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"
18 #include "subvolume.h"
21 static void trace_move_extent_finish2(struct bch_fs *c, struct bkey_s_c k)
23 if (trace_move_extent_finish_enabled()) {
24 struct printbuf buf = PRINTBUF;
26 bch2_bkey_val_to_text(&buf, c, k);
27 trace_move_extent_finish(c, buf.buf);
32 static void trace_move_extent_fail2(struct data_update *m,
34 struct bkey_s_c wrote,
35 struct bkey_i *insert,
38 struct bch_fs *c = m->op.c;
39 struct bkey_s_c old = bkey_i_to_s_c(m->k.k);
40 const union bch_extent_entry *entry;
41 struct bch_extent_ptr *ptr;
42 struct extent_ptr_decoded p;
43 struct printbuf buf = PRINTBUF;
44 unsigned i, rewrites_found = 0;
46 if (!trace_move_extent_fail_enabled())
53 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry) {
54 if (((1U << i) & m->data_opts.rewrite_ptrs) &&
55 (ptr = bch2_extent_has_ptr(old, p, bkey_i_to_s(insert))) &&
57 rewrites_found |= 1U << i;
62 prt_printf(&buf, "\nrewrite ptrs: %u%u%u%u",
63 (m->data_opts.rewrite_ptrs & (1 << 0)) != 0,
64 (m->data_opts.rewrite_ptrs & (1 << 1)) != 0,
65 (m->data_opts.rewrite_ptrs & (1 << 2)) != 0,
66 (m->data_opts.rewrite_ptrs & (1 << 3)) != 0);
68 prt_printf(&buf, "\nrewrites found: %u%u%u%u",
69 (rewrites_found & (1 << 0)) != 0,
70 (rewrites_found & (1 << 1)) != 0,
71 (rewrites_found & (1 << 2)) != 0,
72 (rewrites_found & (1 << 3)) != 0);
74 prt_str(&buf, "\nold: ");
75 bch2_bkey_val_to_text(&buf, c, old);
77 prt_str(&buf, "\nnew: ");
78 bch2_bkey_val_to_text(&buf, c, new);
80 prt_str(&buf, "\nwrote: ");
81 bch2_bkey_val_to_text(&buf, c, wrote);
84 prt_str(&buf, "\ninsert: ");
85 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
88 trace_move_extent_fail(c, buf.buf);
92 static int __bch2_data_update_index_update(struct btree_trans *trans,
93 struct bch_write_op *op)
95 struct bch_fs *c = op->c;
96 struct btree_iter iter;
97 struct data_update *m =
98 container_of(op, struct data_update, op);
99 struct keylist *keys = &op->insert_keys;
100 struct bkey_buf _new, _insert;
103 bch2_bkey_buf_init(&_new);
104 bch2_bkey_buf_init(&_insert);
105 bch2_bkey_buf_realloc(&_insert, c, U8_MAX);
107 bch2_trans_iter_init(trans, &iter, m->btree_id,
108 bkey_start_pos(&bch2_keylist_front(keys)->k),
109 BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
113 struct bkey_s_c old = bkey_i_to_s_c(m->k.k);
114 struct bkey_i *insert = NULL;
115 struct bkey_i_extent *new;
116 const union bch_extent_entry *entry_c;
117 union bch_extent_entry *entry;
118 struct extent_ptr_decoded p;
119 struct bch_extent_ptr *ptr;
120 const struct bch_extent_ptr *ptr_c;
121 struct bpos next_pos;
122 bool should_check_enospc;
123 s64 i_sectors_delta = 0, disk_sectors_delta = 0;
124 unsigned rewrites_found = 0, durability, i;
126 bch2_trans_begin(trans);
128 k = bch2_btree_iter_peek_slot(&iter);
133 new = bkey_i_to_extent(bch2_keylist_front(keys));
135 if (!bch2_extents_match(k, old)) {
136 trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i),
141 bkey_reassemble(_insert.k, k);
144 bch2_bkey_buf_copy(&_new, c, bch2_keylist_front(keys));
145 new = bkey_i_to_extent(_new.k);
146 bch2_cut_front(iter.pos, &new->k_i);
148 bch2_cut_front(iter.pos, insert);
149 bch2_cut_back(new->k.p, insert);
150 bch2_cut_back(insert->k.p, &new->k_i);
153 * @old: extent that we read from
154 * @insert: key that we're going to update, initialized from
155 * extent currently in btree - same as @old unless we raced with
157 * @new: extent with new pointers that we'll be adding to @insert
159 * Fist, drop rewrite_ptrs from @new:
162 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry_c) {
163 if (((1U << i) & m->data_opts.rewrite_ptrs) &&
164 (ptr = bch2_extent_has_ptr(old, p, bkey_i_to_s(insert))) &&
166 bch2_extent_ptr_set_cached(bkey_i_to_s(insert), ptr);
167 rewrites_found |= 1U << i;
172 if (m->data_opts.rewrite_ptrs &&
174 bch2_bkey_durability(c, k) >= m->op.opts.data_replicas) {
175 trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i), insert, "no rewrites found:");
180 * A replica that we just wrote might conflict with a replica
181 * that we want to keep, due to racing with another move:
183 restart_drop_conflicting_replicas:
184 extent_for_each_ptr(extent_i_to_s(new), ptr)
185 if ((ptr_c = bch2_bkey_has_device_c(bkey_i_to_s_c(insert), ptr->dev)) &&
187 bch2_bkey_drop_ptr_noerror(bkey_i_to_s(&new->k_i), ptr);
188 goto restart_drop_conflicting_replicas;
191 if (!bkey_val_u64s(&new->k)) {
192 trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i), insert, "new replicas conflicted:");
196 /* Now, drop pointers that conflict with what we just wrote: */
197 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry)
198 if ((ptr = bch2_bkey_has_device(bkey_i_to_s(insert), p.ptr.dev)))
199 bch2_bkey_drop_ptr_noerror(bkey_i_to_s(insert), ptr);
201 durability = bch2_bkey_durability(c, bkey_i_to_s_c(insert)) +
202 bch2_bkey_durability(c, bkey_i_to_s_c(&new->k_i));
204 /* Now, drop excess replicas: */
205 restart_drop_extra_replicas:
206 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs(bkey_i_to_s(insert)), p, entry) {
207 unsigned ptr_durability = bch2_extent_ptr_durability(c, &p);
210 durability - ptr_durability >= m->op.opts.data_replicas) {
211 durability -= ptr_durability;
213 bch2_extent_ptr_set_cached(bkey_i_to_s(insert), &entry->ptr);
214 goto restart_drop_extra_replicas;
218 /* Finally, add the pointers we just wrote: */
219 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry)
220 bch2_extent_ptr_decoded_append(insert, &p);
222 bch2_bkey_narrow_crcs(insert, (struct bch_extent_crc_unpacked) { 0 });
223 bch2_extent_normalize(c, bkey_i_to_s(insert));
225 ret = bch2_sum_sector_overwrites(trans, &iter, insert,
226 &should_check_enospc,
228 &disk_sectors_delta);
232 if (disk_sectors_delta > (s64) op->res.sectors) {
233 ret = bch2_disk_reservation_add(c, &op->res,
234 disk_sectors_delta - op->res.sectors,
236 ? BCH_DISK_RESERVATION_NOFAIL : 0);
241 next_pos = insert->k.p;
244 * Check for nonce offset inconsistency:
245 * This is debug code - we've been seeing this bug rarely, and
246 * it's been hard to reproduce, so this should give us some more
247 * information when it does occur:
249 struct printbuf err = PRINTBUF;
250 int invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(insert), __btree_node_type(0, m->btree_id), 0, &err);
254 struct printbuf buf = PRINTBUF;
256 prt_str(&buf, "about to insert invalid key in data update path");
257 prt_str(&buf, "\nold: ");
258 bch2_bkey_val_to_text(&buf, c, old);
259 prt_str(&buf, "\nk: ");
260 bch2_bkey_val_to_text(&buf, c, k);
261 prt_str(&buf, "\nnew: ");
262 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
264 bch2_print_string_as_lines(KERN_ERR, buf.buf);
271 if (trace_data_update_enabled()) {
272 struct printbuf buf = PRINTBUF;
274 prt_str(&buf, "\nold: ");
275 bch2_bkey_val_to_text(&buf, c, old);
276 prt_str(&buf, "\nk: ");
277 bch2_bkey_val_to_text(&buf, c, k);
278 prt_str(&buf, "\nnew: ");
279 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
281 trace_data_update(c, buf.buf);
285 ret = bch2_insert_snapshot_whiteouts(trans, m->btree_id,
286 k.k->p, bkey_start_pos(&insert->k)) ?:
287 bch2_insert_snapshot_whiteouts(trans, m->btree_id,
288 k.k->p, insert->k.p) ?:
289 bch2_bkey_set_needs_rebalance(c, insert, &op->opts) ?:
290 bch2_trans_update(trans, &iter, insert,
291 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
292 bch2_trans_commit(trans, &op->res,
294 BCH_TRANS_COMMIT_no_check_rw|
295 BCH_TRANS_COMMIT_no_enospc|
296 m->data_opts.btree_insert_flags);
298 bch2_btree_iter_set_pos(&iter, next_pos);
300 this_cpu_add(c->counters[BCH_COUNTER_move_extent_finish], new->k.size);
301 trace_move_extent_finish2(c, bkey_i_to_s_c(&new->k_i));
304 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
309 while (bkey_ge(iter.pos, bch2_keylist_front(keys)->k.p)) {
310 bch2_keylist_pop_front(keys);
311 if (bch2_keylist_empty(keys))
317 BUG_ON(k.k->p.offset <= iter.pos.offset);
318 atomic64_inc(&m->stats->keys_raced);
319 atomic64_add(k.k->p.offset - iter.pos.offset,
320 &m->stats->sectors_raced);
323 count_event(c, move_extent_fail);
325 bch2_btree_iter_advance(&iter);
329 bch2_trans_iter_exit(trans, &iter);
330 bch2_bkey_buf_exit(&_insert, c);
331 bch2_bkey_buf_exit(&_new, c);
332 BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart));
336 int bch2_data_update_index_update(struct bch_write_op *op)
338 return bch2_trans_run(op->c, __bch2_data_update_index_update(trans, op));
341 void bch2_data_update_read_done(struct data_update *m,
342 struct bch_extent_crc_unpacked crc)
344 /* write bio must own pages: */
345 BUG_ON(!m->op.wbio.bio.bi_vcnt);
348 m->op.wbio.bio.bi_iter.bi_size = crc.compressed_size << 9;
350 closure_call(&m->op.cl, bch2_write, NULL, NULL);
353 void bch2_data_update_exit(struct data_update *update)
355 struct bch_fs *c = update->op.c;
356 struct bkey_ptrs_c ptrs =
357 bch2_bkey_ptrs_c(bkey_i_to_s_c(update->k.k));
359 bkey_for_each_ptr(ptrs, ptr) {
360 if (c->opts.nocow_enabled)
361 bch2_bucket_nocow_unlock(&c->nocow_locks,
362 PTR_BUCKET_POS(c, ptr), 0);
363 percpu_ref_put(&bch_dev_bkey_exists(c, ptr->dev)->ref);
366 bch2_bkey_buf_exit(&update->k, c);
367 bch2_disk_reservation_put(c, &update->op.res);
368 bch2_bio_free_pages_pool(c, &update->op.wbio.bio);
371 static void bch2_update_unwritten_extent(struct btree_trans *trans,
372 struct data_update *update)
374 struct bch_fs *c = update->op.c;
375 struct bio *bio = &update->op.wbio.bio;
376 struct bkey_i_extent *e;
377 struct write_point *wp;
379 struct btree_iter iter;
383 closure_init_stack(&cl);
384 bch2_keylist_init(&update->op.insert_keys, update->op.inline_keys);
386 while (bio_sectors(bio)) {
387 unsigned sectors = bio_sectors(bio);
389 bch2_trans_iter_init(trans, &iter, update->btree_id, update->op.pos,
391 ret = lockrestart_do(trans, ({
392 k = bch2_btree_iter_peek_slot(&iter);
395 bch2_trans_iter_exit(trans, &iter);
397 if (ret || !bch2_extents_match(k, bkey_i_to_s_c(update->k.k)))
400 e = bkey_extent_init(update->op.insert_keys.top);
401 e->k.p = update->op.pos;
403 ret = bch2_alloc_sectors_start_trans(trans,
406 update->op.write_point,
407 &update->op.devs_have,
408 update->op.nr_replicas,
409 update->op.nr_replicas,
410 update->op.watermark,
412 if (bch2_err_matches(ret, BCH_ERR_operation_blocked)) {
413 bch2_trans_unlock(trans);
418 bch_err_fn_ratelimited(c, ret);
423 sectors = min(sectors, wp->sectors_free);
425 bch2_key_resize(&e->k, sectors);
427 bch2_open_bucket_get(c, wp, &update->op.open_buckets);
428 bch2_alloc_sectors_append_ptrs(c, wp, &e->k_i, sectors, false);
429 bch2_alloc_sectors_done(c, wp);
431 bio_advance(bio, sectors << 9);
432 update->op.pos.offset += sectors;
434 extent_for_each_ptr(extent_i_to_s(e), ptr)
435 ptr->unwritten = true;
436 bch2_keylist_push(&update->op.insert_keys);
438 ret = __bch2_data_update_index_update(trans, &update->op);
440 bch2_open_buckets_put(c, &update->op.open_buckets);
446 if (closure_nr_remaining(&cl) != 1) {
447 bch2_trans_unlock(trans);
452 int bch2_extent_drop_ptrs(struct btree_trans *trans,
453 struct btree_iter *iter,
455 struct data_update_opts data_opts)
457 struct bch_fs *c = trans->c;
461 n = bch2_bkey_make_mut_noupdate(trans, k);
462 ret = PTR_ERR_OR_ZERO(n);
466 while (data_opts.kill_ptrs) {
467 unsigned i = 0, drop = __fls(data_opts.kill_ptrs);
468 struct bch_extent_ptr *ptr;
470 bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, i++ == drop);
471 data_opts.kill_ptrs ^= 1U << drop;
475 * If the new extent no longer has any pointers, bch2_extent_normalize()
476 * will do the appropriate thing with it (turning it into a
477 * KEY_TYPE_error key, or just a discard if it was a cached extent)
479 bch2_extent_normalize(c, bkey_i_to_s(n));
482 * Since we're not inserting through an extent iterator
483 * (BTREE_ITER_ALL_SNAPSHOTS iterators aren't extent iterators),
484 * we aren't using the extent overwrite path to delete, we're
485 * just using the normal key deletion path:
487 if (bkey_deleted(&n->k) && !(iter->flags & BTREE_ITER_IS_EXTENTS))
490 return bch2_trans_relock(trans) ?:
491 bch2_trans_update(trans, iter, n, BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
492 bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc);
495 int bch2_data_update_init(struct btree_trans *trans,
496 struct btree_iter *iter,
497 struct moving_context *ctxt,
498 struct data_update *m,
499 struct write_point_specifier wp,
500 struct bch_io_opts io_opts,
501 struct data_update_opts data_opts,
502 enum btree_id btree_id,
505 struct bch_fs *c = trans->c;
506 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
507 const union bch_extent_entry *entry;
508 struct extent_ptr_decoded p;
509 unsigned i, reserve_sectors = k.k->size * data_opts.extra_replicas;
510 unsigned ptrs_locked = 0;
514 * fs is corrupt we have a key for a snapshot node that doesn't exist,
515 * and we have to check for this because we go rw before repairing the
516 * snapshots table - just skip it, we can move it later.
518 if (unlikely(k.k->p.snapshot && !bch2_snapshot_equiv(c, k.k->p.snapshot)))
519 return -BCH_ERR_data_update_done;
521 bch2_bkey_buf_init(&m->k);
522 bch2_bkey_buf_reassemble(&m->k, c, k);
523 m->btree_id = btree_id;
524 m->data_opts = data_opts;
526 m->stats = ctxt ? ctxt->stats : NULL;
528 bch2_write_op_init(&m->op, c, io_opts);
529 m->op.pos = bkey_start_pos(k.k);
530 m->op.version = k.k->version;
531 m->op.target = data_opts.target;
532 m->op.write_point = wp;
533 m->op.nr_replicas = 0;
534 m->op.flags |= BCH_WRITE_PAGES_STABLE|
535 BCH_WRITE_PAGES_OWNED|
536 BCH_WRITE_DATA_ENCODED|
538 m->data_opts.write_flags;
539 m->op.compression_opt = background_compression(io_opts);
540 m->op.watermark = m->data_opts.btree_insert_flags & BCH_WATERMARK_MASK;
542 bkey_for_each_ptr(ptrs, ptr)
543 percpu_ref_get(&bch_dev_bkey_exists(c, ptr->dev)->ref);
545 unsigned durability_have = 0, durability_removing = 0;
548 bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
551 if (((1U << i) & m->data_opts.rewrite_ptrs)) {
552 BUG_ON(p.ptr.cached);
554 if (crc_is_compressed(p.crc))
555 reserve_sectors += k.k->size;
557 m->op.nr_replicas += bch2_extent_ptr_desired_durability(c, &p);
558 durability_removing += bch2_extent_ptr_desired_durability(c, &p);
559 } else if (!p.ptr.cached &&
560 !((1U << i) & m->data_opts.kill_ptrs)) {
561 bch2_dev_list_add_dev(&m->op.devs_have, p.ptr.dev);
562 durability_have += bch2_extent_ptr_durability(c, &p);
566 * op->csum_type is normally initialized from the fs/file's
567 * current options - but if an extent is encrypted, we require
568 * that it stays encrypted:
570 if (bch2_csum_type_is_encryption(p.crc.csum_type)) {
571 m->op.nonce = p.crc.nonce + p.crc.offset;
572 m->op.csum_type = p.crc.csum_type;
575 if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible)
576 m->op.incompressible = true;
578 if (c->opts.nocow_enabled) {
580 move_ctxt_wait_event(ctxt,
581 (locked = bch2_bucket_nocow_trylock(&c->nocow_locks,
582 PTR_BUCKET_POS(c, &p.ptr), 0)) ||
583 list_empty(&ctxt->ios));
586 bch2_bucket_nocow_lock(&c->nocow_locks,
587 PTR_BUCKET_POS(c, &p.ptr), 0);
589 if (!bch2_bucket_nocow_trylock(&c->nocow_locks,
590 PTR_BUCKET_POS(c, &p.ptr), 0)) {
591 ret = -BCH_ERR_nocow_lock_blocked;
595 ptrs_locked |= (1U << i);
601 unsigned durability_required = max(0, (int) (io_opts.data_replicas - durability_have));
604 * If current extent durability is less than io_opts.data_replicas,
605 * we're not trying to rereplicate the extent up to data_replicas here -
606 * unless extra_replicas was specified
608 * Increasing replication is an explicit operation triggered by
609 * rereplicate, currently, so that users don't get an unexpected -ENOSPC
611 if (!(m->data_opts.write_flags & BCH_WRITE_CACHED) &&
612 !durability_required) {
613 m->data_opts.kill_ptrs |= m->data_opts.rewrite_ptrs;
614 m->data_opts.rewrite_ptrs = 0;
615 /* if iter == NULL, it's just a promote */
617 ret = bch2_extent_drop_ptrs(trans, iter, k, m->data_opts);
621 m->op.nr_replicas = min(durability_removing, durability_required) +
622 m->data_opts.extra_replicas;
625 * If device(s) were set to durability=0 after data was written to them
626 * we can end up with a duribilty=0 extent, and the normal algorithm
627 * that tries not to increase durability doesn't work:
629 if (!(durability_have + durability_removing))
630 m->op.nr_replicas = max((unsigned) m->op.nr_replicas, 1);
632 m->op.nr_replicas_required = m->op.nr_replicas;
634 if (reserve_sectors) {
635 ret = bch2_disk_reservation_add(c, &m->op.res, reserve_sectors,
636 m->data_opts.extra_replicas
638 : BCH_DISK_RESERVATION_NOFAIL);
643 if (bkey_extent_is_unwritten(k)) {
644 bch2_update_unwritten_extent(trans, m);
651 bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
652 if ((1U << i) & ptrs_locked)
653 bch2_bucket_nocow_unlock(&c->nocow_locks,
654 PTR_BUCKET_POS(c, &p.ptr), 0);
655 percpu_ref_put(&bch_dev_bkey_exists(c, p.ptr.dev)->ref);
659 bch2_bkey_buf_exit(&m->k, c);
660 bch2_bio_free_pages_pool(c, &m->op.wbio.bio);
663 bch2_data_update_exit(m);
664 return ret ?: -BCH_ERR_data_update_done;
667 void bch2_data_update_opts_normalize(struct bkey_s_c k, struct data_update_opts *opts)
669 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
672 bkey_for_each_ptr(ptrs, ptr) {
673 if ((opts->rewrite_ptrs & (1U << i)) && ptr->cached) {
674 opts->kill_ptrs |= 1U << i;
675 opts->rewrite_ptrs ^= 1U << i;