1 // SPDX-License-Identifier: GPL-2.0
4 #include "alloc_foreground.h"
5 #include "bkey_on_stack.h"
7 #include "btree_update.h"
8 #include "btree_update_interior.h"
10 #include "disk_groups.h"
13 #include "journal_reclaim.h"
19 #include <linux/ioprio.h>
20 #include <linux/kthread.h>
22 #include <trace/events/bcachefs.h>
24 #define SECTORS_IN_FLIGHT_PER_DEVICE 2048
27 struct list_head list;
31 unsigned read_sectors;
32 unsigned write_sectors;
34 struct bch_read_bio rbio;
36 struct migrate_write write;
37 /* Must be last since it is variable size */
38 struct bio_vec bi_inline_vecs[0];
41 struct moving_context {
42 /* Closure for waiting on all reads and writes to complete */
45 struct bch_move_stats *stats;
47 struct list_head reads;
49 /* in flight sectors: */
50 atomic_t read_sectors;
51 atomic_t write_sectors;
53 wait_queue_head_t wait;
56 static int bch2_migrate_index_update(struct bch_write_op *op)
58 struct bch_fs *c = op->c;
59 struct btree_trans trans;
60 struct btree_iter *iter;
61 struct migrate_write *m =
62 container_of(op, struct migrate_write, op);
63 struct keylist *keys = &op->insert_keys;
66 bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
68 iter = bch2_trans_get_iter(&trans, m->btree_id,
69 bkey_start_pos(&bch2_keylist_front(keys)->k),
70 BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
74 struct bkey_i *insert;
75 struct bkey_i_extent *new;
76 BKEY_PADDED(k) _new, _insert;
77 const union bch_extent_entry *entry;
78 struct extent_ptr_decoded p;
79 bool did_work = false;
82 bch2_trans_reset(&trans, 0);
84 k = bch2_btree_iter_peek_slot(iter);
92 new = bkey_i_to_extent(bch2_keylist_front(keys));
94 if (bversion_cmp(k.k->version, new->k.version) ||
95 !bch2_bkey_matches_ptr(c, k, m->ptr, m->offset))
98 if (m->data_cmd == DATA_REWRITE &&
99 !bch2_bkey_has_device(k, m->data_opts.rewrite_dev))
102 bkey_reassemble(&_insert.k, k);
105 bkey_copy(&_new.k, bch2_keylist_front(keys));
106 new = bkey_i_to_extent(&_new.k);
107 bch2_cut_front(iter->pos, &new->k_i);
109 bch2_cut_front(iter->pos, insert);
110 bch2_cut_back(new->k.p, insert);
111 bch2_cut_back(insert->k.p, &new->k_i);
113 if (m->data_cmd == DATA_REWRITE)
114 bch2_bkey_drop_device(bkey_i_to_s(insert),
115 m->data_opts.rewrite_dev);
117 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry) {
118 if (bch2_bkey_has_device(bkey_i_to_s_c(insert), p.ptr.dev)) {
120 * raced with another move op? extent already
121 * has a pointer to the device we just wrote
127 bch2_extent_ptr_decoded_append(insert, &p);
134 bch2_bkey_narrow_crcs(insert,
135 (struct bch_extent_crc_unpacked) { 0 });
136 bch2_extent_normalize(c, bkey_i_to_s(insert));
137 bch2_bkey_mark_replicas_cached(c, bkey_i_to_s(insert),
138 op->opts.background_target,
139 op->opts.data_replicas);
142 * If we're not fully overwriting @k, and it's compressed, we
143 * need a reservation for all the pointers in @insert
145 nr = bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(insert)) -
148 if (insert->k.size < k.k->size &&
149 bch2_bkey_sectors_compressed(k) &&
151 ret = bch2_disk_reservation_add(c, &op->res,
152 keylist_sectors(keys) * nr, 0);
156 m->nr_ptrs_reserved += nr;
160 bch2_trans_update(&trans, iter, insert, 0);
162 ret = bch2_trans_commit(&trans, &op->res,
165 BTREE_INSERT_USE_RESERVE|
166 m->data_opts.btree_insert_flags);
168 atomic_long_inc(&c->extent_migrate_done);
174 while (bkey_cmp(iter->pos, bch2_keylist_front(keys)->k.p) >= 0) {
175 bch2_keylist_pop_front(keys);
176 if (bch2_keylist_empty(keys))
182 atomic64_add(k.k->p.offset - iter->pos.offset,
183 &m->ctxt->stats->sectors_raced);
184 atomic_long_inc(&c->extent_migrate_raced);
185 trace_move_race(&new->k);
186 bch2_btree_iter_next_slot(iter);
190 bch2_trans_exit(&trans);
191 BUG_ON(ret == -EINTR);
195 void bch2_migrate_read_done(struct migrate_write *m, struct bch_read_bio *rbio)
197 /* write bio must own pages: */
198 BUG_ON(!m->op.wbio.bio.bi_vcnt);
200 m->ptr = rbio->pick.ptr;
201 m->offset = rbio->pos.offset - rbio->pick.crc.offset;
202 m->op.devs_have = rbio->devs_have;
203 m->op.pos = rbio->pos;
204 m->op.version = rbio->version;
205 m->op.crc = rbio->pick.crc;
206 m->op.wbio.bio.bi_iter.bi_size = m->op.crc.compressed_size << 9;
208 if (bch2_csum_type_is_encryption(m->op.crc.csum_type)) {
209 m->op.nonce = m->op.crc.nonce + m->op.crc.offset;
210 m->op.csum_type = m->op.crc.csum_type;
213 if (m->data_cmd == DATA_REWRITE)
214 bch2_dev_list_drop_dev(&m->op.devs_have, m->data_opts.rewrite_dev);
217 int bch2_migrate_write_init(struct bch_fs *c, struct migrate_write *m,
218 struct write_point_specifier wp,
219 struct bch_io_opts io_opts,
220 enum data_cmd data_cmd,
221 struct data_opts data_opts,
222 enum btree_id btree_id,
225 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
226 const union bch_extent_entry *entry;
227 struct extent_ptr_decoded p;
230 m->btree_id = btree_id;
231 m->data_cmd = data_cmd;
232 m->data_opts = data_opts;
233 m->nr_ptrs_reserved = 0;
235 bch2_write_op_init(&m->op, c, io_opts);
237 if (!bch2_bkey_is_incompressible(k))
238 m->op.compression_type =
239 bch2_compression_opt_to_type[io_opts.background_compression ?:
240 io_opts.compression];
242 m->op.incompressible = true;
244 m->op.target = data_opts.target,
245 m->op.write_point = wp;
247 if (m->data_opts.btree_insert_flags & BTREE_INSERT_USE_RESERVE)
248 m->op.alloc_reserve = RESERVE_MOVINGGC;
250 m->op.flags |= BCH_WRITE_ONLY_SPECIFIED_DEVS|
251 BCH_WRITE_PAGES_STABLE|
252 BCH_WRITE_PAGES_OWNED|
253 BCH_WRITE_DATA_ENCODED|
254 BCH_WRITE_FROM_INTERNAL;
256 m->op.nr_replicas = 1;
257 m->op.nr_replicas_required = 1;
258 m->op.index_update_fn = bch2_migrate_index_update;
261 case DATA_ADD_REPLICAS: {
263 * DATA_ADD_REPLICAS is used for moving data to a different
264 * device in the background, and due to compression the new copy
265 * might take up more space than the old copy:
268 int nr = (int) io_opts.data_replicas -
269 bch2_bkey_nr_ptrs_allocated(k);
271 int nr = (int) io_opts.data_replicas;
274 m->op.nr_replicas = m->nr_ptrs_reserved = nr;
276 ret = bch2_disk_reservation_get(c, &m->op.res,
277 k.k->size, m->op.nr_replicas, 0);
284 unsigned compressed_sectors = 0;
286 bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
288 crc_is_compressed(p.crc) &&
289 bch2_dev_in_target(c, p.ptr.dev, data_opts.target))
290 compressed_sectors += p.crc.compressed_size;
292 if (compressed_sectors) {
293 ret = bch2_disk_reservation_add(c, &m->op.res,
295 BCH_DISK_RESERVATION_NOFAIL);
302 m->op.flags |= BCH_WRITE_ALLOC_NOWAIT;
303 m->op.flags |= BCH_WRITE_CACHED;
312 static void move_free(struct closure *cl)
314 struct moving_io *io = container_of(cl, struct moving_io, cl);
315 struct moving_context *ctxt = io->write.ctxt;
316 struct bvec_iter_all iter;
319 bch2_disk_reservation_put(io->write.op.c, &io->write.op.res);
321 bio_for_each_segment_all(bv, &io->write.op.wbio.bio, iter)
323 __free_page(bv->bv_page);
325 wake_up(&ctxt->wait);
330 static void move_write_done(struct closure *cl)
332 struct moving_io *io = container_of(cl, struct moving_io, cl);
334 atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
335 closure_return_with_destructor(cl, move_free);
338 static void move_write(struct closure *cl)
340 struct moving_io *io = container_of(cl, struct moving_io, cl);
342 if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
343 closure_return_with_destructor(cl, move_free);
347 bch2_migrate_read_done(&io->write, &io->rbio);
349 atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
350 closure_call(&io->write.op.cl, bch2_write, NULL, cl);
351 continue_at(cl, move_write_done, NULL);
354 static inline struct moving_io *next_pending_write(struct moving_context *ctxt)
356 struct moving_io *io =
357 list_first_entry_or_null(&ctxt->reads, struct moving_io, list);
359 return io && io->read_completed ? io : NULL;
362 static void move_read_endio(struct bio *bio)
364 struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
365 struct moving_context *ctxt = io->write.ctxt;
367 atomic_sub(io->read_sectors, &ctxt->read_sectors);
368 io->read_completed = true;
370 if (next_pending_write(ctxt))
371 wake_up(&ctxt->wait);
373 closure_put(&ctxt->cl);
376 static void do_pending_writes(struct moving_context *ctxt)
378 struct moving_io *io;
380 while ((io = next_pending_write(ctxt))) {
382 closure_call(&io->cl, move_write, NULL, &ctxt->cl);
386 #define move_ctxt_wait_event(_ctxt, _cond) \
388 do_pending_writes(_ctxt); \
392 __wait_event((_ctxt)->wait, \
393 next_pending_write(_ctxt) || (_cond)); \
396 static void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt)
398 unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
400 move_ctxt_wait_event(ctxt,
401 !atomic_read(&ctxt->write_sectors) ||
402 atomic_read(&ctxt->write_sectors) != sectors_pending);
405 static int bch2_move_extent(struct bch_fs *c,
406 struct moving_context *ctxt,
407 struct write_point_specifier wp,
408 struct bch_io_opts io_opts,
409 enum btree_id btree_id,
411 enum data_cmd data_cmd,
412 struct data_opts data_opts)
414 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
415 struct moving_io *io;
416 const union bch_extent_entry *entry;
417 struct extent_ptr_decoded p;
418 unsigned sectors = k.k->size, pages;
421 move_ctxt_wait_event(ctxt,
422 atomic_read(&ctxt->write_sectors) <
423 SECTORS_IN_FLIGHT_PER_DEVICE);
425 move_ctxt_wait_event(ctxt,
426 atomic_read(&ctxt->read_sectors) <
427 SECTORS_IN_FLIGHT_PER_DEVICE);
429 /* write path might have to decompress data: */
430 bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
431 sectors = max_t(unsigned, sectors, p.crc.uncompressed_size);
433 pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
434 io = kzalloc(sizeof(struct moving_io) +
435 sizeof(struct bio_vec) * pages, GFP_KERNEL);
439 io->write.ctxt = ctxt;
440 io->read_sectors = k.k->size;
441 io->write_sectors = k.k->size;
443 bio_init(&io->write.op.wbio.bio, io->bi_inline_vecs, pages);
444 bio_set_prio(&io->write.op.wbio.bio,
445 IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
447 if (bch2_bio_alloc_pages(&io->write.op.wbio.bio, sectors << 9,
452 io->rbio.opts = io_opts;
453 bio_init(&io->rbio.bio, io->bi_inline_vecs, pages);
454 io->rbio.bio.bi_vcnt = pages;
455 bio_set_prio(&io->rbio.bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
456 io->rbio.bio.bi_iter.bi_size = sectors << 9;
458 bio_set_op_attrs(&io->rbio.bio, REQ_OP_READ, 0);
459 io->rbio.bio.bi_iter.bi_sector = bkey_start_offset(k.k);
460 io->rbio.bio.bi_end_io = move_read_endio;
462 ret = bch2_migrate_write_init(c, &io->write, wp, io_opts,
463 data_cmd, data_opts, btree_id, k);
467 atomic64_inc(&ctxt->stats->keys_moved);
468 atomic64_add(k.k->size, &ctxt->stats->sectors_moved);
470 trace_move_extent(k.k);
472 atomic_add(io->read_sectors, &ctxt->read_sectors);
473 list_add_tail(&io->list, &ctxt->reads);
476 * dropped by move_read_endio() - guards against use after free of
477 * ctxt when doing wakeup
479 closure_get(&ctxt->cl);
480 bch2_read_extent(c, &io->rbio, k, 0,
482 BCH_READ_LAST_FRAGMENT);
485 bio_free_pages(&io->write.op.wbio.bio);
489 trace_move_alloc_fail(k.k);
493 static int __bch2_move_data(struct bch_fs *c,
494 struct moving_context *ctxt,
495 struct bch_ratelimit *rate,
496 struct write_point_specifier wp,
499 move_pred_fn pred, void *arg,
500 struct bch_move_stats *stats,
501 enum btree_id btree_id)
503 bool kthread = (current->flags & PF_KTHREAD) != 0;
504 struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
505 struct bkey_on_stack sk;
506 struct btree_trans trans;
507 struct btree_iter *iter;
509 struct data_opts data_opts;
510 enum data_cmd data_cmd;
511 u64 delay, cur_inum = U64_MAX;
514 bkey_on_stack_init(&sk);
515 bch2_trans_init(&trans, c, 0, 0);
517 stats->data_type = BCH_DATA_USER;
518 stats->btree_id = btree_id;
519 stats->pos = POS_MIN;
521 iter = bch2_trans_get_iter(&trans, btree_id, start,
522 BTREE_ITER_PREFETCH);
525 bch2_ratelimit_reset(rate);
529 delay = rate ? bch2_ratelimit_delay(rate) : 0;
532 bch2_trans_unlock(&trans);
533 set_current_state(TASK_INTERRUPTIBLE);
536 if (kthread && (ret = kthread_should_stop())) {
537 __set_current_state(TASK_RUNNING);
542 schedule_timeout(delay);
544 if (unlikely(freezing(current))) {
545 bch2_trans_unlock(&trans);
546 move_ctxt_wait_event(ctxt, list_empty(&ctxt->reads));
551 k = bch2_btree_iter_peek(iter);
553 stats->pos = iter->pos;
560 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
563 if (!bkey_extent_is_direct_data(k.k))
566 if (btree_id == BTREE_ID_EXTENTS &&
567 cur_inum != k.k->p.inode) {
568 struct bch_inode_unpacked inode;
570 /* don't hold btree locks while looking up inode: */
571 bch2_trans_unlock(&trans);
573 io_opts = bch2_opts_to_inode_opts(c->opts);
574 if (!bch2_inode_find_by_inum(c, k.k->p.inode, &inode))
575 bch2_io_opts_apply(&io_opts, bch2_inode_opts_get(&inode));
576 cur_inum = k.k->p.inode;
580 switch ((data_cmd = pred(c, arg, k, &io_opts, &data_opts))) {
585 case DATA_ADD_REPLICAS:
593 /* unlock before doing IO: */
594 bkey_on_stack_reassemble(&sk, c, k);
595 k = bkey_i_to_s_c(sk.k);
596 bch2_trans_unlock(&trans);
598 ret2 = bch2_move_extent(c, ctxt, wp, io_opts, btree_id, k,
599 data_cmd, data_opts);
601 if (ret2 == -ENOMEM) {
602 /* memory allocation failure, wait for some IO to finish */
603 bch2_move_ctxt_wait_for_io(ctxt);
607 /* XXX signal failure */
612 bch2_ratelimit_increment(rate, k.k->size);
614 atomic64_add(k.k->size * bch2_bkey_nr_ptrs_allocated(k),
615 &stats->sectors_seen);
617 bch2_btree_iter_next(iter);
618 bch2_trans_cond_resched(&trans);
621 ret = bch2_trans_exit(&trans) ?: ret;
622 bkey_on_stack_exit(&sk, c);
627 int bch2_move_data(struct bch_fs *c,
628 struct bch_ratelimit *rate,
629 struct write_point_specifier wp,
632 move_pred_fn pred, void *arg,
633 struct bch_move_stats *stats)
635 struct moving_context ctxt = { .stats = stats };
638 closure_init_stack(&ctxt.cl);
639 INIT_LIST_HEAD(&ctxt.reads);
640 init_waitqueue_head(&ctxt.wait);
642 stats->data_type = BCH_DATA_USER;
644 ret = __bch2_move_data(c, &ctxt, rate, wp, start, end,
645 pred, arg, stats, BTREE_ID_EXTENTS) ?:
646 __bch2_move_data(c, &ctxt, rate, wp, start, end,
647 pred, arg, stats, BTREE_ID_REFLINK);
649 move_ctxt_wait_event(&ctxt, list_empty(&ctxt.reads));
650 closure_sync(&ctxt.cl);
652 EBUG_ON(atomic_read(&ctxt.write_sectors));
655 atomic64_read(&stats->sectors_moved),
656 atomic64_read(&stats->keys_moved));
661 static int bch2_move_btree(struct bch_fs *c,
664 struct bch_move_stats *stats)
666 struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
667 struct btree_trans trans;
668 struct btree_iter *iter;
671 struct data_opts data_opts;
675 bch2_trans_init(&trans, c, 0, 0);
677 stats->data_type = BCH_DATA_BTREE;
679 for (id = 0; id < BTREE_ID_NR; id++) {
680 stats->btree_id = id;
682 for_each_btree_node(&trans, iter, id, POS_MIN,
683 BTREE_ITER_PREFETCH, b) {
684 stats->pos = iter->pos;
686 switch ((cmd = pred(c, arg,
687 bkey_i_to_s_c(&b->key),
688 &io_opts, &data_opts))) {
693 case DATA_ADD_REPLICAS:
700 ret = bch2_btree_node_rewrite(c, iter,
701 b->data->keys.seq, 0) ?: ret;
703 bch2_trans_cond_resched(&trans);
706 ret = bch2_trans_iter_free(&trans, iter) ?: ret;
709 bch2_trans_exit(&trans);
715 static enum data_cmd scrub_pred(struct bch_fs *c, void *arg,
717 struct bch_io_opts *io_opts,
718 struct data_opts *data_opts)
724 static enum data_cmd rereplicate_pred(struct bch_fs *c, void *arg,
726 struct bch_io_opts *io_opts,
727 struct data_opts *data_opts)
729 unsigned nr_good = bch2_bkey_durability(c, k);
730 unsigned replicas = 0;
733 case KEY_TYPE_btree_ptr:
734 replicas = c->opts.metadata_replicas;
736 case KEY_TYPE_extent:
737 replicas = io_opts->data_replicas;
741 if (!nr_good || nr_good >= replicas)
744 data_opts->target = 0;
745 data_opts->btree_insert_flags = 0;
746 return DATA_ADD_REPLICAS;
749 static enum data_cmd migrate_pred(struct bch_fs *c, void *arg,
751 struct bch_io_opts *io_opts,
752 struct data_opts *data_opts)
754 struct bch_ioctl_data *op = arg;
756 if (!bch2_bkey_has_device(k, op->migrate.dev))
759 data_opts->target = 0;
760 data_opts->btree_insert_flags = 0;
761 data_opts->rewrite_dev = op->migrate.dev;
765 int bch2_data_job(struct bch_fs *c,
766 struct bch_move_stats *stats,
767 struct bch_ioctl_data op)
772 case BCH_DATA_OP_REREPLICATE:
773 stats->data_type = BCH_DATA_JOURNAL;
774 ret = bch2_journal_flush_device_pins(&c->journal, -1);
776 ret = bch2_move_btree(c, rereplicate_pred, c, stats) ?: ret;
778 closure_wait_event(&c->btree_interior_update_wait,
779 !bch2_btree_interior_updates_nr_pending(c));
781 ret = bch2_replicas_gc2(c) ?: ret;
783 ret = bch2_move_data(c, NULL,
784 writepoint_hashed((unsigned long) current),
787 rereplicate_pred, c, stats) ?: ret;
788 ret = bch2_replicas_gc2(c) ?: ret;
790 case BCH_DATA_OP_MIGRATE:
791 if (op.migrate.dev >= c->sb.nr_devices)
794 stats->data_type = BCH_DATA_JOURNAL;
795 ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
797 ret = bch2_move_btree(c, migrate_pred, &op, stats) ?: ret;
798 ret = bch2_replicas_gc2(c) ?: ret;
800 ret = bch2_move_data(c, NULL,
801 writepoint_hashed((unsigned long) current),
804 migrate_pred, &op, stats) ?: ret;
805 ret = bch2_replicas_gc2(c) ?: ret;