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 bkey_reassemble(&_insert.k, k);
101 bkey_copy(&_new.k, bch2_keylist_front(keys));
102 new = bkey_i_to_extent(&_new.k);
103 bch2_cut_front(iter->pos, &new->k_i);
105 bch2_cut_front(iter->pos, insert);
106 bch2_cut_back(new->k.p, insert);
107 bch2_cut_back(insert->k.p, &new->k_i);
109 if (m->data_cmd == DATA_REWRITE) {
110 struct bch_extent_ptr *new_ptr, *old_ptr = (void *)
111 bch2_bkey_has_device(bkey_i_to_s_c(insert),
112 m->data_opts.rewrite_dev);
117 extent_for_each_ptr(extent_i_to_s(new), new_ptr)
118 new_ptr->cached = true;
120 bch2_bkey_drop_ptr(bkey_i_to_s(insert), old_ptr);
123 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry) {
124 if (bch2_bkey_has_device(bkey_i_to_s_c(insert), p.ptr.dev)) {
126 * raced with another move op? extent already
127 * has a pointer to the device we just wrote
133 bch2_extent_ptr_decoded_append(insert, &p);
140 bch2_bkey_narrow_crcs(insert,
141 (struct bch_extent_crc_unpacked) { 0 });
142 bch2_extent_normalize(c, bkey_i_to_s(insert));
143 bch2_bkey_mark_replicas_cached(c, bkey_i_to_s(insert),
144 op->opts.background_target,
145 op->opts.data_replicas);
148 * If we're not fully overwriting @k, and it's compressed, we
149 * need a reservation for all the pointers in @insert
151 nr = bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(insert)) -
154 if (insert->k.size < k.k->size &&
155 bch2_bkey_sectors_compressed(k) &&
157 ret = bch2_disk_reservation_add(c, &op->res,
158 keylist_sectors(keys) * nr, 0);
162 m->nr_ptrs_reserved += nr;
166 bch2_trans_update(&trans, iter, insert, 0);
168 ret = bch2_trans_commit(&trans, &op->res,
171 BTREE_INSERT_USE_RESERVE|
172 m->data_opts.btree_insert_flags);
174 atomic_long_inc(&c->extent_migrate_done);
180 while (bkey_cmp(iter->pos, bch2_keylist_front(keys)->k.p) >= 0) {
181 bch2_keylist_pop_front(keys);
182 if (bch2_keylist_empty(keys))
188 BUG_ON(k.k->p.offset <= iter->pos.offset);
189 atomic64_inc(&m->ctxt->stats->keys_raced);
190 atomic64_add(k.k->p.offset - iter->pos.offset,
191 &m->ctxt->stats->sectors_raced);
193 atomic_long_inc(&c->extent_migrate_raced);
194 trace_move_race(&new->k);
195 bch2_btree_iter_next_slot(iter);
199 bch2_trans_exit(&trans);
200 BUG_ON(ret == -EINTR);
204 void bch2_migrate_read_done(struct migrate_write *m, struct bch_read_bio *rbio)
206 /* write bio must own pages: */
207 BUG_ON(!m->op.wbio.bio.bi_vcnt);
209 m->ptr = rbio->pick.ptr;
210 m->offset = rbio->pos.offset - rbio->pick.crc.offset;
211 m->op.devs_have = rbio->devs_have;
212 m->op.pos = rbio->pos;
213 m->op.version = rbio->version;
214 m->op.crc = rbio->pick.crc;
215 m->op.wbio.bio.bi_iter.bi_size = m->op.crc.compressed_size << 9;
217 if (bch2_csum_type_is_encryption(m->op.crc.csum_type)) {
218 m->op.nonce = m->op.crc.nonce + m->op.crc.offset;
219 m->op.csum_type = m->op.crc.csum_type;
222 if (m->data_cmd == DATA_REWRITE)
223 bch2_dev_list_drop_dev(&m->op.devs_have, m->data_opts.rewrite_dev);
226 int bch2_migrate_write_init(struct bch_fs *c, struct migrate_write *m,
227 struct write_point_specifier wp,
228 struct bch_io_opts io_opts,
229 enum data_cmd data_cmd,
230 struct data_opts data_opts,
231 enum btree_id btree_id,
234 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
235 const union bch_extent_entry *entry;
236 struct extent_ptr_decoded p;
239 m->btree_id = btree_id;
240 m->data_cmd = data_cmd;
241 m->data_opts = data_opts;
242 m->nr_ptrs_reserved = 0;
244 bch2_write_op_init(&m->op, c, io_opts);
246 if (!bch2_bkey_is_incompressible(k))
247 m->op.compression_type =
248 bch2_compression_opt_to_type[io_opts.background_compression ?:
249 io_opts.compression];
251 m->op.incompressible = true;
253 m->op.target = data_opts.target,
254 m->op.write_point = wp;
256 if (m->data_opts.btree_insert_flags & BTREE_INSERT_USE_RESERVE) {
257 m->op.alloc_reserve = RESERVE_MOVINGGC;
258 m->op.flags |= BCH_WRITE_ALLOC_NOWAIT;
260 /* XXX: this should probably be passed in */
261 m->op.flags |= BCH_WRITE_ONLY_SPECIFIED_DEVS;
264 m->op.flags |= BCH_WRITE_PAGES_STABLE|
265 BCH_WRITE_PAGES_OWNED|
266 BCH_WRITE_DATA_ENCODED|
267 BCH_WRITE_FROM_INTERNAL;
269 m->op.nr_replicas = data_opts.nr_replicas;
270 m->op.nr_replicas_required = data_opts.nr_replicas;
271 m->op.index_update_fn = bch2_migrate_index_update;
274 case DATA_ADD_REPLICAS: {
276 * DATA_ADD_REPLICAS is used for moving data to a different
277 * device in the background, and due to compression the new copy
278 * might take up more space than the old copy:
281 int nr = (int) io_opts.data_replicas -
282 bch2_bkey_nr_ptrs_allocated(k);
284 int nr = (int) io_opts.data_replicas;
287 m->op.nr_replicas = m->nr_ptrs_reserved = nr;
289 ret = bch2_disk_reservation_get(c, &m->op.res,
290 k.k->size, m->op.nr_replicas, 0);
297 unsigned compressed_sectors = 0;
299 bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
300 if (p.ptr.dev == data_opts.rewrite_dev &&
302 crc_is_compressed(p.crc))
303 compressed_sectors += p.crc.compressed_size;
305 if (compressed_sectors) {
306 ret = bch2_disk_reservation_add(c, &m->op.res,
307 k.k->size * m->op.nr_replicas,
308 BCH_DISK_RESERVATION_NOFAIL);
315 m->op.flags |= BCH_WRITE_ALLOC_NOWAIT;
316 m->op.flags |= BCH_WRITE_CACHED;
325 static void move_free(struct closure *cl)
327 struct moving_io *io = container_of(cl, struct moving_io, cl);
328 struct moving_context *ctxt = io->write.ctxt;
329 struct bvec_iter_all iter;
332 bch2_disk_reservation_put(io->write.op.c, &io->write.op.res);
334 bio_for_each_segment_all(bv, &io->write.op.wbio.bio, iter)
336 __free_page(bv->bv_page);
338 wake_up(&ctxt->wait);
343 static void move_write_done(struct closure *cl)
345 struct moving_io *io = container_of(cl, struct moving_io, cl);
347 atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
348 closure_return_with_destructor(cl, move_free);
351 static void move_write(struct closure *cl)
353 struct moving_io *io = container_of(cl, struct moving_io, cl);
355 if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
356 closure_return_with_destructor(cl, move_free);
360 bch2_migrate_read_done(&io->write, &io->rbio);
362 atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
363 closure_call(&io->write.op.cl, bch2_write, NULL, cl);
364 continue_at(cl, move_write_done, NULL);
367 static inline struct moving_io *next_pending_write(struct moving_context *ctxt)
369 struct moving_io *io =
370 list_first_entry_or_null(&ctxt->reads, struct moving_io, list);
372 return io && io->read_completed ? io : NULL;
375 static void move_read_endio(struct bio *bio)
377 struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
378 struct moving_context *ctxt = io->write.ctxt;
380 atomic_sub(io->read_sectors, &ctxt->read_sectors);
381 io->read_completed = true;
383 if (next_pending_write(ctxt))
384 wake_up(&ctxt->wait);
386 closure_put(&ctxt->cl);
389 static void do_pending_writes(struct moving_context *ctxt)
391 struct moving_io *io;
393 while ((io = next_pending_write(ctxt))) {
395 closure_call(&io->cl, move_write, NULL, &ctxt->cl);
399 #define move_ctxt_wait_event(_ctxt, _cond) \
401 do_pending_writes(_ctxt); \
405 __wait_event((_ctxt)->wait, \
406 next_pending_write(_ctxt) || (_cond)); \
409 static void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt)
411 unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
413 move_ctxt_wait_event(ctxt,
414 !atomic_read(&ctxt->write_sectors) ||
415 atomic_read(&ctxt->write_sectors) != sectors_pending);
418 static int bch2_move_extent(struct btree_trans *trans,
419 struct moving_context *ctxt,
420 struct write_point_specifier wp,
421 struct bch_io_opts io_opts,
422 enum btree_id btree_id,
424 enum data_cmd data_cmd,
425 struct data_opts data_opts)
427 struct bch_fs *c = trans->c;
428 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
429 struct moving_io *io;
430 const union bch_extent_entry *entry;
431 struct extent_ptr_decoded p;
432 unsigned sectors = k.k->size, pages;
435 move_ctxt_wait_event(ctxt,
436 atomic_read(&ctxt->write_sectors) <
437 SECTORS_IN_FLIGHT_PER_DEVICE);
439 move_ctxt_wait_event(ctxt,
440 atomic_read(&ctxt->read_sectors) <
441 SECTORS_IN_FLIGHT_PER_DEVICE);
443 /* write path might have to decompress data: */
444 bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
445 sectors = max_t(unsigned, sectors, p.crc.uncompressed_size);
447 pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
448 io = kzalloc(sizeof(struct moving_io) +
449 sizeof(struct bio_vec) * pages, GFP_KERNEL);
453 io->write.ctxt = ctxt;
454 io->read_sectors = k.k->size;
455 io->write_sectors = k.k->size;
457 bio_init(&io->write.op.wbio.bio, io->bi_inline_vecs, pages);
458 bio_set_prio(&io->write.op.wbio.bio,
459 IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
461 if (bch2_bio_alloc_pages(&io->write.op.wbio.bio, sectors << 9,
466 io->rbio.opts = io_opts;
467 bio_init(&io->rbio.bio, io->bi_inline_vecs, pages);
468 io->rbio.bio.bi_vcnt = pages;
469 bio_set_prio(&io->rbio.bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
470 io->rbio.bio.bi_iter.bi_size = sectors << 9;
472 bio_set_op_attrs(&io->rbio.bio, REQ_OP_READ, 0);
473 io->rbio.bio.bi_iter.bi_sector = bkey_start_offset(k.k);
474 io->rbio.bio.bi_end_io = move_read_endio;
476 ret = bch2_migrate_write_init(c, &io->write, wp, io_opts,
477 data_cmd, data_opts, btree_id, k);
481 atomic64_inc(&ctxt->stats->keys_moved);
482 atomic64_add(k.k->size, &ctxt->stats->sectors_moved);
484 trace_move_extent(k.k);
486 atomic_add(io->read_sectors, &ctxt->read_sectors);
487 list_add_tail(&io->list, &ctxt->reads);
490 * dropped by move_read_endio() - guards against use after free of
491 * ctxt when doing wakeup
493 closure_get(&ctxt->cl);
494 bch2_read_extent(trans, &io->rbio, k, 0,
496 BCH_READ_LAST_FRAGMENT);
499 bio_free_pages(&io->write.op.wbio.bio);
503 trace_move_alloc_fail(k.k);
507 static int __bch2_move_data(struct bch_fs *c,
508 struct moving_context *ctxt,
509 struct bch_ratelimit *rate,
510 struct write_point_specifier wp,
513 move_pred_fn pred, void *arg,
514 struct bch_move_stats *stats,
515 enum btree_id btree_id)
517 bool kthread = (current->flags & PF_KTHREAD) != 0;
518 struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
519 struct bkey_on_stack sk;
520 struct btree_trans trans;
521 struct btree_iter *iter;
523 struct data_opts data_opts;
524 enum data_cmd data_cmd;
525 u64 delay, cur_inum = U64_MAX;
528 bkey_on_stack_init(&sk);
529 bch2_trans_init(&trans, c, 0, 0);
531 stats->data_type = BCH_DATA_user;
532 stats->btree_id = btree_id;
533 stats->pos = POS_MIN;
535 iter = bch2_trans_get_iter(&trans, btree_id, start,
536 BTREE_ITER_PREFETCH);
539 bch2_ratelimit_reset(rate);
543 delay = rate ? bch2_ratelimit_delay(rate) : 0;
546 bch2_trans_unlock(&trans);
547 set_current_state(TASK_INTERRUPTIBLE);
550 if (kthread && (ret = kthread_should_stop())) {
551 __set_current_state(TASK_RUNNING);
556 schedule_timeout(delay);
558 if (unlikely(freezing(current))) {
559 bch2_trans_unlock(&trans);
560 move_ctxt_wait_event(ctxt, list_empty(&ctxt->reads));
565 k = bch2_btree_iter_peek(iter);
567 stats->pos = iter->pos;
574 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
577 if (!bkey_extent_is_direct_data(k.k))
580 if (btree_id == BTREE_ID_EXTENTS &&
581 cur_inum != k.k->p.inode) {
582 struct bch_inode_unpacked inode;
584 /* don't hold btree locks while looking up inode: */
585 bch2_trans_unlock(&trans);
587 io_opts = bch2_opts_to_inode_opts(c->opts);
588 if (!bch2_inode_find_by_inum(c, k.k->p.inode, &inode))
589 bch2_io_opts_apply(&io_opts, bch2_inode_opts_get(&inode));
590 cur_inum = k.k->p.inode;
594 switch ((data_cmd = pred(c, arg, k, &io_opts, &data_opts))) {
599 case DATA_ADD_REPLICAS:
607 /* unlock before doing IO: */
608 bkey_on_stack_reassemble(&sk, c, k);
609 k = bkey_i_to_s_c(sk.k);
610 bch2_trans_unlock(&trans);
612 ret2 = bch2_move_extent(&trans, ctxt, wp, io_opts, btree_id, k,
613 data_cmd, data_opts);
615 if (ret2 == -ENOMEM) {
616 /* memory allocation failure, wait for some IO to finish */
617 bch2_move_ctxt_wait_for_io(ctxt);
621 /* XXX signal failure */
626 bch2_ratelimit_increment(rate, k.k->size);
628 atomic64_add(k.k->size * bch2_bkey_nr_ptrs_allocated(k),
629 &stats->sectors_seen);
631 bch2_btree_iter_next(iter);
632 bch2_trans_cond_resched(&trans);
635 ret = bch2_trans_exit(&trans) ?: ret;
636 bkey_on_stack_exit(&sk, c);
641 int bch2_move_data(struct bch_fs *c,
642 struct bch_ratelimit *rate,
643 struct write_point_specifier wp,
646 move_pred_fn pred, void *arg,
647 struct bch_move_stats *stats)
649 struct moving_context ctxt = { .stats = stats };
652 closure_init_stack(&ctxt.cl);
653 INIT_LIST_HEAD(&ctxt.reads);
654 init_waitqueue_head(&ctxt.wait);
656 stats->data_type = BCH_DATA_user;
658 ret = __bch2_move_data(c, &ctxt, rate, wp, start, end,
659 pred, arg, stats, BTREE_ID_EXTENTS) ?:
660 __bch2_move_data(c, &ctxt, rate, wp, start, end,
661 pred, arg, stats, BTREE_ID_REFLINK);
663 move_ctxt_wait_event(&ctxt, list_empty(&ctxt.reads));
664 closure_sync(&ctxt.cl);
666 EBUG_ON(atomic_read(&ctxt.write_sectors));
669 atomic64_read(&stats->sectors_moved),
670 atomic64_read(&stats->keys_moved));
675 static int bch2_move_btree(struct bch_fs *c,
678 struct bch_move_stats *stats)
680 struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
681 struct btree_trans trans;
682 struct btree_iter *iter;
685 struct data_opts data_opts;
689 bch2_trans_init(&trans, c, 0, 0);
691 stats->data_type = BCH_DATA_btree;
693 for (id = 0; id < BTREE_ID_NR; id++) {
694 stats->btree_id = id;
696 for_each_btree_node(&trans, iter, id, POS_MIN,
697 BTREE_ITER_PREFETCH, b) {
698 stats->pos = iter->pos;
700 switch ((cmd = pred(c, arg,
701 bkey_i_to_s_c(&b->key),
702 &io_opts, &data_opts))) {
707 case DATA_ADD_REPLICAS:
714 ret = bch2_btree_node_rewrite(c, iter,
715 b->data->keys.seq, 0) ?: ret;
717 bch2_trans_cond_resched(&trans);
720 ret = bch2_trans_iter_free(&trans, iter) ?: ret;
723 bch2_trans_exit(&trans);
729 static enum data_cmd scrub_pred(struct bch_fs *c, void *arg,
731 struct bch_io_opts *io_opts,
732 struct data_opts *data_opts)
738 static enum data_cmd rereplicate_pred(struct bch_fs *c, void *arg,
740 struct bch_io_opts *io_opts,
741 struct data_opts *data_opts)
743 unsigned nr_good = bch2_bkey_durability(c, k);
744 unsigned replicas = 0;
747 case KEY_TYPE_btree_ptr:
748 replicas = c->opts.metadata_replicas;
750 case KEY_TYPE_extent:
751 replicas = io_opts->data_replicas;
755 if (!nr_good || nr_good >= replicas)
758 data_opts->target = 0;
759 data_opts->nr_replicas = 1;
760 data_opts->btree_insert_flags = 0;
761 return DATA_ADD_REPLICAS;
764 static enum data_cmd migrate_pred(struct bch_fs *c, void *arg,
766 struct bch_io_opts *io_opts,
767 struct data_opts *data_opts)
769 struct bch_ioctl_data *op = arg;
771 if (!bch2_bkey_has_device(k, op->migrate.dev))
774 data_opts->target = 0;
775 data_opts->nr_replicas = 1;
776 data_opts->btree_insert_flags = 0;
777 data_opts->rewrite_dev = op->migrate.dev;
781 int bch2_data_job(struct bch_fs *c,
782 struct bch_move_stats *stats,
783 struct bch_ioctl_data op)
788 case BCH_DATA_OP_REREPLICATE:
789 stats->data_type = BCH_DATA_journal;
790 ret = bch2_journal_flush_device_pins(&c->journal, -1);
792 ret = bch2_move_btree(c, rereplicate_pred, c, stats) ?: ret;
794 closure_wait_event(&c->btree_interior_update_wait,
795 !bch2_btree_interior_updates_nr_pending(c));
797 ret = bch2_replicas_gc2(c) ?: ret;
799 ret = bch2_move_data(c, NULL,
800 writepoint_hashed((unsigned long) current),
803 rereplicate_pred, c, stats) ?: ret;
804 ret = bch2_replicas_gc2(c) ?: ret;
806 case BCH_DATA_OP_MIGRATE:
807 if (op.migrate.dev >= c->sb.nr_devices)
810 stats->data_type = BCH_DATA_journal;
811 ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
813 ret = bch2_move_btree(c, migrate_pred, &op, stats) ?: ret;
814 ret = bch2_replicas_gc2(c) ?: ret;
816 ret = bch2_move_data(c, NULL,
817 writepoint_hashed((unsigned long) current),
820 migrate_pred, &op, stats) ?: ret;
821 ret = bch2_replicas_gc2(c) ?: ret;