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);
73 struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
74 struct bkey_i *insert;
75 struct bkey_i_extent *new =
76 bkey_i_to_extent(bch2_keylist_front(keys));
77 BKEY_PADDED(k) _new, _insert;
78 const union bch_extent_entry *entry;
79 struct extent_ptr_decoded p;
80 bool did_work = false;
87 if (bversion_cmp(k.k->version, new->k.version) ||
88 !bch2_bkey_matches_ptr(c, k, m->ptr, m->offset))
91 if (m->data_cmd == DATA_REWRITE &&
92 !bch2_bkey_has_device(k, m->data_opts.rewrite_dev))
95 bkey_reassemble(&_insert.k, k);
98 bkey_copy(&_new.k, bch2_keylist_front(keys));
99 new = bkey_i_to_extent(&_new.k);
100 bch2_cut_front(iter->pos, &new->k_i);
102 bch2_cut_front(iter->pos, insert);
103 bch2_cut_back(new->k.p, insert);
104 bch2_cut_back(insert->k.p, &new->k_i);
106 if (m->data_cmd == DATA_REWRITE)
107 bch2_bkey_drop_device(bkey_i_to_s(insert),
108 m->data_opts.rewrite_dev);
110 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry) {
111 if (bch2_bkey_has_device(bkey_i_to_s_c(insert), p.ptr.dev)) {
113 * raced with another move op? extent already
114 * has a pointer to the device we just wrote
120 bch2_extent_ptr_decoded_append(insert, &p);
127 bch2_bkey_narrow_crcs(insert,
128 (struct bch_extent_crc_unpacked) { 0 });
129 bch2_extent_normalize(c, bkey_i_to_s(insert));
130 bch2_bkey_mark_replicas_cached(c, bkey_i_to_s(insert),
131 op->opts.background_target,
132 op->opts.data_replicas);
135 * If we're not fully overwriting @k, and it's compressed, we
136 * need a reservation for all the pointers in @insert
138 nr = bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(insert)) -
141 if (insert->k.size < k.k->size &&
142 bch2_bkey_sectors_compressed(k) &&
144 ret = bch2_disk_reservation_add(c, &op->res,
145 keylist_sectors(keys) * nr, 0);
149 m->nr_ptrs_reserved += nr;
153 bch2_trans_update(&trans, iter, insert, 0);
155 ret = bch2_trans_commit(&trans, &op->res,
158 BTREE_INSERT_USE_RESERVE|
159 m->data_opts.btree_insert_flags);
161 atomic_long_inc(&c->extent_migrate_done);
167 while (bkey_cmp(iter->pos, bch2_keylist_front(keys)->k.p) >= 0) {
168 bch2_keylist_pop_front(keys);
169 if (bch2_keylist_empty(keys))
175 atomic64_add(k.k->p.offset - iter->pos.offset,
176 &m->ctxt->stats->sectors_raced);
177 atomic_long_inc(&c->extent_migrate_raced);
178 trace_move_race(&new->k);
179 bch2_btree_iter_next_slot(iter);
183 bch2_trans_exit(&trans);
184 BUG_ON(ret == -EINTR);
188 void bch2_migrate_read_done(struct migrate_write *m, struct bch_read_bio *rbio)
190 /* write bio must own pages: */
191 BUG_ON(!m->op.wbio.bio.bi_vcnt);
193 m->ptr = rbio->pick.ptr;
194 m->offset = rbio->pos.offset - rbio->pick.crc.offset;
195 m->op.devs_have = rbio->devs_have;
196 m->op.pos = rbio->pos;
197 m->op.version = rbio->version;
198 m->op.crc = rbio->pick.crc;
199 m->op.wbio.bio.bi_iter.bi_size = m->op.crc.compressed_size << 9;
201 if (bch2_csum_type_is_encryption(m->op.crc.csum_type)) {
202 m->op.nonce = m->op.crc.nonce + m->op.crc.offset;
203 m->op.csum_type = m->op.crc.csum_type;
206 if (m->data_cmd == DATA_REWRITE)
207 bch2_dev_list_drop_dev(&m->op.devs_have, m->data_opts.rewrite_dev);
210 int bch2_migrate_write_init(struct bch_fs *c, struct migrate_write *m,
211 struct write_point_specifier wp,
212 struct bch_io_opts io_opts,
213 enum data_cmd data_cmd,
214 struct data_opts data_opts,
215 enum btree_id btree_id,
218 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
219 const union bch_extent_entry *entry;
220 struct extent_ptr_decoded p;
223 m->btree_id = btree_id;
224 m->data_cmd = data_cmd;
225 m->data_opts = data_opts;
226 m->nr_ptrs_reserved = 0;
228 bch2_write_op_init(&m->op, c, io_opts);
230 if (!bch2_bkey_is_incompressible(k))
231 m->op.compression_type =
232 bch2_compression_opt_to_type[io_opts.background_compression ?:
233 io_opts.compression];
235 m->op.incompressible = true;
237 m->op.target = data_opts.target,
238 m->op.write_point = wp;
240 if (m->data_opts.btree_insert_flags & BTREE_INSERT_USE_RESERVE)
241 m->op.alloc_reserve = RESERVE_MOVINGGC;
243 m->op.flags |= BCH_WRITE_ONLY_SPECIFIED_DEVS|
244 BCH_WRITE_PAGES_STABLE|
245 BCH_WRITE_PAGES_OWNED|
246 BCH_WRITE_DATA_ENCODED|
247 BCH_WRITE_FROM_INTERNAL;
249 m->op.nr_replicas = 1;
250 m->op.nr_replicas_required = 1;
251 m->op.index_update_fn = bch2_migrate_index_update;
254 case DATA_ADD_REPLICAS: {
256 * DATA_ADD_REPLICAS is used for moving data to a different
257 * device in the background, and due to compression the new copy
258 * might take up more space than the old copy:
261 int nr = (int) io_opts.data_replicas -
262 bch2_bkey_nr_ptrs_allocated(k);
264 int nr = (int) io_opts.data_replicas;
267 m->op.nr_replicas = m->nr_ptrs_reserved = nr;
269 ret = bch2_disk_reservation_get(c, &m->op.res,
270 k.k->size, m->op.nr_replicas, 0);
277 unsigned compressed_sectors = 0;
279 bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
281 crc_is_compressed(p.crc) &&
282 bch2_dev_in_target(c, p.ptr.dev, data_opts.target))
283 compressed_sectors += p.crc.compressed_size;
285 if (compressed_sectors) {
286 ret = bch2_disk_reservation_add(c, &m->op.res,
288 BCH_DISK_RESERVATION_NOFAIL);
295 m->op.flags |= BCH_WRITE_ALLOC_NOWAIT;
296 m->op.flags |= BCH_WRITE_CACHED;
305 static void move_free(struct closure *cl)
307 struct moving_io *io = container_of(cl, struct moving_io, cl);
308 struct moving_context *ctxt = io->write.ctxt;
309 struct bvec_iter_all iter;
312 bch2_disk_reservation_put(io->write.op.c, &io->write.op.res);
314 bio_for_each_segment_all(bv, &io->write.op.wbio.bio, iter)
316 __free_page(bv->bv_page);
318 wake_up(&ctxt->wait);
323 static void move_write_done(struct closure *cl)
325 struct moving_io *io = container_of(cl, struct moving_io, cl);
327 atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
328 closure_return_with_destructor(cl, move_free);
331 static void move_write(struct closure *cl)
333 struct moving_io *io = container_of(cl, struct moving_io, cl);
335 if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
336 closure_return_with_destructor(cl, move_free);
340 bch2_migrate_read_done(&io->write, &io->rbio);
342 atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
343 closure_call(&io->write.op.cl, bch2_write, NULL, cl);
344 continue_at(cl, move_write_done, NULL);
347 static inline struct moving_io *next_pending_write(struct moving_context *ctxt)
349 struct moving_io *io =
350 list_first_entry_or_null(&ctxt->reads, struct moving_io, list);
352 return io && io->read_completed ? io : NULL;
355 static void move_read_endio(struct bio *bio)
357 struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
358 struct moving_context *ctxt = io->write.ctxt;
360 atomic_sub(io->read_sectors, &ctxt->read_sectors);
361 io->read_completed = true;
363 if (next_pending_write(ctxt))
364 wake_up(&ctxt->wait);
366 closure_put(&ctxt->cl);
369 static void do_pending_writes(struct moving_context *ctxt)
371 struct moving_io *io;
373 while ((io = next_pending_write(ctxt))) {
375 closure_call(&io->cl, move_write, NULL, &ctxt->cl);
379 #define move_ctxt_wait_event(_ctxt, _cond) \
381 do_pending_writes(_ctxt); \
385 __wait_event((_ctxt)->wait, \
386 next_pending_write(_ctxt) || (_cond)); \
389 static void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt)
391 unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
393 move_ctxt_wait_event(ctxt,
394 !atomic_read(&ctxt->write_sectors) ||
395 atomic_read(&ctxt->write_sectors) != sectors_pending);
398 static int bch2_move_extent(struct bch_fs *c,
399 struct moving_context *ctxt,
400 struct write_point_specifier wp,
401 struct bch_io_opts io_opts,
402 enum btree_id btree_id,
404 enum data_cmd data_cmd,
405 struct data_opts data_opts)
407 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
408 struct moving_io *io;
409 const union bch_extent_entry *entry;
410 struct extent_ptr_decoded p;
411 unsigned sectors = k.k->size, pages;
414 move_ctxt_wait_event(ctxt,
415 atomic_read(&ctxt->write_sectors) <
416 SECTORS_IN_FLIGHT_PER_DEVICE);
418 move_ctxt_wait_event(ctxt,
419 atomic_read(&ctxt->read_sectors) <
420 SECTORS_IN_FLIGHT_PER_DEVICE);
422 /* write path might have to decompress data: */
423 bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
424 sectors = max_t(unsigned, sectors, p.crc.uncompressed_size);
426 pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
427 io = kzalloc(sizeof(struct moving_io) +
428 sizeof(struct bio_vec) * pages, GFP_KERNEL);
432 io->write.ctxt = ctxt;
433 io->read_sectors = k.k->size;
434 io->write_sectors = k.k->size;
436 bio_init(&io->write.op.wbio.bio, io->bi_inline_vecs, pages);
437 bio_set_prio(&io->write.op.wbio.bio,
438 IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
440 if (bch2_bio_alloc_pages(&io->write.op.wbio.bio, sectors << 9,
445 io->rbio.opts = io_opts;
446 bio_init(&io->rbio.bio, io->bi_inline_vecs, pages);
447 io->rbio.bio.bi_vcnt = pages;
448 bio_set_prio(&io->rbio.bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
449 io->rbio.bio.bi_iter.bi_size = sectors << 9;
451 bio_set_op_attrs(&io->rbio.bio, REQ_OP_READ, 0);
452 io->rbio.bio.bi_iter.bi_sector = bkey_start_offset(k.k);
453 io->rbio.bio.bi_end_io = move_read_endio;
455 ret = bch2_migrate_write_init(c, &io->write, wp, io_opts,
456 data_cmd, data_opts, btree_id, k);
460 atomic64_inc(&ctxt->stats->keys_moved);
461 atomic64_add(k.k->size, &ctxt->stats->sectors_moved);
463 trace_move_extent(k.k);
465 atomic_add(io->read_sectors, &ctxt->read_sectors);
466 list_add_tail(&io->list, &ctxt->reads);
469 * dropped by move_read_endio() - guards against use after free of
470 * ctxt when doing wakeup
472 closure_get(&ctxt->cl);
473 bch2_read_extent(c, &io->rbio, k, 0,
475 BCH_READ_LAST_FRAGMENT);
478 bio_free_pages(&io->write.op.wbio.bio);
482 trace_move_alloc_fail(k.k);
486 static int __bch2_move_data(struct bch_fs *c,
487 struct moving_context *ctxt,
488 struct bch_ratelimit *rate,
489 struct write_point_specifier wp,
492 move_pred_fn pred, void *arg,
493 struct bch_move_stats *stats,
494 enum btree_id btree_id)
496 bool kthread = (current->flags & PF_KTHREAD) != 0;
497 struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
498 struct bkey_on_stack sk;
499 struct btree_trans trans;
500 struct btree_iter *iter;
502 struct data_opts data_opts;
503 enum data_cmd data_cmd;
504 u64 delay, cur_inum = U64_MAX;
507 bkey_on_stack_init(&sk);
508 bch2_trans_init(&trans, c, 0, 0);
510 stats->data_type = BCH_DATA_USER;
511 stats->btree_id = btree_id;
512 stats->pos = POS_MIN;
514 iter = bch2_trans_get_iter(&trans, btree_id, start,
515 BTREE_ITER_PREFETCH);
518 bch2_ratelimit_reset(rate);
522 delay = rate ? bch2_ratelimit_delay(rate) : 0;
525 bch2_trans_unlock(&trans);
526 set_current_state(TASK_INTERRUPTIBLE);
529 if (kthread && (ret = kthread_should_stop())) {
530 __set_current_state(TASK_RUNNING);
535 schedule_timeout(delay);
537 if (unlikely(freezing(current))) {
538 bch2_trans_unlock(&trans);
539 move_ctxt_wait_event(ctxt, list_empty(&ctxt->reads));
544 k = bch2_btree_iter_peek(iter);
546 stats->pos = iter->pos;
553 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
556 if (!bkey_extent_is_direct_data(k.k))
559 if (btree_id == BTREE_ID_EXTENTS &&
560 cur_inum != k.k->p.inode) {
561 struct bch_inode_unpacked inode;
563 /* don't hold btree locks while looking up inode: */
564 bch2_trans_unlock(&trans);
566 io_opts = bch2_opts_to_inode_opts(c->opts);
567 if (!bch2_inode_find_by_inum(c, k.k->p.inode, &inode))
568 bch2_io_opts_apply(&io_opts, bch2_inode_opts_get(&inode));
569 cur_inum = k.k->p.inode;
573 switch ((data_cmd = pred(c, arg, k, &io_opts, &data_opts))) {
578 case DATA_ADD_REPLICAS:
586 /* unlock before doing IO: */
587 bkey_on_stack_reassemble(&sk, c, k);
588 k = bkey_i_to_s_c(sk.k);
589 bch2_trans_unlock(&trans);
591 ret2 = bch2_move_extent(c, ctxt, wp, io_opts, btree_id, k,
592 data_cmd, data_opts);
594 if (ret2 == -ENOMEM) {
595 /* memory allocation failure, wait for some IO to finish */
596 bch2_move_ctxt_wait_for_io(ctxt);
600 /* XXX signal failure */
605 bch2_ratelimit_increment(rate, k.k->size);
607 atomic64_add(k.k->size * bch2_bkey_nr_ptrs_allocated(k),
608 &stats->sectors_seen);
610 bch2_btree_iter_next(iter);
611 bch2_trans_cond_resched(&trans);
614 ret = bch2_trans_exit(&trans) ?: ret;
615 bkey_on_stack_exit(&sk, c);
620 int bch2_move_data(struct bch_fs *c,
621 struct bch_ratelimit *rate,
622 struct write_point_specifier wp,
625 move_pred_fn pred, void *arg,
626 struct bch_move_stats *stats)
628 struct moving_context ctxt = { .stats = stats };
631 closure_init_stack(&ctxt.cl);
632 INIT_LIST_HEAD(&ctxt.reads);
633 init_waitqueue_head(&ctxt.wait);
635 stats->data_type = BCH_DATA_USER;
637 ret = __bch2_move_data(c, &ctxt, rate, wp, start, end,
638 pred, arg, stats, BTREE_ID_EXTENTS) ?:
639 __bch2_move_data(c, &ctxt, rate, wp, start, end,
640 pred, arg, stats, BTREE_ID_REFLINK);
642 move_ctxt_wait_event(&ctxt, list_empty(&ctxt.reads));
643 closure_sync(&ctxt.cl);
645 EBUG_ON(atomic_read(&ctxt.write_sectors));
648 atomic64_read(&stats->sectors_moved),
649 atomic64_read(&stats->keys_moved));
654 static int bch2_move_btree(struct bch_fs *c,
657 struct bch_move_stats *stats)
659 struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
660 struct btree_trans trans;
661 struct btree_iter *iter;
664 struct data_opts data_opts;
668 bch2_trans_init(&trans, c, 0, 0);
670 stats->data_type = BCH_DATA_BTREE;
672 for (id = 0; id < BTREE_ID_NR; id++) {
673 stats->btree_id = id;
675 for_each_btree_node(&trans, iter, id, POS_MIN,
676 BTREE_ITER_PREFETCH, b) {
677 stats->pos = iter->pos;
679 switch ((cmd = pred(c, arg,
680 bkey_i_to_s_c(&b->key),
681 &io_opts, &data_opts))) {
686 case DATA_ADD_REPLICAS:
693 ret = bch2_btree_node_rewrite(c, iter,
694 b->data->keys.seq, 0) ?: ret;
696 bch2_trans_cond_resched(&trans);
699 ret = bch2_trans_iter_free(&trans, iter) ?: ret;
702 bch2_trans_exit(&trans);
708 static enum data_cmd scrub_pred(struct bch_fs *c, void *arg,
710 struct bch_io_opts *io_opts,
711 struct data_opts *data_opts)
717 static enum data_cmd rereplicate_pred(struct bch_fs *c, void *arg,
719 struct bch_io_opts *io_opts,
720 struct data_opts *data_opts)
722 unsigned nr_good = bch2_bkey_durability(c, k);
723 unsigned replicas = 0;
726 case KEY_TYPE_btree_ptr:
727 replicas = c->opts.metadata_replicas;
729 case KEY_TYPE_extent:
730 replicas = io_opts->data_replicas;
734 if (!nr_good || nr_good >= replicas)
737 data_opts->target = 0;
738 data_opts->btree_insert_flags = 0;
739 return DATA_ADD_REPLICAS;
742 static enum data_cmd migrate_pred(struct bch_fs *c, void *arg,
744 struct bch_io_opts *io_opts,
745 struct data_opts *data_opts)
747 struct bch_ioctl_data *op = arg;
749 if (!bch2_bkey_has_device(k, op->migrate.dev))
752 data_opts->target = 0;
753 data_opts->btree_insert_flags = 0;
754 data_opts->rewrite_dev = op->migrate.dev;
758 int bch2_data_job(struct bch_fs *c,
759 struct bch_move_stats *stats,
760 struct bch_ioctl_data op)
765 case BCH_DATA_OP_REREPLICATE:
766 stats->data_type = BCH_DATA_JOURNAL;
767 ret = bch2_journal_flush_device_pins(&c->journal, -1);
769 ret = bch2_move_btree(c, rereplicate_pred, c, stats) ?: ret;
772 closure_wait_event(&c->btree_interior_update_wait,
773 !bch2_btree_interior_updates_nr_pending(c) ||
774 c->btree_roots_dirty);
775 if (!bch2_btree_interior_updates_nr_pending(c))
777 bch2_journal_meta(&c->journal);
780 ret = bch2_replicas_gc2(c) ?: ret;
782 ret = bch2_move_data(c, NULL,
783 writepoint_hashed((unsigned long) current),
786 rereplicate_pred, c, stats) ?: ret;
787 ret = bch2_replicas_gc2(c) ?: ret;
789 case BCH_DATA_OP_MIGRATE:
790 if (op.migrate.dev >= c->sb.nr_devices)
793 stats->data_type = BCH_DATA_JOURNAL;
794 ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
796 ret = bch2_move_btree(c, migrate_pred, &op, stats) ?: ret;
797 ret = bch2_replicas_gc2(c) ?: ret;
799 ret = bch2_move_data(c, NULL,
800 writepoint_hashed((unsigned long) current),
803 migrate_pred, &op, stats) ?: ret;
804 ret = bch2_replicas_gc2(c) ?: ret;