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 "rebalance.h"
17 #include "subvolume.h"
20 static void trace_move_extent_finish2(struct bch_fs *c, struct bkey_s_c k)
22 if (trace_move_extent_finish_enabled()) {
23 struct printbuf buf = PRINTBUF;
25 bch2_bkey_val_to_text(&buf, c, k);
26 trace_move_extent_finish(c, buf.buf);
31 static void trace_move_extent_fail2(struct data_update *m,
33 struct bkey_s_c wrote,
34 struct bkey_i *insert,
37 struct bch_fs *c = m->op.c;
38 struct bkey_s_c old = bkey_i_to_s_c(m->k.k);
39 const union bch_extent_entry *entry;
40 struct bch_extent_ptr *ptr;
41 struct extent_ptr_decoded p;
42 struct printbuf buf = PRINTBUF;
43 unsigned i, rewrites_found = 0;
45 if (!trace_move_extent_fail_enabled())
52 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry) {
53 if (((1U << i) & m->data_opts.rewrite_ptrs) &&
54 (ptr = bch2_extent_has_ptr(old, p, bkey_i_to_s(insert))) &&
56 rewrites_found |= 1U << i;
61 prt_printf(&buf, "\nrewrite ptrs: %u%u%u%u",
62 (m->data_opts.rewrite_ptrs & (1 << 0)) != 0,
63 (m->data_opts.rewrite_ptrs & (1 << 1)) != 0,
64 (m->data_opts.rewrite_ptrs & (1 << 2)) != 0,
65 (m->data_opts.rewrite_ptrs & (1 << 3)) != 0);
67 prt_printf(&buf, "\nrewrites found: %u%u%u%u",
68 (rewrites_found & (1 << 0)) != 0,
69 (rewrites_found & (1 << 1)) != 0,
70 (rewrites_found & (1 << 2)) != 0,
71 (rewrites_found & (1 << 3)) != 0);
73 prt_str(&buf, "\nold: ");
74 bch2_bkey_val_to_text(&buf, c, old);
76 prt_str(&buf, "\nnew: ");
77 bch2_bkey_val_to_text(&buf, c, new);
79 prt_str(&buf, "\nwrote: ");
80 bch2_bkey_val_to_text(&buf, c, wrote);
83 prt_str(&buf, "\ninsert: ");
84 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
87 trace_move_extent_fail(c, buf.buf);
91 static int __bch2_data_update_index_update(struct btree_trans *trans,
92 struct bch_write_op *op)
94 struct bch_fs *c = op->c;
95 struct btree_iter iter;
96 struct data_update *m =
97 container_of(op, struct data_update, op);
98 struct keylist *keys = &op->insert_keys;
99 struct bkey_buf _new, _insert;
102 bch2_bkey_buf_init(&_new);
103 bch2_bkey_buf_init(&_insert);
104 bch2_bkey_buf_realloc(&_insert, c, U8_MAX);
106 bch2_trans_iter_init(trans, &iter, m->btree_id,
107 bkey_start_pos(&bch2_keylist_front(keys)->k),
108 BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
112 struct bkey_s_c old = bkey_i_to_s_c(m->k.k);
113 struct bkey_i *insert = NULL;
114 struct bkey_i_extent *new;
115 const union bch_extent_entry *entry_c;
116 union bch_extent_entry *entry;
117 struct extent_ptr_decoded p;
118 struct bch_extent_ptr *ptr;
119 const struct bch_extent_ptr *ptr_c;
120 struct bpos next_pos;
121 bool should_check_enospc;
122 s64 i_sectors_delta = 0, disk_sectors_delta = 0;
123 unsigned rewrites_found = 0, durability, i;
125 bch2_trans_begin(trans);
127 k = bch2_btree_iter_peek_slot(&iter);
132 new = bkey_i_to_extent(bch2_keylist_front(keys));
134 if (!bch2_extents_match(k, old)) {
135 trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i),
140 bkey_reassemble(_insert.k, k);
143 bch2_bkey_buf_copy(&_new, c, bch2_keylist_front(keys));
144 new = bkey_i_to_extent(_new.k);
145 bch2_cut_front(iter.pos, &new->k_i);
147 bch2_cut_front(iter.pos, insert);
148 bch2_cut_back(new->k.p, insert);
149 bch2_cut_back(insert->k.p, &new->k_i);
152 * @old: extent that we read from
153 * @insert: key that we're going to update, initialized from
154 * extent currently in btree - same as @old unless we raced with
156 * @new: extent with new pointers that we'll be adding to @insert
158 * Fist, drop rewrite_ptrs from @new:
161 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry_c) {
162 if (((1U << i) & m->data_opts.rewrite_ptrs) &&
163 (ptr = bch2_extent_has_ptr(old, p, bkey_i_to_s(insert))) &&
165 bch2_extent_ptr_set_cached(bkey_i_to_s(insert), ptr);
166 rewrites_found |= 1U << i;
171 if (m->data_opts.rewrite_ptrs &&
173 bch2_bkey_durability(c, k) >= m->op.opts.data_replicas) {
174 trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i), insert, "no rewrites found:");
179 * A replica that we just wrote might conflict with a replica
180 * that we want to keep, due to racing with another move:
182 restart_drop_conflicting_replicas:
183 extent_for_each_ptr(extent_i_to_s(new), ptr)
184 if ((ptr_c = bch2_bkey_has_device_c(bkey_i_to_s_c(insert), ptr->dev)) &&
186 bch2_bkey_drop_ptr_noerror(bkey_i_to_s(&new->k_i), ptr);
187 goto restart_drop_conflicting_replicas;
190 if (!bkey_val_u64s(&new->k)) {
191 trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i), insert, "new replicas conflicted:");
195 /* Now, drop pointers that conflict with what we just wrote: */
196 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry)
197 if ((ptr = bch2_bkey_has_device(bkey_i_to_s(insert), p.ptr.dev)))
198 bch2_bkey_drop_ptr_noerror(bkey_i_to_s(insert), ptr);
200 durability = bch2_bkey_durability(c, bkey_i_to_s_c(insert)) +
201 bch2_bkey_durability(c, bkey_i_to_s_c(&new->k_i));
203 /* Now, drop excess replicas: */
204 restart_drop_extra_replicas:
205 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs(bkey_i_to_s(insert)), p, entry) {
206 unsigned ptr_durability = bch2_extent_ptr_durability(c, &p);
209 durability - ptr_durability >= m->op.opts.data_replicas) {
210 durability -= ptr_durability;
212 bch2_extent_ptr_set_cached(bkey_i_to_s(insert), &entry->ptr);
213 goto restart_drop_extra_replicas;
217 /* Finally, add the pointers we just wrote: */
218 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry)
219 bch2_extent_ptr_decoded_append(insert, &p);
221 bch2_bkey_narrow_crcs(insert, (struct bch_extent_crc_unpacked) { 0 });
222 bch2_extent_normalize(c, bkey_i_to_s(insert));
224 ret = bch2_sum_sector_overwrites(trans, &iter, insert,
225 &should_check_enospc,
227 &disk_sectors_delta);
231 if (disk_sectors_delta > (s64) op->res.sectors) {
232 ret = bch2_disk_reservation_add(c, &op->res,
233 disk_sectors_delta - op->res.sectors,
235 ? BCH_DISK_RESERVATION_NOFAIL : 0);
240 next_pos = insert->k.p;
243 * Check for nonce offset inconsistency:
244 * This is debug code - we've been seeing this bug rarely, and
245 * it's been hard to reproduce, so this should give us some more
246 * information when it does occur:
248 struct printbuf err = PRINTBUF;
249 int invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(insert), __btree_node_type(0, m->btree_id), 0, &err);
253 struct printbuf buf = PRINTBUF;
255 prt_str(&buf, "about to insert invalid key in data update path");
256 prt_str(&buf, "\nold: ");
257 bch2_bkey_val_to_text(&buf, c, old);
258 prt_str(&buf, "\nk: ");
259 bch2_bkey_val_to_text(&buf, c, k);
260 prt_str(&buf, "\nnew: ");
261 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
263 bch2_print_string_as_lines(KERN_ERR, buf.buf);
270 if (trace_data_update_enabled()) {
271 struct printbuf buf = PRINTBUF;
273 prt_str(&buf, "\nold: ");
274 bch2_bkey_val_to_text(&buf, c, old);
275 prt_str(&buf, "\nk: ");
276 bch2_bkey_val_to_text(&buf, c, k);
277 prt_str(&buf, "\nnew: ");
278 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
280 trace_data_update(c, buf.buf);
284 ret = bch2_insert_snapshot_whiteouts(trans, m->btree_id,
285 k.k->p, bkey_start_pos(&insert->k)) ?:
286 bch2_insert_snapshot_whiteouts(trans, m->btree_id,
287 k.k->p, insert->k.p) ?:
288 bch2_bkey_set_needs_rebalance(c, insert,
289 op->opts.background_target,
290 op->opts.background_compression) ?:
291 bch2_trans_update(trans, &iter, insert,
292 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
293 bch2_trans_commit(trans, &op->res,
295 BCH_TRANS_COMMIT_no_check_rw|
296 BCH_TRANS_COMMIT_no_enospc|
297 m->data_opts.btree_insert_flags);
299 bch2_btree_iter_set_pos(&iter, next_pos);
301 this_cpu_add(c->counters[BCH_COUNTER_move_extent_finish], new->k.size);
302 trace_move_extent_finish2(c, bkey_i_to_s_c(&new->k_i));
305 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
310 while (bkey_ge(iter.pos, bch2_keylist_front(keys)->k.p)) {
311 bch2_keylist_pop_front(keys);
312 if (bch2_keylist_empty(keys))
317 if (m->stats && m->stats) {
318 BUG_ON(k.k->p.offset <= iter.pos.offset);
319 atomic64_inc(&m->stats->keys_raced);
320 atomic64_add(k.k->p.offset - iter.pos.offset,
321 &m->stats->sectors_raced);
324 this_cpu_inc(c->counters[BCH_COUNTER_move_extent_fail]);
326 bch2_btree_iter_advance(&iter);
330 bch2_trans_iter_exit(trans, &iter);
331 bch2_bkey_buf_exit(&_insert, c);
332 bch2_bkey_buf_exit(&_new, c);
333 BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart));
337 int bch2_data_update_index_update(struct bch_write_op *op)
339 return bch2_trans_run(op->c, __bch2_data_update_index_update(trans, op));
342 void bch2_data_update_read_done(struct data_update *m,
343 struct bch_extent_crc_unpacked crc)
345 /* write bio must own pages: */
346 BUG_ON(!m->op.wbio.bio.bi_vcnt);
349 m->op.wbio.bio.bi_iter.bi_size = crc.compressed_size << 9;
351 closure_call(&m->op.cl, bch2_write, NULL, NULL);
354 void bch2_data_update_exit(struct data_update *update)
356 struct bch_fs *c = update->op.c;
357 struct bkey_ptrs_c ptrs =
358 bch2_bkey_ptrs_c(bkey_i_to_s_c(update->k.k));
359 const struct bch_extent_ptr *ptr;
361 bkey_for_each_ptr(ptrs, ptr) {
362 if (c->opts.nocow_enabled)
363 bch2_bucket_nocow_unlock(&c->nocow_locks,
364 PTR_BUCKET_POS(c, ptr), 0);
365 percpu_ref_put(&bch_dev_bkey_exists(c, ptr->dev)->ref);
368 bch2_bkey_buf_exit(&update->k, c);
369 bch2_disk_reservation_put(c, &update->op.res);
370 bch2_bio_free_pages_pool(c, &update->op.wbio.bio);
373 static void bch2_update_unwritten_extent(struct btree_trans *trans,
374 struct data_update *update)
376 struct bch_fs *c = update->op.c;
377 struct bio *bio = &update->op.wbio.bio;
378 struct bkey_i_extent *e;
379 struct write_point *wp;
380 struct bch_extent_ptr *ptr;
382 struct btree_iter iter;
386 closure_init_stack(&cl);
387 bch2_keylist_init(&update->op.insert_keys, update->op.inline_keys);
389 while (bio_sectors(bio)) {
390 unsigned sectors = bio_sectors(bio);
392 bch2_trans_iter_init(trans, &iter, update->btree_id, update->op.pos,
394 ret = lockrestart_do(trans, ({
395 k = bch2_btree_iter_peek_slot(&iter);
398 bch2_trans_iter_exit(trans, &iter);
400 if (ret || !bch2_extents_match(k, bkey_i_to_s_c(update->k.k)))
403 e = bkey_extent_init(update->op.insert_keys.top);
404 e->k.p = update->op.pos;
406 ret = bch2_alloc_sectors_start_trans(trans,
409 update->op.write_point,
410 &update->op.devs_have,
411 update->op.nr_replicas,
412 update->op.nr_replicas,
413 update->op.watermark,
415 if (bch2_err_matches(ret, BCH_ERR_operation_blocked)) {
416 bch2_trans_unlock(trans);
424 sectors = min(sectors, wp->sectors_free);
426 bch2_key_resize(&e->k, sectors);
428 bch2_open_bucket_get(c, wp, &update->op.open_buckets);
429 bch2_alloc_sectors_append_ptrs(c, wp, &e->k_i, sectors, false);
430 bch2_alloc_sectors_done(c, wp);
432 bio_advance(bio, sectors << 9);
433 update->op.pos.offset += sectors;
435 extent_for_each_ptr(extent_i_to_s(e), ptr)
436 ptr->unwritten = true;
437 bch2_keylist_push(&update->op.insert_keys);
439 ret = __bch2_data_update_index_update(trans, &update->op);
441 bch2_open_buckets_put(c, &update->op.open_buckets);
447 if (closure_nr_remaining(&cl) != 1) {
448 bch2_trans_unlock(trans);
453 int bch2_extent_drop_ptrs(struct btree_trans *trans,
454 struct btree_iter *iter,
456 struct data_update_opts data_opts)
458 struct bch_fs *c = trans->c;
462 n = bch2_bkey_make_mut_noupdate(trans, k);
463 ret = PTR_ERR_OR_ZERO(n);
467 while (data_opts.kill_ptrs) {
468 unsigned i = 0, drop = __fls(data_opts.kill_ptrs);
469 struct bch_extent_ptr *ptr;
471 bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, i++ == drop);
472 data_opts.kill_ptrs ^= 1U << drop;
476 * If the new extent no longer has any pointers, bch2_extent_normalize()
477 * will do the appropriate thing with it (turning it into a
478 * KEY_TYPE_error key, or just a discard if it was a cached extent)
480 bch2_extent_normalize(c, bkey_i_to_s(n));
483 * Since we're not inserting through an extent iterator
484 * (BTREE_ITER_ALL_SNAPSHOTS iterators aren't extent iterators),
485 * we aren't using the extent overwrite path to delete, we're
486 * just using the normal key deletion path:
488 if (bkey_deleted(&n->k))
491 return bch2_trans_relock(trans) ?:
492 bch2_trans_update(trans, iter, n, BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
493 bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc);
496 int bch2_data_update_init(struct btree_trans *trans,
497 struct btree_iter *iter,
498 struct moving_context *ctxt,
499 struct data_update *m,
500 struct write_point_specifier wp,
501 struct bch_io_opts io_opts,
502 struct data_update_opts data_opts,
503 enum btree_id btree_id,
506 struct bch_fs *c = trans->c;
507 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
508 const union bch_extent_entry *entry;
509 struct extent_ptr_decoded p;
510 const struct bch_extent_ptr *ptr;
511 unsigned i, reserve_sectors = k.k->size * data_opts.extra_replicas;
512 unsigned ptrs_locked = 0;
515 bch2_bkey_buf_init(&m->k);
516 bch2_bkey_buf_reassemble(&m->k, c, k);
517 m->btree_id = btree_id;
518 m->data_opts = data_opts;
520 m->stats = ctxt ? ctxt->stats : NULL;
522 bch2_write_op_init(&m->op, c, io_opts);
523 m->op.pos = bkey_start_pos(k.k);
524 m->op.version = k.k->version;
525 m->op.target = data_opts.target;
526 m->op.write_point = wp;
527 m->op.nr_replicas = 0;
528 m->op.flags |= BCH_WRITE_PAGES_STABLE|
529 BCH_WRITE_PAGES_OWNED|
530 BCH_WRITE_DATA_ENCODED|
532 m->data_opts.write_flags;
533 m->op.compression_opt = io_opts.background_compression ?: io_opts.compression;
534 m->op.watermark = m->data_opts.btree_insert_flags & BCH_WATERMARK_MASK;
536 bkey_for_each_ptr(ptrs, ptr)
537 percpu_ref_get(&bch_dev_bkey_exists(c, ptr->dev)->ref);
539 unsigned durability_have = 0, durability_removing = 0;
542 bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
545 if (((1U << i) & m->data_opts.rewrite_ptrs)) {
546 BUG_ON(p.ptr.cached);
548 if (crc_is_compressed(p.crc))
549 reserve_sectors += k.k->size;
551 m->op.nr_replicas += bch2_extent_ptr_desired_durability(c, &p);
552 durability_removing += bch2_extent_ptr_desired_durability(c, &p);
553 } else if (!p.ptr.cached &&
554 !((1U << i) & m->data_opts.kill_ptrs)) {
555 bch2_dev_list_add_dev(&m->op.devs_have, p.ptr.dev);
556 durability_have += bch2_extent_ptr_durability(c, &p);
560 * op->csum_type is normally initialized from the fs/file's
561 * current options - but if an extent is encrypted, we require
562 * that it stays encrypted:
564 if (bch2_csum_type_is_encryption(p.crc.csum_type)) {
565 m->op.nonce = p.crc.nonce + p.crc.offset;
566 m->op.csum_type = p.crc.csum_type;
569 if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible)
570 m->op.incompressible = true;
572 if (c->opts.nocow_enabled) {
574 move_ctxt_wait_event(ctxt,
575 (locked = bch2_bucket_nocow_trylock(&c->nocow_locks,
576 PTR_BUCKET_POS(c, &p.ptr), 0)) ||
577 !atomic_read(&ctxt->read_sectors));
580 bch2_bucket_nocow_lock(&c->nocow_locks,
581 PTR_BUCKET_POS(c, &p.ptr), 0);
583 if (!bch2_bucket_nocow_trylock(&c->nocow_locks,
584 PTR_BUCKET_POS(c, &p.ptr), 0)) {
585 ret = -BCH_ERR_nocow_lock_blocked;
589 ptrs_locked |= (1U << i);
596 * If current extent durability is less than io_opts.data_replicas,
597 * we're not trying to rereplicate the extent up to data_replicas here -
598 * unless extra_replicas was specified
600 * Increasing replication is an explicit operation triggered by
601 * rereplicate, currently, so that users don't get an unexpected -ENOSPC
603 if (durability_have >= io_opts.data_replicas) {
604 m->data_opts.kill_ptrs |= m->data_opts.rewrite_ptrs;
605 m->data_opts.rewrite_ptrs = 0;
606 /* if iter == NULL, it's just a promote */
608 ret = bch2_extent_drop_ptrs(trans, iter, k, data_opts);
612 m->op.nr_replicas = min(durability_removing, io_opts.data_replicas - durability_have) +
613 m->data_opts.extra_replicas;
614 m->op.nr_replicas_required = m->op.nr_replicas;
616 BUG_ON(!m->op.nr_replicas);
618 if (reserve_sectors) {
619 ret = bch2_disk_reservation_add(c, &m->op.res, reserve_sectors,
620 m->data_opts.extra_replicas
622 : BCH_DISK_RESERVATION_NOFAIL);
627 if (bkey_extent_is_unwritten(k)) {
628 bch2_update_unwritten_extent(trans, m);
635 bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
636 if ((1U << i) & ptrs_locked)
637 bch2_bucket_nocow_unlock(&c->nocow_locks,
638 PTR_BUCKET_POS(c, &p.ptr), 0);
639 percpu_ref_put(&bch_dev_bkey_exists(c, p.ptr.dev)->ref);
643 bch2_bkey_buf_exit(&m->k, c);
644 bch2_bio_free_pages_pool(c, &m->op.wbio.bio);
647 bch2_data_update_exit(m);
648 return ret ?: -BCH_ERR_data_update_done;
651 void bch2_data_update_opts_normalize(struct bkey_s_c k, struct data_update_opts *opts)
653 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
654 const struct bch_extent_ptr *ptr;
657 bkey_for_each_ptr(ptrs, ptr) {
658 if ((opts->rewrite_ptrs & (1U << i)) && ptr->cached) {
659 opts->kill_ptrs |= 1U << i;
660 opts->rewrite_ptrs ^= 1U << i;