1 // SPDX-License-Identifier: GPL-2.0
6 #include "btree_cache.h"
7 #include "btree_journal_iter.h"
8 #include "journal_io.h"
10 #include <linux/sort.h>
13 * For managing keys we read from the journal: until journal replay works normal
14 * btree lookups need to be able to find and return keys from the journal where
15 * they overwrite what's in the btree, so we have a special iterator and
16 * operations for the regular btree iter code to use:
19 static int __journal_key_cmp(enum btree_id l_btree_id,
22 const struct journal_key *r)
24 return (cmp_int(l_btree_id, r->btree_id) ?:
25 cmp_int(l_level, r->level) ?:
26 bpos_cmp(l_pos, r->k->k.p));
29 static int journal_key_cmp(const struct journal_key *l, const struct journal_key *r)
31 return __journal_key_cmp(l->btree_id, l->level, l->k->k.p, r);
34 static inline size_t idx_to_pos(struct journal_keys *keys, size_t idx)
36 size_t gap_size = keys->size - keys->nr;
43 static inline struct journal_key *idx_to_key(struct journal_keys *keys, size_t idx)
45 return keys->d + idx_to_pos(keys, idx);
48 static size_t __bch2_journal_key_search(struct journal_keys *keys,
49 enum btree_id id, unsigned level,
52 size_t l = 0, r = keys->nr, m;
55 m = l + ((r - l) >> 1);
56 if (__journal_key_cmp(id, level, pos, idx_to_key(keys, m)) > 0)
62 BUG_ON(l < keys->nr &&
63 __journal_key_cmp(id, level, pos, idx_to_key(keys, l)) > 0);
66 __journal_key_cmp(id, level, pos, idx_to_key(keys, l - 1)) <= 0);
71 static size_t bch2_journal_key_search(struct journal_keys *keys,
72 enum btree_id id, unsigned level,
75 return idx_to_pos(keys, __bch2_journal_key_search(keys, id, level, pos));
78 /* Returns first non-overwritten key >= search key: */
79 struct bkey_i *bch2_journal_keys_peek_upto(struct bch_fs *c, enum btree_id btree_id,
80 unsigned level, struct bpos pos,
81 struct bpos end_pos, size_t *idx)
83 struct journal_keys *keys = &c->journal_keys;
85 struct journal_key *k;
87 BUG_ON(*idx > keys->nr);
90 *idx = __bch2_journal_key_search(keys, btree_id, level, pos);
93 __journal_key_cmp(btree_id, level, end_pos, idx_to_key(keys, *idx - 1)) <= 0) {
102 while ((k = *idx < keys->nr ? idx_to_key(keys, *idx) : NULL)) {
103 if (__journal_key_cmp(btree_id, level, end_pos, k) < 0)
106 if (k->overwritten) {
111 if (__journal_key_cmp(btree_id, level, pos, k) <= 0)
125 struct bkey_i *bch2_journal_keys_peek_slot(struct bch_fs *c, enum btree_id btree_id,
126 unsigned level, struct bpos pos)
130 return bch2_journal_keys_peek_upto(c, btree_id, level, pos, pos, &idx);
133 static void journal_iters_fix(struct bch_fs *c)
135 struct journal_keys *keys = &c->journal_keys;
136 /* The key we just inserted is immediately before the gap: */
137 size_t gap_end = keys->gap + (keys->size - keys->nr);
138 struct btree_and_journal_iter *iter;
141 * If an iterator points one after the key we just inserted, decrement
142 * the iterator so it points at the key we just inserted - if the
143 * decrement was unnecessary, bch2_btree_and_journal_iter_peek() will
146 list_for_each_entry(iter, &c->journal_iters, journal.list)
147 if (iter->journal.idx == gap_end)
148 iter->journal.idx = keys->gap - 1;
151 static void journal_iters_move_gap(struct bch_fs *c, size_t old_gap, size_t new_gap)
153 struct journal_keys *keys = &c->journal_keys;
154 struct journal_iter *iter;
155 size_t gap_size = keys->size - keys->nr;
157 list_for_each_entry(iter, &c->journal_iters, list) {
158 if (iter->idx > old_gap)
159 iter->idx -= gap_size;
160 if (iter->idx >= new_gap)
161 iter->idx += gap_size;
165 int bch2_journal_key_insert_take(struct bch_fs *c, enum btree_id id,
166 unsigned level, struct bkey_i *k)
168 struct journal_key n = {
174 * Ensure these keys are done last by journal replay, to unblock
177 .journal_seq = U32_MAX,
179 struct journal_keys *keys = &c->journal_keys;
180 size_t idx = bch2_journal_key_search(keys, id, level, k->k.p);
182 BUG_ON(test_bit(BCH_FS_rw, &c->flags));
184 if (idx < keys->size &&
185 journal_key_cmp(&n, &keys->d[idx]) == 0) {
186 if (keys->d[idx].allocated)
187 kfree(keys->d[idx].k);
193 idx -= keys->size - keys->nr;
195 if (keys->nr == keys->size) {
196 struct journal_keys new_keys = {
198 .size = max_t(size_t, keys->size, 8) * 2,
201 new_keys.d = kvmalloc_array(new_keys.size, sizeof(new_keys.d[0]), GFP_KERNEL);
203 bch_err(c, "%s: error allocating new key array (size %zu)",
204 __func__, new_keys.size);
205 return -BCH_ERR_ENOMEM_journal_key_insert;
208 /* Since @keys was full, there was no gap: */
209 memcpy(new_keys.d, keys->d, sizeof(keys->d[0]) * keys->nr);
211 keys->d = new_keys.d;
212 keys->nr = new_keys.nr;
213 keys->size = new_keys.size;
215 /* And now the gap is at the end: */
216 keys->gap = keys->nr;
219 journal_iters_move_gap(c, keys->gap, idx);
221 move_gap(keys->d, keys->nr, keys->size, keys->gap, idx);
225 keys->d[keys->gap++] = n;
227 journal_iters_fix(c);
233 * Can only be used from the recovery thread while we're still RO - can't be
234 * used once we've got RW, as journal_keys is at that point used by multiple
237 int bch2_journal_key_insert(struct bch_fs *c, enum btree_id id,
238 unsigned level, struct bkey_i *k)
243 n = kmalloc(bkey_bytes(&k->k), GFP_KERNEL);
245 return -BCH_ERR_ENOMEM_journal_key_insert;
248 ret = bch2_journal_key_insert_take(c, id, level, n);
254 int bch2_journal_key_delete(struct bch_fs *c, enum btree_id id,
255 unsigned level, struct bpos pos)
257 struct bkey_i whiteout;
259 bkey_init(&whiteout.k);
262 return bch2_journal_key_insert(c, id, level, &whiteout);
265 void bch2_journal_key_overwritten(struct bch_fs *c, enum btree_id btree,
266 unsigned level, struct bpos pos)
268 struct journal_keys *keys = &c->journal_keys;
269 size_t idx = bch2_journal_key_search(keys, btree, level, pos);
271 if (idx < keys->size &&
272 keys->d[idx].btree_id == btree &&
273 keys->d[idx].level == level &&
274 bpos_eq(keys->d[idx].k->k.p, pos))
275 keys->d[idx].overwritten = true;
278 static void bch2_journal_iter_advance(struct journal_iter *iter)
280 if (iter->idx < iter->keys->size) {
282 if (iter->idx == iter->keys->gap)
283 iter->idx += iter->keys->size - iter->keys->nr;
287 static struct bkey_s_c bch2_journal_iter_peek(struct journal_iter *iter)
289 struct journal_key *k = iter->keys->d + iter->idx;
291 while (k < iter->keys->d + iter->keys->size &&
292 k->btree_id == iter->btree_id &&
293 k->level == iter->level) {
295 return bkey_i_to_s_c(k->k);
297 bch2_journal_iter_advance(iter);
298 k = iter->keys->d + iter->idx;
301 return bkey_s_c_null;
304 static void bch2_journal_iter_exit(struct journal_iter *iter)
306 list_del(&iter->list);
309 static void bch2_journal_iter_init(struct bch_fs *c,
310 struct journal_iter *iter,
311 enum btree_id id, unsigned level,
316 iter->keys = &c->journal_keys;
317 iter->idx = bch2_journal_key_search(&c->journal_keys, id, level, pos);
320 static struct bkey_s_c bch2_journal_iter_peek_btree(struct btree_and_journal_iter *iter)
322 return bch2_btree_node_iter_peek_unpack(&iter->node_iter,
323 iter->b, &iter->unpacked);
326 static void bch2_journal_iter_advance_btree(struct btree_and_journal_iter *iter)
328 bch2_btree_node_iter_advance(&iter->node_iter, iter->b);
331 void bch2_btree_and_journal_iter_advance(struct btree_and_journal_iter *iter)
333 if (bpos_eq(iter->pos, SPOS_MAX))
336 iter->pos = bpos_successor(iter->pos);
339 static void btree_and_journal_iter_prefetch(struct btree_and_journal_iter *_iter)
341 struct btree_and_journal_iter iter = *_iter;
342 struct bch_fs *c = iter.trans->c;
343 unsigned level = iter.journal.level;
345 unsigned nr = test_bit(BCH_FS_started, &c->flags)
346 ? (level > 1 ? 0 : 2)
347 : (level > 1 ? 1 : 16);
349 iter.prefetch = false;
350 bch2_bkey_buf_init(&tmp);
353 bch2_btree_and_journal_iter_advance(&iter);
354 struct bkey_s_c k = bch2_btree_and_journal_iter_peek(&iter);
358 bch2_bkey_buf_reassemble(&tmp, c, k);
359 bch2_btree_node_prefetch(iter.trans, NULL, tmp.k, iter.journal.btree_id, level - 1);
362 bch2_bkey_buf_exit(&tmp, c);
365 struct bkey_s_c bch2_btree_and_journal_iter_peek(struct btree_and_journal_iter *iter)
367 struct bkey_s_c btree_k, journal_k, ret;
369 if (iter->prefetch && iter->journal.level)
370 btree_and_journal_iter_prefetch(iter);
373 return bkey_s_c_null;
375 while ((btree_k = bch2_journal_iter_peek_btree(iter)).k &&
376 bpos_lt(btree_k.k->p, iter->pos))
377 bch2_journal_iter_advance_btree(iter);
379 while ((journal_k = bch2_journal_iter_peek(&iter->journal)).k &&
380 bpos_lt(journal_k.k->p, iter->pos))
381 bch2_journal_iter_advance(&iter->journal);
384 (!btree_k.k || bpos_le(journal_k.k->p, btree_k.k->p))
388 if (ret.k && iter->b && bpos_gt(ret.k->p, iter->b->data->max_key))
392 iter->pos = ret.k->p;
393 if (bkey_deleted(ret.k)) {
394 bch2_btree_and_journal_iter_advance(iter);
398 iter->pos = SPOS_MAX;
405 void bch2_btree_and_journal_iter_exit(struct btree_and_journal_iter *iter)
407 bch2_journal_iter_exit(&iter->journal);
410 void __bch2_btree_and_journal_iter_init_node_iter(struct btree_trans *trans,
411 struct btree_and_journal_iter *iter,
413 struct btree_node_iter node_iter,
416 memset(iter, 0, sizeof(*iter));
420 iter->node_iter = node_iter;
421 bch2_journal_iter_init(trans->c, &iter->journal, b->c.btree_id, b->c.level, pos);
422 INIT_LIST_HEAD(&iter->journal.list);
423 iter->pos = b->data->min_key;
424 iter->at_end = false;
428 * this version is used by btree_gc before filesystem has gone RW and
429 * multithreaded, so uses the journal_iters list:
431 void bch2_btree_and_journal_iter_init_node_iter(struct btree_trans *trans,
432 struct btree_and_journal_iter *iter,
435 struct btree_node_iter node_iter;
437 bch2_btree_node_iter_init_from_start(&node_iter, b);
438 __bch2_btree_and_journal_iter_init_node_iter(trans, iter, b, node_iter, b->data->min_key);
439 list_add(&iter->journal.list, &trans->c->journal_iters);
442 /* sort and dedup all keys in the journal: */
444 void bch2_journal_entries_free(struct bch_fs *c)
446 struct journal_replay **i;
447 struct genradix_iter iter;
449 genradix_for_each(&c->journal_entries, iter, i)
451 genradix_free(&c->journal_entries);
455 * When keys compare equal, oldest compares first:
457 static int journal_sort_key_cmp(const void *_l, const void *_r)
459 const struct journal_key *l = _l;
460 const struct journal_key *r = _r;
462 return journal_key_cmp(l, r) ?:
463 cmp_int(l->journal_seq, r->journal_seq) ?:
464 cmp_int(l->journal_offset, r->journal_offset);
467 void bch2_journal_keys_put(struct bch_fs *c)
469 struct journal_keys *keys = &c->journal_keys;
470 struct journal_key *i;
472 BUG_ON(atomic_read(&keys->ref) <= 0);
474 if (!atomic_dec_and_test(&keys->ref))
477 move_gap(keys->d, keys->nr, keys->size, keys->gap, keys->nr);
478 keys->gap = keys->nr;
480 for (i = keys->d; i < keys->d + keys->nr; i++)
486 keys->nr = keys->gap = keys->size = 0;
488 bch2_journal_entries_free(c);
491 static void __journal_keys_sort(struct journal_keys *keys)
493 struct journal_key *src, *dst;
495 sort(keys->d, keys->nr, sizeof(keys->d[0]), journal_sort_key_cmp, NULL);
498 while (src < keys->d + keys->nr) {
499 while (src + 1 < keys->d + keys->nr &&
500 !journal_key_cmp(src, src + 1))
506 keys->nr = dst - keys->d;
509 int bch2_journal_keys_sort(struct bch_fs *c)
511 struct genradix_iter iter;
512 struct journal_replay *i, **_i;
513 struct jset_entry *entry;
515 struct journal_keys *keys = &c->journal_keys;
516 size_t nr_keys = 0, nr_read = 0;
518 genradix_for_each(&c->journal_entries, iter, _i) {
524 for_each_jset_key(k, entry, &i->j)
531 keys->size = roundup_pow_of_two(nr_keys);
533 keys->d = kvmalloc_array(keys->size, sizeof(keys->d[0]), GFP_KERNEL);
535 bch_err(c, "Failed to allocate buffer for sorted journal keys (%zu keys); trying slowpath",
540 keys->d = kvmalloc_array(keys->size, sizeof(keys->d[0]), GFP_KERNEL);
541 } while (!keys->d && keys->size > nr_keys / 8);
544 bch_err(c, "Failed to allocate %zu size buffer for sorted journal keys; exiting",
546 return -BCH_ERR_ENOMEM_journal_keys_sort;
550 genradix_for_each(&c->journal_entries, iter, _i) {
558 for_each_jset_key(k, entry, &i->j) {
559 if (keys->nr == keys->size) {
560 __journal_keys_sort(keys);
562 if (keys->nr > keys->size * 7 / 8) {
563 bch_err(c, "Too many journal keys for slowpath; have %zu compacted, buf size %zu, processed %zu/%zu",
564 keys->nr, keys->size, nr_read, nr_keys);
565 return -BCH_ERR_ENOMEM_journal_keys_sort;
569 keys->d[keys->nr++] = (struct journal_key) {
570 .btree_id = entry->btree_id,
571 .level = entry->level,
573 .journal_seq = le64_to_cpu(i->j.seq),
574 .journal_offset = k->_data - i->j._data,
581 __journal_keys_sort(keys);
582 keys->gap = keys->nr;
584 bch_verbose(c, "Journal keys: %zu read, %zu after sorting and compacting", nr_keys, keys->nr);