xref: /linux/fs/bcachefs/move.c (revision 3f41368fbfe1b3d5922d317fe1a0a0cab6846802)
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_io.h"
10 #include "btree_update.h"
11 #include "btree_update_interior.h"
12 #include "btree_write_buffer.h"
13 #include "compress.h"
14 #include "disk_groups.h"
15 #include "ec.h"
16 #include "errcode.h"
17 #include "error.h"
18 #include "inode.h"
19 #include "io_read.h"
20 #include "io_write.h"
21 #include "journal_reclaim.h"
22 #include "keylist.h"
23 #include "move.h"
24 #include "replicas.h"
25 #include "snapshot.h"
26 #include "super-io.h"
27 #include "trace.h"
28 
29 #include <linux/ioprio.h>
30 #include <linux/kthread.h>
31 
32 const char * const bch2_data_ops_strs[] = {
33 #define x(t, n, ...) [n] = #t,
34 	BCH_DATA_OPS()
35 #undef x
36 	NULL
37 };
38 
39 static void bch2_data_update_opts_to_text(struct printbuf *out, struct bch_fs *c,
40 					  struct bch_io_opts *io_opts,
41 					  struct data_update_opts *data_opts)
42 {
43 	printbuf_tabstop_push(out, 20);
44 	prt_str(out, "rewrite ptrs:\t");
45 	bch2_prt_u64_base2(out, data_opts->rewrite_ptrs);
46 	prt_newline(out);
47 
48 	prt_str(out, "kill ptrs:\t");
49 	bch2_prt_u64_base2(out, data_opts->kill_ptrs);
50 	prt_newline(out);
51 
52 	prt_str(out, "target:\t");
53 	bch2_target_to_text(out, c, data_opts->target);
54 	prt_newline(out);
55 
56 	prt_str(out, "compression:\t");
57 	bch2_compression_opt_to_text(out, background_compression(*io_opts));
58 	prt_newline(out);
59 
60 	prt_str(out, "extra replicas:\t");
61 	prt_u64(out, data_opts->extra_replicas);
62 }
63 
64 static void trace_move_extent2(struct bch_fs *c, struct bkey_s_c k,
65 			       struct bch_io_opts *io_opts,
66 			       struct data_update_opts *data_opts)
67 {
68 	if (trace_move_extent_enabled()) {
69 		struct printbuf buf = PRINTBUF;
70 
71 		bch2_bkey_val_to_text(&buf, c, k);
72 		prt_newline(&buf);
73 		bch2_data_update_opts_to_text(&buf, c, io_opts, data_opts);
74 		trace_move_extent(c, buf.buf);
75 		printbuf_exit(&buf);
76 	}
77 }
78 
79 static void trace_move_extent_read2(struct bch_fs *c, struct bkey_s_c k)
80 {
81 	if (trace_move_extent_read_enabled()) {
82 		struct printbuf buf = PRINTBUF;
83 
84 		bch2_bkey_val_to_text(&buf, c, k);
85 		trace_move_extent_read(c, buf.buf);
86 		printbuf_exit(&buf);
87 	}
88 }
89 
90 struct moving_io {
91 	struct list_head		read_list;
92 	struct list_head		io_list;
93 	struct move_bucket_in_flight	*b;
94 	struct closure			cl;
95 	bool				read_completed;
96 
97 	unsigned			read_sectors;
98 	unsigned			write_sectors;
99 
100 	struct bch_read_bio		rbio;
101 
102 	struct data_update		write;
103 	/* Must be last since it is variable size */
104 	struct bio_vec			bi_inline_vecs[];
105 };
106 
107 static void move_free(struct moving_io *io)
108 {
109 	struct moving_context *ctxt = io->write.ctxt;
110 
111 	if (io->b)
112 		atomic_dec(&io->b->count);
113 
114 	bch2_data_update_exit(&io->write);
115 
116 	mutex_lock(&ctxt->lock);
117 	list_del(&io->io_list);
118 	wake_up(&ctxt->wait);
119 	mutex_unlock(&ctxt->lock);
120 
121 	kfree(io);
122 }
123 
124 static void move_write_done(struct bch_write_op *op)
125 {
126 	struct moving_io *io = container_of(op, struct moving_io, write.op);
127 	struct moving_context *ctxt = io->write.ctxt;
128 
129 	if (io->write.op.error)
130 		ctxt->write_error = true;
131 
132 	atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
133 	atomic_dec(&io->write.ctxt->write_ios);
134 	move_free(io);
135 	closure_put(&ctxt->cl);
136 }
137 
138 static void move_write(struct moving_io *io)
139 {
140 	if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
141 		move_free(io);
142 		return;
143 	}
144 
145 	if (trace_move_extent_write_enabled()) {
146 		struct bch_fs *c = io->write.op.c;
147 		struct printbuf buf = PRINTBUF;
148 
149 		bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(io->write.k.k));
150 		trace_move_extent_write(c, buf.buf);
151 		printbuf_exit(&buf);
152 	}
153 
154 	closure_get(&io->write.ctxt->cl);
155 	atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
156 	atomic_inc(&io->write.ctxt->write_ios);
157 
158 	bch2_data_update_read_done(&io->write, io->rbio.pick.crc);
159 }
160 
161 struct moving_io *bch2_moving_ctxt_next_pending_write(struct moving_context *ctxt)
162 {
163 	struct moving_io *io =
164 		list_first_entry_or_null(&ctxt->reads, struct moving_io, read_list);
165 
166 	return io && io->read_completed ? io : NULL;
167 }
168 
169 static void move_read_endio(struct bio *bio)
170 {
171 	struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
172 	struct moving_context *ctxt = io->write.ctxt;
173 
174 	atomic_sub(io->read_sectors, &ctxt->read_sectors);
175 	atomic_dec(&ctxt->read_ios);
176 	io->read_completed = true;
177 
178 	wake_up(&ctxt->wait);
179 	closure_put(&ctxt->cl);
180 }
181 
182 void bch2_moving_ctxt_do_pending_writes(struct moving_context *ctxt)
183 {
184 	struct moving_io *io;
185 
186 	while ((io = bch2_moving_ctxt_next_pending_write(ctxt))) {
187 		bch2_trans_unlock_long(ctxt->trans);
188 		list_del(&io->read_list);
189 		move_write(io);
190 	}
191 }
192 
193 void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt)
194 {
195 	unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
196 
197 	move_ctxt_wait_event(ctxt,
198 		!atomic_read(&ctxt->write_sectors) ||
199 		atomic_read(&ctxt->write_sectors) != sectors_pending);
200 }
201 
202 void bch2_moving_ctxt_flush_all(struct moving_context *ctxt)
203 {
204 	move_ctxt_wait_event(ctxt, list_empty(&ctxt->reads));
205 	bch2_trans_unlock_long(ctxt->trans);
206 	closure_sync(&ctxt->cl);
207 }
208 
209 void bch2_moving_ctxt_exit(struct moving_context *ctxt)
210 {
211 	struct bch_fs *c = ctxt->trans->c;
212 
213 	bch2_moving_ctxt_flush_all(ctxt);
214 
215 	EBUG_ON(atomic_read(&ctxt->write_sectors));
216 	EBUG_ON(atomic_read(&ctxt->write_ios));
217 	EBUG_ON(atomic_read(&ctxt->read_sectors));
218 	EBUG_ON(atomic_read(&ctxt->read_ios));
219 
220 	mutex_lock(&c->moving_context_lock);
221 	list_del(&ctxt->list);
222 	mutex_unlock(&c->moving_context_lock);
223 
224 	bch2_trans_put(ctxt->trans);
225 	memset(ctxt, 0, sizeof(*ctxt));
226 }
227 
228 void bch2_moving_ctxt_init(struct moving_context *ctxt,
229 			   struct bch_fs *c,
230 			   struct bch_ratelimit *rate,
231 			   struct bch_move_stats *stats,
232 			   struct write_point_specifier wp,
233 			   bool wait_on_copygc)
234 {
235 	memset(ctxt, 0, sizeof(*ctxt));
236 
237 	ctxt->trans	= bch2_trans_get(c);
238 	ctxt->fn	= (void *) _RET_IP_;
239 	ctxt->rate	= rate;
240 	ctxt->stats	= stats;
241 	ctxt->wp	= wp;
242 	ctxt->wait_on_copygc = wait_on_copygc;
243 
244 	closure_init_stack(&ctxt->cl);
245 
246 	mutex_init(&ctxt->lock);
247 	INIT_LIST_HEAD(&ctxt->reads);
248 	INIT_LIST_HEAD(&ctxt->ios);
249 	init_waitqueue_head(&ctxt->wait);
250 
251 	mutex_lock(&c->moving_context_lock);
252 	list_add(&ctxt->list, &c->moving_context_list);
253 	mutex_unlock(&c->moving_context_lock);
254 }
255 
256 void bch2_move_stats_exit(struct bch_move_stats *stats, struct bch_fs *c)
257 {
258 	trace_move_data(c, stats);
259 }
260 
261 void bch2_move_stats_init(struct bch_move_stats *stats, const char *name)
262 {
263 	memset(stats, 0, sizeof(*stats));
264 	stats->data_type = BCH_DATA_user;
265 	scnprintf(stats->name, sizeof(stats->name), "%s", name);
266 }
267 
268 int bch2_move_extent(struct moving_context *ctxt,
269 		     struct move_bucket_in_flight *bucket_in_flight,
270 		     struct btree_iter *iter,
271 		     struct bkey_s_c k,
272 		     struct bch_io_opts io_opts,
273 		     struct data_update_opts data_opts)
274 {
275 	struct btree_trans *trans = ctxt->trans;
276 	struct bch_fs *c = trans->c;
277 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
278 	struct moving_io *io;
279 	const union bch_extent_entry *entry;
280 	struct extent_ptr_decoded p;
281 	unsigned sectors = k.k->size, pages;
282 	int ret = -ENOMEM;
283 
284 	trace_move_extent2(c, k, &io_opts, &data_opts);
285 
286 	if (ctxt->stats)
287 		ctxt->stats->pos = BBPOS(iter->btree_id, iter->pos);
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, iter, ctxt, &io->write, ctxt->wp,
339 				    io_opts, data_opts, iter->btree_id, k);
340 	if (ret)
341 		goto err_free_pages;
342 
343 	io->write.op.end_io = move_write_done;
344 
345 	if (ctxt->rate)
346 		bch2_ratelimit_increment(ctxt->rate, k.k->size);
347 
348 	if (ctxt->stats) {
349 		atomic64_inc(&ctxt->stats->keys_moved);
350 		atomic64_add(k.k->size, &ctxt->stats->sectors_moved);
351 	}
352 
353 	if (bucket_in_flight) {
354 		io->b = bucket_in_flight;
355 		atomic_inc(&io->b->count);
356 	}
357 
358 	this_cpu_add(c->counters[BCH_COUNTER_io_move], k.k->size);
359 	this_cpu_add(c->counters[BCH_COUNTER_move_extent_read], k.k->size);
360 	trace_move_extent_read2(c, k);
361 
362 	mutex_lock(&ctxt->lock);
363 	atomic_add(io->read_sectors, &ctxt->read_sectors);
364 	atomic_inc(&ctxt->read_ios);
365 
366 	list_add_tail(&io->read_list, &ctxt->reads);
367 	list_add_tail(&io->io_list, &ctxt->ios);
368 	mutex_unlock(&ctxt->lock);
369 
370 	/*
371 	 * dropped by move_read_endio() - guards against use after free of
372 	 * ctxt when doing wakeup
373 	 */
374 	closure_get(&ctxt->cl);
375 	bch2_read_extent(trans, &io->rbio,
376 			 bkey_start_pos(k.k),
377 			 iter->btree_id, k, 0,
378 			 BCH_READ_NODECODE|
379 			 BCH_READ_LAST_FRAGMENT);
380 	return 0;
381 err_free_pages:
382 	bio_free_pages(&io->write.op.wbio.bio);
383 err_free:
384 	kfree(io);
385 err:
386 	if (ret == -BCH_ERR_data_update_done)
387 		return 0;
388 
389 	if (bch2_err_matches(ret, EROFS) ||
390 	    bch2_err_matches(ret, BCH_ERR_transaction_restart))
391 		return ret;
392 
393 	count_event(c, move_extent_start_fail);
394 
395 	if (trace_move_extent_start_fail_enabled()) {
396 		struct printbuf buf = PRINTBUF;
397 
398 		bch2_bkey_val_to_text(&buf, c, k);
399 		prt_str(&buf, ": ");
400 		prt_str(&buf, bch2_err_str(ret));
401 		trace_move_extent_start_fail(c, buf.buf);
402 		printbuf_exit(&buf);
403 	}
404 	return ret;
405 }
406 
407 struct bch_io_opts *bch2_move_get_io_opts(struct btree_trans *trans,
408 			  struct per_snapshot_io_opts *io_opts,
409 			  struct bkey_s_c extent_k)
410 {
411 	struct bch_fs *c = trans->c;
412 	u32 restart_count = trans->restart_count;
413 	int ret = 0;
414 
415 	if (io_opts->cur_inum != extent_k.k->p.inode) {
416 		io_opts->d.nr = 0;
417 
418 		ret = for_each_btree_key(trans, iter, BTREE_ID_inodes, POS(0, extent_k.k->p.inode),
419 					 BTREE_ITER_all_snapshots, k, ({
420 			if (k.k->p.offset != extent_k.k->p.inode)
421 				break;
422 
423 			if (!bkey_is_inode(k.k))
424 				continue;
425 
426 			struct bch_inode_unpacked inode;
427 			BUG_ON(bch2_inode_unpack(k, &inode));
428 
429 			struct snapshot_io_opts_entry e = { .snapshot = k.k->p.snapshot };
430 			bch2_inode_opts_get(&e.io_opts, trans->c, &inode);
431 
432 			darray_push(&io_opts->d, e);
433 		}));
434 		io_opts->cur_inum = extent_k.k->p.inode;
435 	}
436 
437 	ret = ret ?: trans_was_restarted(trans, restart_count);
438 	if (ret)
439 		return ERR_PTR(ret);
440 
441 	if (extent_k.k->p.snapshot)
442 		darray_for_each(io_opts->d, i)
443 			if (bch2_snapshot_is_ancestor(c, extent_k.k->p.snapshot, i->snapshot))
444 				return &i->io_opts;
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 	bool is_kthread = current->flags & PF_KTHREAD;
486 	u64 delay;
487 
488 	if (ctxt->wait_on_copygc && c->copygc_running) {
489 		bch2_moving_ctxt_flush_all(ctxt);
490 		wait_event_killable(c->copygc_running_wq,
491 				    !c->copygc_running ||
492 				    (is_kthread && kthread_should_stop()));
493 	}
494 
495 	do {
496 		delay = ctxt->rate ? bch2_ratelimit_delay(ctxt->rate) : 0;
497 
498 		if (is_kthread && kthread_should_stop())
499 			return 1;
500 
501 		if (delay)
502 			move_ctxt_wait_event_timeout(ctxt,
503 					freezing(current) ||
504 					(is_kthread && kthread_should_stop()),
505 					delay);
506 
507 		if (unlikely(freezing(current))) {
508 			bch2_moving_ctxt_flush_all(ctxt);
509 			try_to_freeze();
510 		}
511 	} while (delay);
512 
513 	/*
514 	 * XXX: these limits really ought to be per device, SSDs and hard drives
515 	 * will want different limits
516 	 */
517 	move_ctxt_wait_event(ctxt,
518 		atomic_read(&ctxt->write_sectors) < c->opts.move_bytes_in_flight >> 9 &&
519 		atomic_read(&ctxt->read_sectors) < c->opts.move_bytes_in_flight >> 9 &&
520 		atomic_read(&ctxt->write_ios) < c->opts.move_ios_in_flight &&
521 		atomic_read(&ctxt->read_ios) < c->opts.move_ios_in_flight);
522 
523 	return 0;
524 }
525 
526 static int bch2_move_data_btree(struct moving_context *ctxt,
527 				struct bpos start,
528 				struct bpos end,
529 				move_pred_fn pred, void *arg,
530 				enum btree_id btree_id)
531 {
532 	struct btree_trans *trans = ctxt->trans;
533 	struct bch_fs *c = trans->c;
534 	struct per_snapshot_io_opts snapshot_io_opts;
535 	struct bch_io_opts *io_opts;
536 	struct bkey_buf sk;
537 	struct btree_iter iter;
538 	struct bkey_s_c k;
539 	struct data_update_opts data_opts;
540 	int ret = 0, ret2;
541 
542 	per_snapshot_io_opts_init(&snapshot_io_opts, c);
543 	bch2_bkey_buf_init(&sk);
544 
545 	if (ctxt->stats) {
546 		ctxt->stats->data_type	= BCH_DATA_user;
547 		ctxt->stats->pos	= BBPOS(btree_id, start);
548 	}
549 
550 	bch2_trans_iter_init(trans, &iter, btree_id, start,
551 			     BTREE_ITER_prefetch|
552 			     BTREE_ITER_all_snapshots);
553 
554 	if (ctxt->rate)
555 		bch2_ratelimit_reset(ctxt->rate);
556 
557 	while (!bch2_move_ratelimit(ctxt)) {
558 		bch2_trans_begin(trans);
559 
560 		k = bch2_btree_iter_peek(&iter);
561 		if (!k.k)
562 			break;
563 
564 		ret = bkey_err(k);
565 		if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
566 			continue;
567 		if (ret)
568 			break;
569 
570 		if (bkey_ge(bkey_start_pos(k.k), end))
571 			break;
572 
573 		if (ctxt->stats)
574 			ctxt->stats->pos = BBPOS(iter.btree_id, iter.pos);
575 
576 		if (!bkey_extent_is_direct_data(k.k))
577 			goto next_nondata;
578 
579 		io_opts = bch2_move_get_io_opts(trans, &snapshot_io_opts, k);
580 		ret = PTR_ERR_OR_ZERO(io_opts);
581 		if (ret)
582 			continue;
583 
584 		memset(&data_opts, 0, sizeof(data_opts));
585 		if (!pred(c, arg, k, io_opts, &data_opts))
586 			goto next;
587 
588 		/*
589 		 * The iterator gets unlocked by __bch2_read_extent - need to
590 		 * save a copy of @k elsewhere:
591 		 */
592 		bch2_bkey_buf_reassemble(&sk, c, k);
593 		k = bkey_i_to_s_c(sk.k);
594 
595 		ret2 = bch2_move_extent(ctxt, NULL, &iter, k, *io_opts, data_opts);
596 		if (ret2) {
597 			if (bch2_err_matches(ret2, BCH_ERR_transaction_restart))
598 				continue;
599 
600 			if (ret2 == -ENOMEM) {
601 				/* memory allocation failure, wait for some IO to finish */
602 				bch2_move_ctxt_wait_for_io(ctxt);
603 				continue;
604 			}
605 
606 			/* XXX signal failure */
607 			goto next;
608 		}
609 next:
610 		if (ctxt->stats)
611 			atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
612 next_nondata:
613 		bch2_btree_iter_advance(&iter);
614 	}
615 
616 	bch2_trans_iter_exit(trans, &iter);
617 	bch2_bkey_buf_exit(&sk, c);
618 	per_snapshot_io_opts_exit(&snapshot_io_opts);
619 
620 	return ret;
621 }
622 
623 int __bch2_move_data(struct moving_context *ctxt,
624 		     struct bbpos start,
625 		     struct bbpos end,
626 		     move_pred_fn pred, void *arg)
627 {
628 	struct bch_fs *c = ctxt->trans->c;
629 	enum btree_id id;
630 	int ret = 0;
631 
632 	for (id = start.btree;
633 	     id <= min_t(unsigned, end.btree, btree_id_nr_alive(c) - 1);
634 	     id++) {
635 		ctxt->stats->pos = BBPOS(id, POS_MIN);
636 
637 		if (!btree_type_has_ptrs(id) ||
638 		    !bch2_btree_id_root(c, id)->b)
639 			continue;
640 
641 		ret = bch2_move_data_btree(ctxt,
642 				       id == start.btree ? start.pos : POS_MIN,
643 				       id == end.btree   ? end.pos   : POS_MAX,
644 				       pred, arg, id);
645 		if (ret)
646 			break;
647 	}
648 
649 	return ret;
650 }
651 
652 int bch2_move_data(struct bch_fs *c,
653 		   struct bbpos start,
654 		   struct bbpos end,
655 		   struct bch_ratelimit *rate,
656 		   struct bch_move_stats *stats,
657 		   struct write_point_specifier wp,
658 		   bool wait_on_copygc,
659 		   move_pred_fn pred, void *arg)
660 {
661 
662 	struct moving_context ctxt;
663 	int ret;
664 
665 	bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
666 	ret = __bch2_move_data(&ctxt, start, end, pred, arg);
667 	bch2_moving_ctxt_exit(&ctxt);
668 
669 	return ret;
670 }
671 
672 int bch2_evacuate_bucket(struct moving_context *ctxt,
673 			   struct move_bucket_in_flight *bucket_in_flight,
674 			   struct bpos bucket, int gen,
675 			   struct data_update_opts _data_opts)
676 {
677 	struct btree_trans *trans = ctxt->trans;
678 	struct bch_fs *c = trans->c;
679 	bool is_kthread = current->flags & PF_KTHREAD;
680 	struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
681 	struct btree_iter iter;
682 	struct bkey_buf sk;
683 	struct bch_backpointer bp;
684 	struct bch_alloc_v4 a_convert;
685 	const struct bch_alloc_v4 *a;
686 	struct bkey_s_c k;
687 	struct data_update_opts data_opts;
688 	unsigned dirty_sectors, bucket_size;
689 	u64 fragmentation;
690 	struct bpos bp_pos = POS_MIN;
691 	int ret = 0;
692 
693 	struct bch_dev *ca = bch2_dev_tryget(c, bucket.inode);
694 	if (!ca)
695 		return 0;
696 
697 	trace_bucket_evacuate(c, &bucket);
698 
699 	bch2_bkey_buf_init(&sk);
700 
701 	/*
702 	 * We're not run in a context that handles transaction restarts:
703 	 */
704 	bch2_trans_begin(trans);
705 
706 	bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
707 			     bucket, BTREE_ITER_cached);
708 	ret = lockrestart_do(trans,
709 			bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
710 	bch2_trans_iter_exit(trans, &iter);
711 
712 	bch_err_msg(c, ret, "looking up alloc key");
713 	if (ret)
714 		goto err;
715 
716 	a = bch2_alloc_to_v4(k, &a_convert);
717 	dirty_sectors = bch2_bucket_sectors_dirty(*a);
718 	bucket_size = ca->mi.bucket_size;
719 	fragmentation = a->fragmentation_lru;
720 
721 	ret = bch2_btree_write_buffer_tryflush(trans);
722 	bch_err_msg(c, ret, "flushing btree write buffer");
723 	if (ret)
724 		goto err;
725 
726 	while (!(ret = bch2_move_ratelimit(ctxt))) {
727 		if (is_kthread && kthread_should_stop())
728 			break;
729 
730 		bch2_trans_begin(trans);
731 
732 		ret = bch2_get_next_backpointer(trans, ca, bucket, gen,
733 						&bp_pos, &bp,
734 						BTREE_ITER_cached);
735 		if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
736 			continue;
737 		if (ret)
738 			goto err;
739 		if (bkey_eq(bp_pos, POS_MAX))
740 			break;
741 
742 		if (!bp.level) {
743 			k = bch2_backpointer_get_key(trans, &iter, bp_pos, bp, 0);
744 			ret = bkey_err(k);
745 			if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
746 				continue;
747 			if (ret)
748 				goto err;
749 			if (!k.k)
750 				goto next;
751 
752 			bch2_bkey_buf_reassemble(&sk, c, k);
753 			k = bkey_i_to_s_c(sk.k);
754 
755 			ret = bch2_move_get_io_opts_one(trans, &io_opts, k);
756 			if (ret) {
757 				bch2_trans_iter_exit(trans, &iter);
758 				continue;
759 			}
760 
761 			data_opts = _data_opts;
762 			data_opts.target	= io_opts.background_target;
763 			data_opts.rewrite_ptrs = 0;
764 
765 			unsigned i = 0;
766 			bkey_for_each_ptr(bch2_bkey_ptrs_c(k), ptr) {
767 				if (ptr->dev == bucket.inode) {
768 					data_opts.rewrite_ptrs |= 1U << i;
769 					if (ptr->cached) {
770 						bch2_trans_iter_exit(trans, &iter);
771 						goto next;
772 					}
773 				}
774 				i++;
775 			}
776 
777 			ret = bch2_move_extent(ctxt, bucket_in_flight,
778 					       &iter, k, io_opts, data_opts);
779 			bch2_trans_iter_exit(trans, &iter);
780 
781 			if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
782 				continue;
783 			if (ret == -ENOMEM) {
784 				/* memory allocation failure, wait for some IO to finish */
785 				bch2_move_ctxt_wait_for_io(ctxt);
786 				continue;
787 			}
788 			if (ret)
789 				goto err;
790 
791 			if (ctxt->stats)
792 				atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
793 		} else {
794 			struct btree *b;
795 
796 			b = bch2_backpointer_get_node(trans, &iter, bp_pos, bp);
797 			ret = PTR_ERR_OR_ZERO(b);
798 			if (ret == -BCH_ERR_backpointer_to_overwritten_btree_node)
799 				continue;
800 			if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
801 				continue;
802 			if (ret)
803 				goto err;
804 			if (!b)
805 				goto next;
806 
807 			unsigned sectors = btree_ptr_sectors_written(&b->key);
808 
809 			ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
810 			bch2_trans_iter_exit(trans, &iter);
811 
812 			if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
813 				continue;
814 			if (ret)
815 				goto err;
816 
817 			if (ctxt->rate)
818 				bch2_ratelimit_increment(ctxt->rate, sectors);
819 			if (ctxt->stats) {
820 				atomic64_add(sectors, &ctxt->stats->sectors_seen);
821 				atomic64_add(sectors, &ctxt->stats->sectors_moved);
822 			}
823 		}
824 next:
825 		bp_pos = bpos_nosnap_successor(bp_pos);
826 	}
827 
828 	trace_evacuate_bucket(c, &bucket, dirty_sectors, bucket_size, fragmentation, ret);
829 err:
830 	bch2_dev_put(ca);
831 	bch2_bkey_buf_exit(&sk, c);
832 	return ret;
833 }
834 
835 typedef bool (*move_btree_pred)(struct bch_fs *, void *,
836 				struct btree *, struct bch_io_opts *,
837 				struct data_update_opts *);
838 
839 static int bch2_move_btree(struct bch_fs *c,
840 			   struct bbpos start,
841 			   struct bbpos end,
842 			   move_btree_pred pred, void *arg,
843 			   struct bch_move_stats *stats)
844 {
845 	bool kthread = (current->flags & PF_KTHREAD) != 0;
846 	struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
847 	struct moving_context ctxt;
848 	struct btree_trans *trans;
849 	struct btree_iter iter;
850 	struct btree *b;
851 	enum btree_id btree;
852 	struct data_update_opts data_opts;
853 	int ret = 0;
854 
855 	bch2_moving_ctxt_init(&ctxt, c, NULL, stats,
856 			      writepoint_ptr(&c->btree_write_point),
857 			      true);
858 	trans = ctxt.trans;
859 
860 	stats->data_type = BCH_DATA_btree;
861 
862 	for (btree = start.btree;
863 	     btree <= min_t(unsigned, end.btree, btree_id_nr_alive(c) - 1);
864 	     btree ++) {
865 		stats->pos = BBPOS(btree, POS_MIN);
866 
867 		if (!bch2_btree_id_root(c, btree)->b)
868 			continue;
869 
870 		bch2_trans_node_iter_init(trans, &iter, btree, POS_MIN, 0, 0,
871 					  BTREE_ITER_prefetch);
872 retry:
873 		ret = 0;
874 		while (bch2_trans_begin(trans),
875 		       (b = bch2_btree_iter_peek_node(&iter)) &&
876 		       !(ret = PTR_ERR_OR_ZERO(b))) {
877 			if (kthread && kthread_should_stop())
878 				break;
879 
880 			if ((cmp_int(btree, end.btree) ?:
881 			     bpos_cmp(b->key.k.p, end.pos)) > 0)
882 				break;
883 
884 			stats->pos = BBPOS(iter.btree_id, iter.pos);
885 
886 			if (!pred(c, arg, b, &io_opts, &data_opts))
887 				goto next;
888 
889 			ret = bch2_btree_node_rewrite(trans, &iter, b, 0) ?: ret;
890 			if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
891 				continue;
892 			if (ret)
893 				break;
894 next:
895 			bch2_btree_iter_next_node(&iter);
896 		}
897 		if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
898 			goto retry;
899 
900 		bch2_trans_iter_exit(trans, &iter);
901 
902 		if (kthread && kthread_should_stop())
903 			break;
904 	}
905 
906 	bch_err_fn(c, ret);
907 	bch2_moving_ctxt_exit(&ctxt);
908 	bch2_btree_interior_updates_flush(c);
909 
910 	return ret;
911 }
912 
913 static bool rereplicate_pred(struct bch_fs *c, void *arg,
914 			     struct bkey_s_c k,
915 			     struct bch_io_opts *io_opts,
916 			     struct data_update_opts *data_opts)
917 {
918 	unsigned nr_good = bch2_bkey_durability(c, k);
919 	unsigned replicas = bkey_is_btree_ptr(k.k)
920 		? c->opts.metadata_replicas
921 		: io_opts->data_replicas;
922 
923 	if (!nr_good || nr_good >= replicas)
924 		return false;
925 
926 	data_opts->target		= 0;
927 	data_opts->extra_replicas	= replicas - nr_good;
928 	data_opts->btree_insert_flags	= 0;
929 	return true;
930 }
931 
932 static bool migrate_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 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
938 	struct bch_ioctl_data *op = arg;
939 	unsigned i = 0;
940 
941 	data_opts->rewrite_ptrs		= 0;
942 	data_opts->target		= 0;
943 	data_opts->extra_replicas	= 0;
944 	data_opts->btree_insert_flags	= 0;
945 
946 	bkey_for_each_ptr(ptrs, ptr) {
947 		if (ptr->dev == op->migrate.dev)
948 			data_opts->rewrite_ptrs |= 1U << i;
949 		i++;
950 	}
951 
952 	return data_opts->rewrite_ptrs != 0;
953 }
954 
955 static bool rereplicate_btree_pred(struct bch_fs *c, void *arg,
956 				   struct btree *b,
957 				   struct bch_io_opts *io_opts,
958 				   struct data_update_opts *data_opts)
959 {
960 	return rereplicate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
961 }
962 
963 static bool migrate_btree_pred(struct bch_fs *c, void *arg,
964 			       struct btree *b,
965 			       struct bch_io_opts *io_opts,
966 			       struct data_update_opts *data_opts)
967 {
968 	return migrate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
969 }
970 
971 /*
972  * Ancient versions of bcachefs produced packed formats which could represent
973  * keys that the in memory format cannot represent; this checks for those
974  * formats so we can get rid of them.
975  */
976 static bool bformat_needs_redo(struct bkey_format *f)
977 {
978 	for (unsigned i = 0; i < f->nr_fields; i++)
979 		if (bch2_bkey_format_field_overflows(f, i))
980 			return true;
981 
982 	return false;
983 }
984 
985 static bool rewrite_old_nodes_pred(struct bch_fs *c, void *arg,
986 				   struct btree *b,
987 				   struct bch_io_opts *io_opts,
988 				   struct data_update_opts *data_opts)
989 {
990 	if (b->version_ondisk != c->sb.version ||
991 	    btree_node_need_rewrite(b) ||
992 	    bformat_needs_redo(&b->format)) {
993 		data_opts->target		= 0;
994 		data_opts->extra_replicas	= 0;
995 		data_opts->btree_insert_flags	= 0;
996 		return true;
997 	}
998 
999 	return false;
1000 }
1001 
1002 int bch2_scan_old_btree_nodes(struct bch_fs *c, struct bch_move_stats *stats)
1003 {
1004 	int ret;
1005 
1006 	ret = bch2_move_btree(c,
1007 			      BBPOS_MIN,
1008 			      BBPOS_MAX,
1009 			      rewrite_old_nodes_pred, c, stats);
1010 	if (!ret) {
1011 		mutex_lock(&c->sb_lock);
1012 		c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_extents_above_btree_updates_done);
1013 		c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_bformat_overflow_done);
1014 		c->disk_sb.sb->version_min = c->disk_sb.sb->version;
1015 		bch2_write_super(c);
1016 		mutex_unlock(&c->sb_lock);
1017 	}
1018 
1019 	bch_err_fn(c, ret);
1020 	return ret;
1021 }
1022 
1023 static bool drop_extra_replicas_pred(struct bch_fs *c, void *arg,
1024 			     struct bkey_s_c k,
1025 			     struct bch_io_opts *io_opts,
1026 			     struct data_update_opts *data_opts)
1027 {
1028 	unsigned durability = bch2_bkey_durability(c, k);
1029 	unsigned replicas = bkey_is_btree_ptr(k.k)
1030 		? c->opts.metadata_replicas
1031 		: io_opts->data_replicas;
1032 	const union bch_extent_entry *entry;
1033 	struct extent_ptr_decoded p;
1034 	unsigned i = 0;
1035 
1036 	rcu_read_lock();
1037 	bkey_for_each_ptr_decode(k.k, bch2_bkey_ptrs_c(k), p, entry) {
1038 		unsigned d = bch2_extent_ptr_durability(c, &p);
1039 
1040 		if (d && durability - d >= replicas) {
1041 			data_opts->kill_ptrs |= BIT(i);
1042 			durability -= d;
1043 		}
1044 
1045 		i++;
1046 	}
1047 	rcu_read_unlock();
1048 
1049 	return data_opts->kill_ptrs != 0;
1050 }
1051 
1052 static bool drop_extra_replicas_btree_pred(struct bch_fs *c, void *arg,
1053 				   struct btree *b,
1054 				   struct bch_io_opts *io_opts,
1055 				   struct data_update_opts *data_opts)
1056 {
1057 	return drop_extra_replicas_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
1058 }
1059 
1060 int bch2_data_job(struct bch_fs *c,
1061 		  struct bch_move_stats *stats,
1062 		  struct bch_ioctl_data op)
1063 {
1064 	struct bbpos start	= BBPOS(op.start_btree, op.start_pos);
1065 	struct bbpos end	= BBPOS(op.end_btree, op.end_pos);
1066 	int ret = 0;
1067 
1068 	if (op.op >= BCH_DATA_OP_NR)
1069 		return -EINVAL;
1070 
1071 	bch2_move_stats_init(stats, bch2_data_ops_strs[op.op]);
1072 
1073 	switch (op.op) {
1074 	case BCH_DATA_OP_rereplicate:
1075 		stats->data_type = BCH_DATA_journal;
1076 		ret = bch2_journal_flush_device_pins(&c->journal, -1);
1077 		ret = bch2_move_btree(c, start, end,
1078 				      rereplicate_btree_pred, c, stats) ?: ret;
1079 		ret = bch2_move_data(c, start, end,
1080 				     NULL,
1081 				     stats,
1082 				     writepoint_hashed((unsigned long) current),
1083 				     true,
1084 				     rereplicate_pred, c) ?: ret;
1085 		ret = bch2_replicas_gc2(c) ?: ret;
1086 		break;
1087 	case BCH_DATA_OP_migrate:
1088 		if (op.migrate.dev >= c->sb.nr_devices)
1089 			return -EINVAL;
1090 
1091 		stats->data_type = BCH_DATA_journal;
1092 		ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
1093 		ret = bch2_move_btree(c, start, end,
1094 				      migrate_btree_pred, &op, stats) ?: ret;
1095 		ret = bch2_move_data(c, start, end,
1096 				     NULL,
1097 				     stats,
1098 				     writepoint_hashed((unsigned long) current),
1099 				     true,
1100 				     migrate_pred, &op) ?: ret;
1101 		ret = bch2_replicas_gc2(c) ?: ret;
1102 		break;
1103 	case BCH_DATA_OP_rewrite_old_nodes:
1104 		ret = bch2_scan_old_btree_nodes(c, stats);
1105 		break;
1106 	case BCH_DATA_OP_drop_extra_replicas:
1107 		ret = bch2_move_btree(c, start, end,
1108 				drop_extra_replicas_btree_pred, c, stats) ?: ret;
1109 		ret = bch2_move_data(c, start, end, NULL, stats,
1110 				writepoint_hashed((unsigned long) current),
1111 				true,
1112 				drop_extra_replicas_pred, c) ?: ret;
1113 		ret = bch2_replicas_gc2(c) ?: ret;
1114 		break;
1115 	default:
1116 		ret = -EINVAL;
1117 	}
1118 
1119 	bch2_move_stats_exit(stats, c);
1120 	return ret;
1121 }
1122 
1123 void bch2_move_stats_to_text(struct printbuf *out, struct bch_move_stats *stats)
1124 {
1125 	prt_printf(out, "%s: data type==", stats->name);
1126 	bch2_prt_data_type(out, stats->data_type);
1127 	prt_str(out, " pos=");
1128 	bch2_bbpos_to_text(out, stats->pos);
1129 	prt_newline(out);
1130 	printbuf_indent_add(out, 2);
1131 
1132 	prt_printf(out, "keys moved:  %llu\n",	atomic64_read(&stats->keys_moved));
1133 	prt_printf(out, "keys raced:  %llu\n",	atomic64_read(&stats->keys_raced));
1134 	prt_printf(out, "bytes seen:  ");
1135 	prt_human_readable_u64(out, atomic64_read(&stats->sectors_seen) << 9);
1136 	prt_newline(out);
1137 
1138 	prt_printf(out, "bytes moved: ");
1139 	prt_human_readable_u64(out, atomic64_read(&stats->sectors_moved) << 9);
1140 	prt_newline(out);
1141 
1142 	prt_printf(out, "bytes raced: ");
1143 	prt_human_readable_u64(out, atomic64_read(&stats->sectors_raced) << 9);
1144 	prt_newline(out);
1145 
1146 	printbuf_indent_sub(out, 2);
1147 }
1148 
1149 static void bch2_moving_ctxt_to_text(struct printbuf *out, struct bch_fs *c, struct moving_context *ctxt)
1150 {
1151 	struct moving_io *io;
1152 
1153 	bch2_move_stats_to_text(out, ctxt->stats);
1154 	printbuf_indent_add(out, 2);
1155 
1156 	prt_printf(out, "reads: ios %u/%u sectors %u/%u\n",
1157 		   atomic_read(&ctxt->read_ios),
1158 		   c->opts.move_ios_in_flight,
1159 		   atomic_read(&ctxt->read_sectors),
1160 		   c->opts.move_bytes_in_flight >> 9);
1161 
1162 	prt_printf(out, "writes: ios %u/%u sectors %u/%u\n",
1163 		   atomic_read(&ctxt->write_ios),
1164 		   c->opts.move_ios_in_flight,
1165 		   atomic_read(&ctxt->write_sectors),
1166 		   c->opts.move_bytes_in_flight >> 9);
1167 
1168 	printbuf_indent_add(out, 2);
1169 
1170 	mutex_lock(&ctxt->lock);
1171 	list_for_each_entry(io, &ctxt->ios, io_list)
1172 		bch2_write_op_to_text(out, &io->write.op);
1173 	mutex_unlock(&ctxt->lock);
1174 
1175 	printbuf_indent_sub(out, 4);
1176 }
1177 
1178 void bch2_fs_moving_ctxts_to_text(struct printbuf *out, struct bch_fs *c)
1179 {
1180 	struct moving_context *ctxt;
1181 
1182 	mutex_lock(&c->moving_context_lock);
1183 	list_for_each_entry(ctxt, &c->moving_context_list, list)
1184 		bch2_moving_ctxt_to_text(out, c, ctxt);
1185 	mutex_unlock(&c->moving_context_lock);
1186 }
1187 
1188 void bch2_fs_move_init(struct bch_fs *c)
1189 {
1190 	INIT_LIST_HEAD(&c->moving_context_list);
1191 	mutex_init(&c->moving_context_lock);
1192 }
1193