1 // SPDX-License-Identifier: GPL-2.0
4 #include "alloc_foreground.h"
6 #include "btree_update.h"
8 #include "data_update.h"
15 #include "nocow_locking.h"
16 #include "subvolume.h"
19 static int insert_snapshot_whiteouts(struct btree_trans *trans,
24 struct bch_fs *c = trans->c;
25 struct btree_iter iter, iter2;
26 struct bkey_s_c k, k2;
28 struct bkey_i *update;
31 if (!btree_type_has_snapshots(id))
36 if (!bch2_snapshot_has_children(c, old_pos.snapshot))
39 bch2_trans_iter_init(trans, &iter, id, old_pos,
40 BTREE_ITER_NOT_EXTENTS|
41 BTREE_ITER_ALL_SNAPSHOTS);
43 k = bch2_btree_iter_prev(&iter);
51 if (!bkey_eq(old_pos, k.k->p))
54 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, old_pos.snapshot) &&
55 !snapshot_list_has_ancestor(c, &s, k.k->p.snapshot)) {
56 struct bpos whiteout_pos = new_pos;
58 whiteout_pos.snapshot = k.k->p.snapshot;
60 k2 = bch2_bkey_get_iter(trans, &iter2, id, whiteout_pos,
61 BTREE_ITER_NOT_EXTENTS|
65 if (!ret && k2.k->type == KEY_TYPE_deleted) {
66 update = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
67 ret = PTR_ERR_OR_ZERO(update);
71 bkey_init(&update->k);
72 update->k.p = whiteout_pos;
73 update->k.type = KEY_TYPE_whiteout;
75 ret = bch2_trans_update(trans, &iter2, update,
76 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
78 bch2_trans_iter_exit(trans, &iter2);
83 ret = snapshot_list_add(c, &s, k.k->p.snapshot);
88 bch2_trans_iter_exit(trans, &iter);
94 static void trace_move_extent_finish2(struct bch_fs *c, struct bkey_s_c k)
96 if (trace_move_extent_finish_enabled()) {
97 struct printbuf buf = PRINTBUF;
99 bch2_bkey_val_to_text(&buf, c, k);
100 trace_move_extent_finish(c, buf.buf);
105 static void trace_move_extent_fail2(struct data_update *m,
107 struct bkey_s_c wrote,
108 struct bkey_i *insert,
111 struct bch_fs *c = m->op.c;
112 struct bkey_s_c old = bkey_i_to_s_c(m->k.k);
113 const union bch_extent_entry *entry;
114 struct bch_extent_ptr *ptr;
115 struct extent_ptr_decoded p;
116 struct printbuf buf = PRINTBUF;
117 unsigned i, rewrites_found = 0;
119 if (!trace_move_extent_fail_enabled())
126 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry) {
128 new_s.k = (void *) new.k;
129 new_s.v = (void *) new.v;
131 if (((1U << i) & m->data_opts.rewrite_ptrs) &&
132 (ptr = bch2_extent_has_ptr(old, p, bkey_i_to_s(insert))) &&
134 rewrites_found |= 1U << i;
139 prt_printf(&buf, "\nrewrite ptrs: %u%u%u%u",
140 (m->data_opts.rewrite_ptrs & (1 << 0)) != 0,
141 (m->data_opts.rewrite_ptrs & (1 << 1)) != 0,
142 (m->data_opts.rewrite_ptrs & (1 << 2)) != 0,
143 (m->data_opts.rewrite_ptrs & (1 << 3)) != 0);
145 prt_printf(&buf, "\nrewrites found: %u%u%u%u",
146 (rewrites_found & (1 << 0)) != 0,
147 (rewrites_found & (1 << 1)) != 0,
148 (rewrites_found & (1 << 2)) != 0,
149 (rewrites_found & (1 << 3)) != 0);
151 prt_str(&buf, "\nold: ");
152 bch2_bkey_val_to_text(&buf, c, old);
154 prt_str(&buf, "\nnew: ");
155 bch2_bkey_val_to_text(&buf, c, new);
157 prt_str(&buf, "\nwrote: ");
158 bch2_bkey_val_to_text(&buf, c, wrote);
161 prt_str(&buf, "\ninsert: ");
162 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
165 trace_move_extent_fail(c, buf.buf);
169 static int __bch2_data_update_index_update(struct btree_trans *trans,
170 struct bch_write_op *op)
172 struct bch_fs *c = op->c;
173 struct btree_iter iter;
174 struct data_update *m =
175 container_of(op, struct data_update, op);
176 struct keylist *keys = &op->insert_keys;
177 struct bkey_buf _new, _insert;
180 bch2_bkey_buf_init(&_new);
181 bch2_bkey_buf_init(&_insert);
182 bch2_bkey_buf_realloc(&_insert, c, U8_MAX);
184 bch2_trans_iter_init(trans, &iter, m->btree_id,
185 bkey_start_pos(&bch2_keylist_front(keys)->k),
186 BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
190 struct bkey_s_c old = bkey_i_to_s_c(m->k.k);
191 struct bkey_i *insert = NULL;
192 struct bkey_i_extent *new;
193 const union bch_extent_entry *entry_c;
194 union bch_extent_entry *entry;
195 struct extent_ptr_decoded p;
196 struct bch_extent_ptr *ptr;
197 const struct bch_extent_ptr *ptr_c;
198 struct bpos next_pos;
199 bool should_check_enospc;
200 s64 i_sectors_delta = 0, disk_sectors_delta = 0;
201 unsigned rewrites_found = 0, durability, i;
203 bch2_trans_begin(trans);
205 k = bch2_btree_iter_peek_slot(&iter);
210 new = bkey_i_to_extent(bch2_keylist_front(keys));
212 if (!bch2_extents_match(k, old)) {
213 trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i),
218 bkey_reassemble(_insert.k, k);
221 bch2_bkey_buf_copy(&_new, c, bch2_keylist_front(keys));
222 new = bkey_i_to_extent(_new.k);
223 bch2_cut_front(iter.pos, &new->k_i);
225 bch2_cut_front(iter.pos, insert);
226 bch2_cut_back(new->k.p, insert);
227 bch2_cut_back(insert->k.p, &new->k_i);
230 * @old: extent that we read from
231 * @insert: key that we're going to update, initialized from
232 * extent currently in btree - same as @old unless we raced with
234 * @new: extent with new pointers that we'll be adding to @insert
236 * Fist, drop rewrite_ptrs from @new:
239 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry_c) {
240 if (((1U << i) & m->data_opts.rewrite_ptrs) &&
241 (ptr = bch2_extent_has_ptr(old, p, bkey_i_to_s(insert))) &&
243 bch2_bkey_drop_ptr_noerror(bkey_i_to_s(insert), ptr);
246 bch2_extent_ptr_set_cached(bkey_i_to_s(insert), ptr);
248 rewrites_found |= 1U << i;
253 if (m->data_opts.rewrite_ptrs &&
255 bch2_bkey_durability(c, k) >= m->op.opts.data_replicas) {
256 trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i), insert, "no rewrites found:");
261 * A replica that we just wrote might conflict with a replica
262 * that we want to keep, due to racing with another move:
264 restart_drop_conflicting_replicas:
265 extent_for_each_ptr(extent_i_to_s(new), ptr)
266 if ((ptr_c = bch2_bkey_has_device_c(bkey_i_to_s_c(insert), ptr->dev)) &&
268 bch2_bkey_drop_ptr_noerror(bkey_i_to_s(&new->k_i), ptr);
269 goto restart_drop_conflicting_replicas;
272 if (!bkey_val_u64s(&new->k)) {
273 trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i), insert, "new replicas conflicted:");
277 /* Now, drop pointers that conflict with what we just wrote: */
278 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry)
279 if ((ptr = bch2_bkey_has_device(bkey_i_to_s(insert), p.ptr.dev)))
280 bch2_bkey_drop_ptr_noerror(bkey_i_to_s(insert), ptr);
282 durability = bch2_bkey_durability(c, bkey_i_to_s_c(insert)) +
283 bch2_bkey_durability(c, bkey_i_to_s_c(&new->k_i));
285 /* Now, drop excess replicas: */
286 restart_drop_extra_replicas:
287 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs(bkey_i_to_s(insert)), p, entry) {
288 unsigned ptr_durability = bch2_extent_ptr_durability(c, &p);
291 durability - ptr_durability >= m->op.opts.data_replicas) {
292 durability -= ptr_durability;
293 bch2_bkey_drop_ptr_noerror(bkey_i_to_s(insert), &entry->ptr);
295 * Currently, we're dropping unneeded replicas
296 * instead of marking them as cached, since
297 * cached data in stripe buckets prevents them
299 bch2_extent_ptr_set_cached(bkey_i_to_s(insert), &entry->ptr);
301 goto restart_drop_extra_replicas;
305 /* Finally, add the pointers we just wrote: */
306 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry)
307 bch2_extent_ptr_decoded_append(insert, &p);
309 bch2_bkey_narrow_crcs(insert, (struct bch_extent_crc_unpacked) { 0 });
310 bch2_extent_normalize(c, bkey_i_to_s(insert));
312 ret = bch2_sum_sector_overwrites(trans, &iter, insert,
313 &should_check_enospc,
315 &disk_sectors_delta);
319 if (disk_sectors_delta > (s64) op->res.sectors) {
320 ret = bch2_disk_reservation_add(c, &op->res,
321 disk_sectors_delta - op->res.sectors,
323 ? BCH_DISK_RESERVATION_NOFAIL : 0);
328 next_pos = insert->k.p;
330 if (!bkey_eq(bkey_start_pos(&insert->k), bkey_start_pos(k.k))) {
331 ret = insert_snapshot_whiteouts(trans, m->btree_id, k.k->p,
332 bkey_start_pos(&insert->k));
337 if (!bkey_eq(insert->k.p, k.k->p)) {
338 ret = insert_snapshot_whiteouts(trans, m->btree_id,
339 k.k->p, insert->k.p);
344 ret = bch2_trans_update(trans, &iter, insert,
345 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
346 bch2_trans_commit(trans, &op->res,
348 BTREE_INSERT_NOCHECK_RW|
350 m->data_opts.btree_insert_flags);
352 bch2_btree_iter_set_pos(&iter, next_pos);
354 this_cpu_add(c->counters[BCH_COUNTER_move_extent_finish], new->k.size);
355 trace_move_extent_finish2(c, bkey_i_to_s_c(&new->k_i));
358 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
363 while (bkey_ge(iter.pos, bch2_keylist_front(keys)->k.p)) {
364 bch2_keylist_pop_front(keys);
365 if (bch2_keylist_empty(keys))
370 if (m->ctxt && m->ctxt->stats) {
371 BUG_ON(k.k->p.offset <= iter.pos.offset);
372 atomic64_inc(&m->ctxt->stats->keys_raced);
373 atomic64_add(k.k->p.offset - iter.pos.offset,
374 &m->ctxt->stats->sectors_raced);
377 this_cpu_add(c->counters[BCH_COUNTER_move_extent_fail], new->k.size);
379 bch2_btree_iter_advance(&iter);
383 bch2_trans_iter_exit(trans, &iter);
384 bch2_bkey_buf_exit(&_insert, c);
385 bch2_bkey_buf_exit(&_new, c);
386 BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart));
390 int bch2_data_update_index_update(struct bch_write_op *op)
392 return bch2_trans_run(op->c, __bch2_data_update_index_update(&trans, op));
395 void bch2_data_update_read_done(struct data_update *m,
396 struct bch_extent_crc_unpacked crc)
398 /* write bio must own pages: */
399 BUG_ON(!m->op.wbio.bio.bi_vcnt);
402 m->op.wbio.bio.bi_iter.bi_size = crc.compressed_size << 9;
404 closure_call(&m->op.cl, bch2_write, NULL, NULL);
407 void bch2_data_update_exit(struct data_update *update)
409 struct bch_fs *c = update->op.c;
410 struct bkey_ptrs_c ptrs =
411 bch2_bkey_ptrs_c(bkey_i_to_s_c(update->k.k));
412 const struct bch_extent_ptr *ptr;
414 bkey_for_each_ptr(ptrs, ptr) {
415 if (c->opts.nocow_enabled)
416 bch2_bucket_nocow_unlock(&c->nocow_locks,
417 PTR_BUCKET_POS(c, ptr), 0);
418 percpu_ref_put(&bch_dev_bkey_exists(c, ptr->dev)->ref);
421 bch2_bkey_buf_exit(&update->k, c);
422 bch2_disk_reservation_put(c, &update->op.res);
423 bch2_bio_free_pages_pool(c, &update->op.wbio.bio);
426 void bch2_update_unwritten_extent(struct btree_trans *trans,
427 struct data_update *update)
429 struct bch_fs *c = update->op.c;
430 struct bio *bio = &update->op.wbio.bio;
431 struct bkey_i_extent *e;
432 struct write_point *wp;
433 struct bch_extent_ptr *ptr;
435 struct btree_iter iter;
439 closure_init_stack(&cl);
440 bch2_keylist_init(&update->op.insert_keys, update->op.inline_keys);
442 while (bio_sectors(bio)) {
443 unsigned sectors = bio_sectors(bio);
445 bch2_trans_iter_init(trans, &iter, update->btree_id, update->op.pos,
447 ret = lockrestart_do(trans, ({
448 k = bch2_btree_iter_peek_slot(&iter);
451 bch2_trans_iter_exit(trans, &iter);
453 if (ret || !bch2_extents_match(k, bkey_i_to_s_c(update->k.k)))
456 e = bkey_extent_init(update->op.insert_keys.top);
457 e->k.p = update->op.pos;
459 ret = bch2_alloc_sectors_start_trans(trans,
462 update->op.write_point,
463 &update->op.devs_have,
464 update->op.nr_replicas,
465 update->op.nr_replicas,
466 update->op.alloc_reserve,
468 if (bch2_err_matches(ret, BCH_ERR_operation_blocked)) {
469 bch2_trans_unlock(trans);
477 sectors = min(sectors, wp->sectors_free);
479 bch2_key_resize(&e->k, sectors);
481 bch2_open_bucket_get(c, wp, &update->op.open_buckets);
482 bch2_alloc_sectors_append_ptrs(c, wp, &e->k_i, sectors, false);
483 bch2_alloc_sectors_done(c, wp);
485 bio_advance(bio, sectors << 9);
486 update->op.pos.offset += sectors;
488 extent_for_each_ptr(extent_i_to_s(e), ptr)
489 ptr->unwritten = true;
490 bch2_keylist_push(&update->op.insert_keys);
492 ret = __bch2_data_update_index_update(trans, &update->op);
494 bch2_open_buckets_put(c, &update->op.open_buckets);
500 if ((atomic_read(&cl.remaining) & CLOSURE_REMAINING_MASK) != 1) {
501 bch2_trans_unlock(trans);
506 int bch2_data_update_init(struct btree_trans *trans,
507 struct moving_context *ctxt,
508 struct data_update *m,
509 struct write_point_specifier wp,
510 struct bch_io_opts io_opts,
511 struct data_update_opts data_opts,
512 enum btree_id btree_id,
515 struct bch_fs *c = trans->c;
516 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
517 const union bch_extent_entry *entry;
518 struct extent_ptr_decoded p;
519 const struct bch_extent_ptr *ptr;
520 unsigned i, reserve_sectors = k.k->size * data_opts.extra_replicas;
521 unsigned ptrs_locked = 0;
524 bch2_bkey_buf_init(&m->k);
525 bch2_bkey_buf_reassemble(&m->k, c, k);
526 m->btree_id = btree_id;
527 m->data_opts = data_opts;
529 bch2_write_op_init(&m->op, c, io_opts);
530 m->op.pos = bkey_start_pos(k.k);
531 m->op.version = k.k->version;
532 m->op.target = data_opts.target;
533 m->op.write_point = wp;
534 m->op.nr_replicas = 0;
535 m->op.flags |= BCH_WRITE_PAGES_STABLE|
536 BCH_WRITE_PAGES_OWNED|
537 BCH_WRITE_DATA_ENCODED|
539 m->data_opts.write_flags;
540 m->op.compression_type =
541 bch2_compression_opt_to_type[io_opts.background_compression ?:
542 io_opts.compression];
543 if (m->data_opts.btree_insert_flags & BTREE_INSERT_USE_RESERVE)
544 m->op.alloc_reserve = RESERVE_movinggc;
546 bkey_for_each_ptr(ptrs, ptr)
547 percpu_ref_get(&bch_dev_bkey_exists(c, ptr->dev)->ref);
550 bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
553 if (((1U << i) & m->data_opts.rewrite_ptrs)) {
554 BUG_ON(p.ptr.cached);
556 if (crc_is_compressed(p.crc))
557 reserve_sectors += k.k->size;
559 m->op.nr_replicas += bch2_extent_ptr_durability(c, &p);
560 } else if (!p.ptr.cached) {
561 bch2_dev_list_add_dev(&m->op.devs_have, p.ptr.dev);
565 * op->csum_type is normally initialized from the fs/file's
566 * current options - but if an extent is encrypted, we require
567 * that it stays encrypted:
569 if (bch2_csum_type_is_encryption(p.crc.csum_type)) {
570 m->op.nonce = p.crc.nonce + p.crc.offset;
571 m->op.csum_type = p.crc.csum_type;
574 if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible)
575 m->op.incompressible = true;
577 if (c->opts.nocow_enabled) {
579 move_ctxt_wait_event(ctxt, trans,
580 (locked = bch2_bucket_nocow_trylock(&c->nocow_locks,
581 PTR_BUCKET_POS(c, &p.ptr), 0)) ||
582 !atomic_read(&ctxt->read_sectors));
585 bch2_bucket_nocow_lock(&c->nocow_locks,
586 PTR_BUCKET_POS(c, &p.ptr), 0);
588 if (!bch2_bucket_nocow_trylock(&c->nocow_locks,
589 PTR_BUCKET_POS(c, &p.ptr), 0)) {
590 ret = -BCH_ERR_nocow_lock_blocked;
594 ptrs_locked |= (1U << i);
600 if (reserve_sectors) {
601 ret = bch2_disk_reservation_add(c, &m->op.res, reserve_sectors,
602 m->data_opts.extra_replicas
604 : BCH_DISK_RESERVATION_NOFAIL);
609 m->op.nr_replicas += m->data_opts.extra_replicas;
610 m->op.nr_replicas_required = m->op.nr_replicas;
612 BUG_ON(!m->op.nr_replicas);
614 /* Special handling required: */
615 if (bkey_extent_is_unwritten(k))
616 return -BCH_ERR_unwritten_extent_update;
620 bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
621 if ((1U << i) & ptrs_locked)
622 bch2_bucket_nocow_unlock(&c->nocow_locks,
623 PTR_BUCKET_POS(c, &p.ptr), 0);
624 percpu_ref_put(&bch_dev_bkey_exists(c, p.ptr.dev)->ref);
628 bch2_bkey_buf_exit(&m->k, c);
629 bch2_bio_free_pages_pool(c, &m->op.wbio.bio);
633 void bch2_data_update_opts_normalize(struct bkey_s_c k, struct data_update_opts *opts)
635 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
636 const struct bch_extent_ptr *ptr;
639 bkey_for_each_ptr(ptrs, ptr) {
640 if ((opts->rewrite_ptrs & (1U << i)) && ptr->cached) {
641 opts->kill_ptrs |= 1U << i;
642 opts->rewrite_ptrs ^= 1U << i;