 	data_opts->target		= 0;
963 	data_opts->extra_replicas	= 0;
964 	data_opts->btree_insert_flags	= 0;
965 
966 	bkey_for_each_ptr(ptrs, ptr) {
967 		if (ptr->dev == op->migrate.dev)
968 			data_opts->rewrite_ptrs |= 1U << i;
969 		i++;
970 	}
971 
972 	return data_opts->rewrite_ptrs != 0;
973 }
974 
975 static bool rereplicate_btree_pred(struct bch_fs *c, void *arg,
976 				   struct btree *b,
977 				   struct bch_io_opts *io_opts,
978 				   struct data_update_opts *data_opts)
979 {
980 	return rereplicate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
981 }
982 
983 static bool migrate_btree_pred(struct bch_fs *c, void *arg,
984 			       struct btree *b,
985 			       struct bch_io_opts *io_opts,
986 			       struct data_update_opts *data_opts)
987 {
988 	return migrate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
989 }
990 
991 static bool bformat_needs_redo(struct bkey_format *f)
992 {
993 	unsigned i;
994 
995 	for (i = 0; i < f->nr_fields; i++) {
996 		unsigned unpacked_bits = bch2_bkey_format_current.bits_per_field[i];
997 		u64 unpacked_mask = ~((~0ULL << 1) << (unpacked_bits - 1));
998 		u64 field_offset = le64_to_cpu(f->field_offset[i]);
999 
1000 		if (f->bits_per_field[i] > unpacked_bits)
1001 			return true;
1002 
1003 		if ((f->bits_per_field[i] == unpacked_bits) && field_offset)
1004 			return true;
1005 
1006 		if (((field_offset + ((1ULL << f->bits_per_field[i]) - 1)) &
1007 		     unpacked_mask) <
1008 		    field_offset)
1009 			return true;
1010 	}
1011 
1012 	return false;
1013 }
1014 
1015 static bool rewrite_old_nodes_pred(struct bch_fs *c, void *arg,
1016 				   struct btree *b,
1017 				   struct bch_io_opts *io_opts,
1018 				   struct data_update_opts *data_opts)
1019 {
1020 	if (b->version_ondisk != c->sb.version ||
1021 	    btree_node_need_rewrite(b) ||
1022 	    bformat_needs_redo(&b->format)) {
1023 		data_opts->target		= 0;
1024 		data_opts->extra_replicas	= 0;
1025 		data_opts->btree_insert_flags	= 0;
1026 		return true;
1027 	}
1028 
1029 	return false;
1030 }
1031 
1032 int bch2_scan_old_btree_nodes(struct bch_fs *c, struct bch_move_stats *stats)
1033 {
1034 	int ret;
1035 
1036 	ret = bch2_move_btree(c,
1037 			      0,		POS_MIN,
1038 			      BTREE_ID_NR,	SPOS_MAX,
1039 			      rewrite_old_nodes_pred, c, stats);
1040 	if (!ret) {
1041 		mutex_lock(&c->sb_lock);
1042 		c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_extents_above_btree_updates_done);
1043 		c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_bformat_overflow_done);
1044 		c->disk_sb.sb->version_min = c->disk_sb.sb->version;
1045 		bch2_write_super(c);
1046 		mutex_unlock(&c->sb_lock);
1047 	}
1048 
1049 	bch_err_fn(c, ret);
1050 	return ret;
1051 }
1052 
1053 int bch2_data_job(struct bch_fs *c,
1054 		  struct bch_move_stats *stats,
1055 		  struct bch_ioctl_data op)
1056 {
1057 	int ret = 0;
1058 
1059 	switch (op.op) {
1060 	case BCH_DATA_OP_REREPLICATE:
1061 		bch2_move_stats_init(stats, "rereplicate");
1062 		stats->data_type = BCH_DATA_journal;
1063 		ret = bch2_journal_flush_device_pins(&c->journal, -1);
1064 
1065 		ret = bch2_move_btree(c,
1066 				      op.start_btree,	op.start_pos,
1067 				      op.end_btree,	op.end_pos,
1068 				      rereplicate_btree_pred, c, stats) ?: ret;
1069 		ret = bch2_replicas_gc2(c) ?: ret;
1070 
1071 		ret = bch2_move_data(c,
1072 				     (struct bbpos) { op.start_btree,	op.start_pos },
1073 				     (struct bbpos) { op.end_btree,	op.end_pos },
1074 				     NULL,
1075 				     stats,
1076 				     writepoint_hashed((unsigned long) current),
1077 				     true,
1078 				     rereplicate_pred, c) ?: ret;
1079 		ret = bch2_replicas_gc2(c) ?: ret;
1080 
1081 		bch2_move_stats_exit(stats, c);
1082 		break;
1083 	case BCH_DATA_OP_MIGRATE:
1084 		if (op.migrate.dev >= c->sb.nr_devices)
1085 			return -EINVAL;
1086 
1087 		bch2_move_stats_init(stats, "migrate");
1088 		stats->data_type = BCH_DATA_journal;
1089 		ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
1090 
1091 		ret = bch2_move_btree(c,
1092 				      op.start_btree,	op.start_pos,
1093 				      op.end_btree,	op.end_pos,
1094 				      migrate_btree_pred, &op, stats) ?: ret;
1095 		ret = bch2_replicas_gc2(c) ?: ret;
1096 
1097 		ret = bch2_move_data(c,
1098 				     (struct bbpos) { op.start_btree,	op.start_pos },
1099 				     (struct bbpos) { op.end_btree,	op.end_pos },
1100 				     NULL,
1101 				     stats,
1102 				     writepoint_hashed((unsigned long) current),
1103 				     true,
1104 				     migrate_pred, &op) ?: ret;
1105 		ret = bch2_replicas_gc2(c) ?: ret;
1106 
1107 		bch2_move_stats_exit(stats, c);
1108 		break;
1109 	case BCH_DATA_OP_REWRITE_OLD_NODES:
1110 		bch2_move_stats_init(stats, "rewrite_old_nodes");
1111 		ret = bch2_scan_old_btree_nodes(c, stats);
1112 		bch2_move_stats_exit(stats, c);
1113 		break;
1114 	default:
1115 		ret = -EINVAL;
1116 	}
1117 
1118 	return ret;
1119 }
1120 
1121 void bch2_move_stats_to_text(struct printbuf *out, struct bch_move_stats *stats)
1122 {
1123 	prt_printf(out, "%s: data type=%s pos=",
1124 		   stats->name,
1125 		   bch2_data_types[stats->data_type]);
1126 	bch2_bbpos_to_text(out, stats->pos);
1127 	prt_newline(out);
1128 	printbuf_indent_add(out, 2);
1129 
1130 	prt_str(out, "keys moved:  ");
1131 	prt_u64(out, atomic64_read(&stats->keys_moved));
1132 	prt_newline(out);
1133 
1134 	prt_str(out, "keys raced:  ");
1135 	prt_u64(out, atomic64_read(&stats->keys_raced));
1136 	prt_newline(out);
1137 
1138 	prt_str(out, "bytes seen:  ");
1139 	prt_human_readable_u64(out, atomic64_read(&stats->sectors_seen) << 9);
1140 	prt_newline(out);
1141 
1142 	prt_str(out, "bytes moved: ");
1143 	prt_human_readable_u64(out, atomic64_read(&stats->sectors_moved) << 9);
1144 	prt_newline(out);
1145 
1146 	prt_str(out, "bytes raced: ");
1147 	prt_human_readable_u64(out, atomic64_read(&stats->sectors_raced) << 9);
1148 	prt_newline(out);
1149 
1150 	printbuf_indent_sub(out, 2);
1151 }
1152 
1153 static void bch2_moving_ctxt_to_text(struct printbuf *out, struct bch_fs *c, struct moving_context *ctxt)
1154 {
1155 	struct moving_io *io;
1156 
1157 	bch2_move_stats_to_text(out, ctxt->stats);
1158 	printbuf_indent_add(out, 2);
1159 
1160 	prt_printf(out, "reads: ios %u/%u sectors %u/%u",
1161 		   atomic_read(&ctxt->read_ios),
1162 		   c->opts.move_ios_in_flight,
1163 		   atomic_read(&ctxt->read_sectors),
1164 		   c->opts.move_bytes_in_flight >> 9);
1165 	prt_newline(out);
1166 
1167 	prt_printf(out, "writes: ios %u/%u sectors %u/%u",
1168 		   atomic_read(&ctxt->write_ios),
1169 		   c->opts.move_ios_in_flight,
1170 		   atomic_read(&ctxt->write_sectors),
1171 		   c->opts.move_bytes_in_flight >> 9);
1172 	prt_newline(out);
1173 
1174 	printbuf_indent_add(out, 2);
1175 
1176 	mutex_lock(&ctxt->lock);
1177 	list_for_each_entry(io, &ctxt->ios, io_list)
1178 		bch2_write_op_to_text(out, &io->write.op);
1179 	mutex_unlock(&ctxt->lock);
1180 
1181 	printbuf_indent_sub(out, 4);
1182 }
1183 
1184 void bch2_fs_moving_ctxts_to_text(struct printbuf *out, struct bch_fs *c)
1185 {
1186 	struct moving_context *ctxt;
1187 
1188 	mutex_lock(&c->moving_context_lock);
1189 	list_for_each_entry(ctxt, &c->moving_context_list, list)
1190 		bch2_moving_ctxt_to_text(out, c, ctxt);
1191 	mutex_unlock(&c->moving_context_lock);
1192 }
1193 
1194 void bch2_fs_move_init(struct bch_fs *c)
1195 {
1196 	INIT_LIST_HEAD(&c->moving_context_list);
1197 	mutex_init(&c->moving_context_lock);
1198 }
1199