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 BUG_ON(k.k->p.offset <= iter->pos.offset);
183 atomic64_inc(&m->ctxt->stats->keys_raced);
184 atomic64_add(k.k->p.offset - iter->pos.offset,
185 &m->ctxt->stats->sectors_raced);
187 atomic_long_inc(&c->extent_migrate_raced);
188 trace_move_race(&new->k);
189 bch2_btree_iter_next_slot(iter);
193 bch2_trans_exit(&trans);
194 BUG_ON(ret == -EINTR);
198 void bch2_migrate_read_done(struct migrate_write *m, struct bch_read_bio *rbio)
200 /* write bio must own pages: */
201 BUG_ON(!m->op.wbio.bio.bi_vcnt);
203 m->ptr = rbio->pick.ptr;
204 m->offset = rbio->pos.offset - rbio->pick.crc.offset;
205 m->op.devs_have = rbio->devs_have;
206 m->op.pos = rbio->pos;
207 m->op.version = rbio->version;
208 m->op.crc = rbio->pick.crc;
209 m->op.wbio.bio.bi_iter.bi_size = m->op.crc.compressed_size << 9;
211 if (bch2_csum_type_is_encryption(m->op.crc.csum_type)) {
212 m->op.nonce = m->op.crc.nonce + m->op.crc.offset;
213 m->op.csum_type = m->op.crc.csum_type;
216 if (m->data_cmd == DATA_REWRITE)
217 bch2_dev_list_drop_dev(&m->op.devs_have, m->data_opts.rewrite_dev);
220 int bch2_migrate_write_init(struct bch_fs *c, struct migrate_write *m,
221 struct write_point_specifier wp,
222 struct bch_io_opts io_opts,
223 enum data_cmd data_cmd,
224 struct data_opts data_opts,
225 enum btree_id btree_id,
228 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
229 const union bch_extent_entry *entry;
230 struct extent_ptr_decoded p;
233 m->btree_id = btree_id;
234 m->data_cmd = data_cmd;
235 m->data_opts = data_opts;
236 m->nr_ptrs_reserved = 0;
238 bch2_write_op_init(&m->op, c, io_opts);
240 if (!bch2_bkey_is_incompressible(k))
241 m->op.compression_type =
242 bch2_compression_opt_to_type[io_opts.background_compression ?:
243 io_opts.compression];
245 m->op.incompressible = true;
247 m->op.target = data_opts.target,
248 m->op.write_point = wp;
250 if (m->data_opts.btree_insert_flags & BTREE_INSERT_USE_RESERVE)
251 m->op.alloc_reserve = RESERVE_MOVINGGC;
253 m->op.flags |= BCH_WRITE_ONLY_SPECIFIED_DEVS|
254 BCH_WRITE_PAGES_STABLE|
255 BCH_WRITE_PAGES_OWNED|
256 BCH_WRITE_DATA_ENCODED|
257 BCH_WRITE_FROM_INTERNAL;
259 m->op.nr_replicas = 1;
260 m->op.nr_replicas_required = 1;
261 m->op.index_update_fn = bch2_migrate_index_update;
264 case DATA_ADD_REPLICAS: {
266 * DATA_ADD_REPLICAS is used for moving data to a different
267 * device in the background, and due to compression the new copy
268 * might take up more space than the old copy:
271 int nr = (int) io_opts.data_replicas -
272 bch2_bkey_nr_ptrs_allocated(k);
274 int nr = (int) io_opts.data_replicas;
277 m->op.nr_replicas = m->nr_ptrs_reserved = nr;
279 ret = bch2_disk_reservation_get(c, &m->op.res,
280 k.k->size, m->op.nr_replicas, 0);
287 unsigned compressed_sectors = 0;
289 bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
291 crc_is_compressed(p.crc) &&
292 bch2_dev_in_target(c, p.ptr.dev, data_opts.target))
293 compressed_sectors += p.crc.compressed_size;
295 if (compressed_sectors) {
296 ret = bch2_disk_reservation_add(c, &m->op.res,
298 BCH_DISK_RESERVATION_NOFAIL);
305 m->op.flags |= BCH_WRITE_ALLOC_NOWAIT;
306 m->op.flags |= BCH_WRITE_CACHED;
315 static void move_free(struct closure *cl)
317 struct moving_io *io = container_of(cl, struct moving_io, cl);
318 struct moving_context *ctxt = io->write.ctxt;
319 struct bvec_iter_all iter;
322 bch2_disk_reservation_put(io->write.op.c, &io->write.op.res);
324 bio_for_each_segment_all(bv, &io->write.op.wbio.bio, iter)
326 __free_page(bv->bv_page);
328 wake_up(&ctxt->wait);
333 static void move_write_done(struct closure *cl)
335 struct moving_io *io = container_of(cl, struct moving_io, cl);
337 atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
338 closure_return_with_destructor(cl, move_free);
341 static void move_write(struct closure *cl)
343 struct moving_io *io = container_of(cl, struct moving_io, cl);
345 if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
346 closure_return_with_destructor(cl, move_free);
350 bch2_migrate_read_done(&io->write, &io->rbio);
352 atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
353 closure_call(&io->write.op.cl, bch2_write, NULL, cl);
354 continue_at(cl, move_write_done, NULL);
357 static inline struct moving_io *next_pending_write(struct moving_context *ctxt)
359 struct moving_io *io =
360 list_first_entry_or_null(&ctxt->reads, struct moving_io, list);
362 return io && io->read_completed ? io : NULL;
365 static void move_read_endio(struct bio *bio)
367 struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
368 struct moving_context *ctxt = io->write.ctxt;
370 atomic_sub(io->read_sectors, &ctxt->read_sectors);
371 io->read_completed = true;
373 if (next_pending_write(ctxt))
374 wake_up(&ctxt->wait);
376 closure_put(&ctxt->cl);
379 static void do_pending_writes(struct moving_context *ctxt)
381 struct moving_io *io;
383 while ((io = next_pending_write(ctxt))) {
385 closure_call(&io->cl, move_write, NULL, &ctxt->cl);
389 #define move_ctxt_wait_event(_ctxt, _cond) \
391 do_pending_writes(_ctxt); \
395 __wait_event((_ctxt)->wait, \
396 next_pending_write(_ctxt) || (_cond)); \
399 static void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt)
401 unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
403 move_ctxt_wait_event(ctxt,
404 !atomic_read(&ctxt->write_sectors) ||
405 atomic_read(&ctxt->write_sectors) != sectors_pending);
408 static int bch2_move_extent(struct bch_fs *c,
409 struct moving_context *ctxt,
410 struct write_point_specifier wp,
411 struct bch_io_opts io_opts,
412 enum btree_id btree_id,
414 enum data_cmd data_cmd,
415 struct data_opts data_opts)
417 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
418 struct moving_io *io;
419 const union bch_extent_entry *entry;
420 struct extent_ptr_decoded p;
421 unsigned sectors = k.k->size, pages;
424 move_ctxt_wait_event(ctxt,
425 atomic_read(&ctxt->write_sectors) <
426 SECTORS_IN_FLIGHT_PER_DEVICE);
428 move_ctxt_wait_event(ctxt,
429 atomic_read(&ctxt->read_sectors) <
430 SECTORS_IN_FLIGHT_PER_DEVICE);
432 /* write path might have to decompress data: */
433 bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
434 sectors = max_t(unsigned, sectors, p.crc.uncompressed_size);
436 pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
437 io = kzalloc(sizeof(struct moving_io) +
438 sizeof(struct bio_vec) * pages, GFP_KERNEL);
442 io->write.ctxt = ctxt;
443 io->read_sectors = k.k->size;
444 io->write_sectors = k.k->size;
446 bio_init(&io->write.op.wbio.bio, io->bi_inline_vecs, pages);
447 bio_set_prio(&io->write.op.wbio.bio,
448 IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
450 if (bch2_bio_alloc_pages(&io->write.op.wbio.bio, sectors << 9,
455 io->rbio.opts = io_opts;
456 bio_init(&io->rbio.bio, io->bi_inline_vecs, pages);
457 io->rbio.bio.bi_vcnt = pages;
458 bio_set_prio(&io->rbio.bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
459 io->rbio.bio.bi_iter.bi_size = sectors << 9;
461 bio_set_op_attrs(&io->rbio.bio, REQ_OP_READ, 0);
462 io->rbio.bio.bi_iter.bi_sector = bkey_start_offset(k.k);
463 io->rbio.bio.bi_end_io = move_read_endio;
465 ret = bch2_migrate_write_init(c, &io->write, wp, io_opts,
466 data_cmd, data_opts, btree_id, k);
470 atomic64_inc(&ctxt->stats->keys_moved);
471 atomic64_add(k.k->size, &ctxt->stats->sectors_moved);
473 trace_move_extent(k.k);
475 atomic_add(io->read_sectors, &ctxt->read_sectors);
476 list_add_tail(&io->list, &ctxt->reads);
479 * dropped by move_read_endio() - guards against use after free of
480 * ctxt when doing wakeup
482 closure_get(&ctxt->cl);
483 bch2_read_extent(c, &io->rbio, k, 0,
485 BCH_READ_LAST_FRAGMENT);
488 bio_free_pages(&io->write.op.wbio.bio);
492 trace_move_alloc_fail(k.k);
496 static int __bch2_move_data(struct bch_fs *c,
497 struct moving_context *ctxt,
498 struct bch_ratelimit *rate,
499 struct write_point_specifier wp,
502 move_pred_fn pred, void *arg,
503 struct bch_move_stats *stats,
504 enum btree_id btree_id)
506 bool kthread = (current->flags & PF_KTHREAD) != 0;
507 struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
508 struct bkey_on_stack sk;
509 struct btree_trans trans;
510 struct btree_iter *iter;
512 struct data_opts data_opts;
513 enum data_cmd data_cmd;
514 u64 delay, cur_inum = U64_MAX;
517 bkey_on_stack_init(&sk);
518 bch2_trans_init(&trans, c, 0, 0);
520 stats->data_type = BCH_DATA_USER;
521 stats->btree_id = btree_id;
522 stats->pos = POS_MIN;
524 iter = bch2_trans_get_iter(&trans, btree_id, start,
525 BTREE_ITER_PREFETCH);
528 bch2_ratelimit_reset(rate);
532 delay = rate ? bch2_ratelimit_delay(rate) : 0;
535 bch2_trans_unlock(&trans);
536 set_current_state(TASK_INTERRUPTIBLE);
539 if (kthread && (ret = kthread_should_stop())) {
540 __set_current_state(TASK_RUNNING);
545 schedule_timeout(delay);
547 if (unlikely(freezing(current))) {
548 bch2_trans_unlock(&trans);
549 move_ctxt_wait_event(ctxt, list_empty(&ctxt->reads));
554 k = bch2_btree_iter_peek(iter);
556 stats->pos = iter->pos;
563 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
566 if (!bkey_extent_is_direct_data(k.k))
569 if (btree_id == BTREE_ID_EXTENTS &&
570 cur_inum != k.k->p.inode) {
571 struct bch_inode_unpacked inode;
573 /* don't hold btree locks while looking up inode: */
574 bch2_trans_unlock(&trans);
576 io_opts = bch2_opts_to_inode_opts(c->opts);
577 if (!bch2_inode_find_by_inum(c, k.k->p.inode, &inode))
578 bch2_io_opts_apply(&io_opts, bch2_inode_opts_get(&inode));
579 cur_inum = k.k->p.inode;
583 switch ((data_cmd = pred(c, arg, k, &io_opts, &data_opts))) {
588 case DATA_ADD_REPLICAS:
596 /* unlock before doing IO: */
597 bkey_on_stack_reassemble(&sk, c, k);
598 k = bkey_i_to_s_c(sk.k);
599 bch2_trans_unlock(&trans);
601 ret2 = bch2_move_extent(c, ctxt, wp, io_opts, btree_id, k,
602 data_cmd, data_opts);
604 if (ret2 == -ENOMEM) {
605 /* memory allocation failure, wait for some IO to finish */
606 bch2_move_ctxt_wait_for_io(ctxt);
610 /* XXX signal failure */
615 bch2_ratelimit_increment(rate, k.k->size);
617 atomic64_add(k.k->size * bch2_bkey_nr_ptrs_allocated(k),
618 &stats->sectors_seen);
620 bch2_btree_iter_next(iter);
621 bch2_trans_cond_resched(&trans);
624 ret = bch2_trans_exit(&trans) ?: ret;
625 bkey_on_stack_exit(&sk, c);
630 int bch2_move_data(struct bch_fs *c,
631 struct bch_ratelimit *rate,
632 struct write_point_specifier wp,
635 move_pred_fn pred, void *arg,
636 struct bch_move_stats *stats)
638 struct moving_context ctxt = { .stats = stats };
641 closure_init_stack(&ctxt.cl);
642 INIT_LIST_HEAD(&ctxt.reads);
643 init_waitqueue_head(&ctxt.wait);
645 stats->data_type = BCH_DATA_USER;
647 ret = __bch2_move_data(c, &ctxt, rate, wp, start, end,
648 pred, arg, stats, BTREE_ID_EXTENTS) ?:
649 __bch2_move_data(c, &ctxt, rate, wp, start, end,
650 pred, arg, stats, BTREE_ID_REFLINK);
652 move_ctxt_wait_event(&ctxt, list_empty(&ctxt.reads));
653 closure_sync(&ctxt.cl);
655 EBUG_ON(atomic_read(&ctxt.write_sectors));
658 atomic64_read(&stats->sectors_moved),
659 atomic64_read(&stats->keys_moved));
664 static int bch2_move_btree(struct bch_fs *c,
667 struct bch_move_stats *stats)
669 struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
670 struct btree_trans trans;
671 struct btree_iter *iter;
674 struct data_opts data_opts;
678 bch2_trans_init(&trans, c, 0, 0);
680 stats->data_type = BCH_DATA_BTREE;
682 for (id = 0; id < BTREE_ID_NR; id++) {
683 stats->btree_id = id;
685 for_each_btree_node(&trans, iter, id, POS_MIN,
686 BTREE_ITER_PREFETCH, b) {
687 stats->pos = iter->pos;
689 switch ((cmd = pred(c, arg,
690 bkey_i_to_s_c(&b->key),
691 &io_opts, &data_opts))) {
696 case DATA_ADD_REPLICAS:
703 ret = bch2_btree_node_rewrite(c, iter,
704 b->data->keys.seq, 0) ?: ret;
706 bch2_trans_cond_resched(&trans);
709 ret = bch2_trans_iter_free(&trans, iter) ?: ret;
712 bch2_trans_exit(&trans);
718 static enum data_cmd scrub_pred(struct bch_fs *c, void *arg,
720 struct bch_io_opts *io_opts,
721 struct data_opts *data_opts)
727 static enum data_cmd rereplicate_pred(struct bch_fs *c, void *arg,
729 struct bch_io_opts *io_opts,
730 struct data_opts *data_opts)
732 unsigned nr_good = bch2_bkey_durability(c, k);
733 unsigned replicas = 0;
736 case KEY_TYPE_btree_ptr:
737 replicas = c->opts.metadata_replicas;
739 case KEY_TYPE_extent:
740 replicas = io_opts->data_replicas;
744 if (!nr_good || nr_good >= replicas)
747 data_opts->target = 0;
748 data_opts->btree_insert_flags = 0;
749 return DATA_ADD_REPLICAS;
752 static enum data_cmd migrate_pred(struct bch_fs *c, void *arg,
754 struct bch_io_opts *io_opts,
755 struct data_opts *data_opts)
757 struct bch_ioctl_data *op = arg;
759 if (!bch2_bkey_has_device(k, op->migrate.dev))
762 data_opts->target = 0;
763 data_opts->btree_insert_flags = 0;
764 data_opts->rewrite_dev = op->migrate.dev;
768 int bch2_data_job(struct bch_fs *c,
769 struct bch_move_stats *stats,
770 struct bch_ioctl_data op)
775 case BCH_DATA_OP_REREPLICATE:
776 stats->data_type = BCH_DATA_JOURNAL;
777 ret = bch2_journal_flush_device_pins(&c->journal, -1);
779 ret = bch2_move_btree(c, rereplicate_pred, c, stats) ?: ret;
781 closure_wait_event(&c->btree_interior_update_wait,
782 !bch2_btree_interior_updates_nr_pending(c));
784 ret = bch2_replicas_gc2(c) ?: ret;
786 ret = bch2_move_data(c, NULL,
787 writepoint_hashed((unsigned long) current),
790 rereplicate_pred, c, stats) ?: ret;
791 ret = bch2_replicas_gc2(c) ?: ret;
793 case BCH_DATA_OP_MIGRATE:
794 if (op.migrate.dev >= c->sb.nr_devices)
797 stats->data_type = BCH_DATA_JOURNAL;
798 ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
800 ret = bch2_move_btree(c, migrate_pred, &op, stats) ?: ret;
801 ret = bch2_replicas_gc2(c) ?: ret;
803 ret = bch2_move_data(c, NULL,
804 writepoint_hashed((unsigned long) current),
807 migrate_pred, &op, stats) ?: ret;
808 ret = bch2_replicas_gc2(c) ?: ret;