1 // SPDX-License-Identifier: GPL-2.0
2
3 #include "bcachefs.h"
4 #include "alloc_foreground.h"
5 #include "bkey_methods.h"
6 #include "btree_cache.h"
7 #include "btree_gc.h"
8 #include "btree_journal_iter.h"
9 #include "btree_update.h"
10 #include "btree_update_interior.h"
11 #include "btree_io.h"
12 #include "btree_iter.h"
13 #include "btree_locking.h"
14 #include "buckets.h"
15 #include "clock.h"
16 #include "error.h"
17 #include "extents.h"
18 #include "journal.h"
19 #include "journal_reclaim.h"
20 #include "keylist.h"
21 #include "replicas.h"
22 #include "super-io.h"
23 #include "trace.h"
24
25 #include <linux/random.h>
26
27 static int bch2_btree_insert_node(struct btree_update *, struct btree_trans *,
28 btree_path_idx_t, struct btree *,
29 struct keylist *, unsigned);
30 static void bch2_btree_update_add_new_node(struct btree_update *, struct btree *);
31
get_unlocked_mut_path(struct btree_trans * trans,enum btree_id btree_id,unsigned level,struct bpos pos)32 static btree_path_idx_t get_unlocked_mut_path(struct btree_trans *trans,
33 enum btree_id btree_id,
34 unsigned level,
35 struct bpos pos)
36 {
37 btree_path_idx_t path_idx = bch2_path_get(trans, btree_id, pos, level + 1, level,
38 BTREE_ITER_NOPRESERVE|
39 BTREE_ITER_INTENT, _RET_IP_);
40 path_idx = bch2_btree_path_make_mut(trans, path_idx, true, _RET_IP_);
41
42 struct btree_path *path = trans->paths + path_idx;
43 bch2_btree_path_downgrade(trans, path);
44 __bch2_btree_path_unlock(trans, path);
45 return path_idx;
46 }
47
48 /* Debug code: */
49
50 /*
51 * Verify that child nodes correctly span parent node's range:
52 */
btree_node_interior_verify(struct bch_fs * c,struct btree * b)53 static void btree_node_interior_verify(struct bch_fs *c, struct btree *b)
54 {
55 #ifdef CONFIG_BCACHEFS_DEBUG
56 struct bpos next_node = b->data->min_key;
57 struct btree_node_iter iter;
58 struct bkey_s_c k;
59 struct bkey_s_c_btree_ptr_v2 bp;
60 struct bkey unpacked;
61 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
62
63 BUG_ON(!b->c.level);
64
65 if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
66 return;
67
68 bch2_btree_node_iter_init_from_start(&iter, b);
69
70 while (1) {
71 k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked);
72 if (k.k->type != KEY_TYPE_btree_ptr_v2)
73 break;
74 bp = bkey_s_c_to_btree_ptr_v2(k);
75
76 if (!bpos_eq(next_node, bp.v->min_key)) {
77 bch2_dump_btree_node(c, b);
78 bch2_bpos_to_text(&buf1, next_node);
79 bch2_bpos_to_text(&buf2, bp.v->min_key);
80 panic("expected next min_key %s got %s\n", buf1.buf, buf2.buf);
81 }
82
83 bch2_btree_node_iter_advance(&iter, b);
84
85 if (bch2_btree_node_iter_end(&iter)) {
86 if (!bpos_eq(k.k->p, b->key.k.p)) {
87 bch2_dump_btree_node(c, b);
88 bch2_bpos_to_text(&buf1, b->key.k.p);
89 bch2_bpos_to_text(&buf2, k.k->p);
90 panic("expected end %s got %s\n", buf1.buf, buf2.buf);
91 }
92 break;
93 }
94
95 next_node = bpos_successor(k.k->p);
96 }
97 #endif
98 }
99
100 /* Calculate ideal packed bkey format for new btree nodes: */
101
__bch2_btree_calc_format(struct bkey_format_state * s,struct btree * b)102 static void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
103 {
104 struct bkey_packed *k;
105 struct bset_tree *t;
106 struct bkey uk;
107
108 for_each_bset(b, t)
109 bset_tree_for_each_key(b, t, k)
110 if (!bkey_deleted(k)) {
111 uk = bkey_unpack_key(b, k);
112 bch2_bkey_format_add_key(s, &uk);
113 }
114 }
115
bch2_btree_calc_format(struct btree * b)116 static struct bkey_format bch2_btree_calc_format(struct btree *b)
117 {
118 struct bkey_format_state s;
119
120 bch2_bkey_format_init(&s);
121 bch2_bkey_format_add_pos(&s, b->data->min_key);
122 bch2_bkey_format_add_pos(&s, b->data->max_key);
123 __bch2_btree_calc_format(&s, b);
124
125 return bch2_bkey_format_done(&s);
126 }
127
btree_node_u64s_with_format(struct btree_nr_keys nr,struct bkey_format * old_f,struct bkey_format * new_f)128 static size_t btree_node_u64s_with_format(struct btree_nr_keys nr,
129 struct bkey_format *old_f,
130 struct bkey_format *new_f)
131 {
132 /* stupid integer promotion rules */
133 ssize_t delta =
134 (((int) new_f->key_u64s - old_f->key_u64s) *
135 (int) nr.packed_keys) +
136 (((int) new_f->key_u64s - BKEY_U64s) *
137 (int) nr.unpacked_keys);
138
139 BUG_ON(delta + nr.live_u64s < 0);
140
141 return nr.live_u64s + delta;
142 }
143
144 /**
145 * bch2_btree_node_format_fits - check if we could rewrite node with a new format
146 *
147 * @c: filesystem handle
148 * @b: btree node to rewrite
149 * @nr: number of keys for new node (i.e. b->nr)
150 * @new_f: bkey format to translate keys to
151 *
152 * Returns: true if all re-packed keys will be able to fit in a new node.
153 *
154 * Assumes all keys will successfully pack with the new format.
155 */
bch2_btree_node_format_fits(struct bch_fs * c,struct btree * b,struct btree_nr_keys nr,struct bkey_format * new_f)156 static bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
157 struct btree_nr_keys nr,
158 struct bkey_format *new_f)
159 {
160 size_t u64s = btree_node_u64s_with_format(nr, &b->format, new_f);
161
162 return __vstruct_bytes(struct btree_node, u64s) < btree_buf_bytes(b);
163 }
164
165 /* Btree node freeing/allocation: */
166
__btree_node_free(struct btree_trans * trans,struct btree * b)167 static void __btree_node_free(struct btree_trans *trans, struct btree *b)
168 {
169 struct bch_fs *c = trans->c;
170
171 trace_and_count(c, btree_node_free, trans, b);
172
173 BUG_ON(btree_node_write_blocked(b));
174 BUG_ON(btree_node_dirty(b));
175 BUG_ON(btree_node_need_write(b));
176 BUG_ON(b == btree_node_root(c, b));
177 BUG_ON(b->ob.nr);
178 BUG_ON(!list_empty(&b->write_blocked));
179 BUG_ON(b->will_make_reachable);
180
181 clear_btree_node_noevict(b);
182
183 mutex_lock(&c->btree_cache.lock);
184 list_move(&b->list, &c->btree_cache.freeable);
185 mutex_unlock(&c->btree_cache.lock);
186 }
187
bch2_btree_node_free_inmem(struct btree_trans * trans,struct btree_path * path,struct btree * b)188 static void bch2_btree_node_free_inmem(struct btree_trans *trans,
189 struct btree_path *path,
190 struct btree *b)
191 {
192 struct bch_fs *c = trans->c;
193 unsigned i, level = b->c.level;
194
195 bch2_btree_node_lock_write_nofail(trans, path, &b->c);
196 bch2_btree_node_hash_remove(&c->btree_cache, b);
197 __btree_node_free(trans, b);
198 six_unlock_write(&b->c.lock);
199 mark_btree_node_locked_noreset(path, level, BTREE_NODE_INTENT_LOCKED);
200
201 trans_for_each_path(trans, path, i)
202 if (path->l[level].b == b) {
203 btree_node_unlock(trans, path, level);
204 path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
205 }
206 }
207
bch2_btree_node_free_never_used(struct btree_update * as,struct btree_trans * trans,struct btree * b)208 static void bch2_btree_node_free_never_used(struct btree_update *as,
209 struct btree_trans *trans,
210 struct btree *b)
211 {
212 struct bch_fs *c = as->c;
213 struct prealloc_nodes *p = &as->prealloc_nodes[b->c.lock.readers != NULL];
214 struct btree_path *path;
215 unsigned i, level = b->c.level;
216
217 BUG_ON(!list_empty(&b->write_blocked));
218 BUG_ON(b->will_make_reachable != (1UL|(unsigned long) as));
219
220 b->will_make_reachable = 0;
221 closure_put(&as->cl);
222
223 clear_btree_node_will_make_reachable(b);
224 clear_btree_node_accessed(b);
225 clear_btree_node_dirty_acct(c, b);
226 clear_btree_node_need_write(b);
227
228 mutex_lock(&c->btree_cache.lock);
229 list_del_init(&b->list);
230 bch2_btree_node_hash_remove(&c->btree_cache, b);
231 mutex_unlock(&c->btree_cache.lock);
232
233 BUG_ON(p->nr >= ARRAY_SIZE(p->b));
234 p->b[p->nr++] = b;
235
236 six_unlock_intent(&b->c.lock);
237
238 trans_for_each_path(trans, path, i)
239 if (path->l[level].b == b) {
240 btree_node_unlock(trans, path, level);
241 path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
242 }
243 }
244
__bch2_btree_node_alloc(struct btree_trans * trans,struct disk_reservation * res,struct closure * cl,bool interior_node,unsigned flags)245 static struct btree *__bch2_btree_node_alloc(struct btree_trans *trans,
246 struct disk_reservation *res,
247 struct closure *cl,
248 bool interior_node,
249 unsigned flags)
250 {
251 struct bch_fs *c = trans->c;
252 struct write_point *wp;
253 struct btree *b;
254 BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
255 struct open_buckets obs = { .nr = 0 };
256 struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
257 enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
258 unsigned nr_reserve = watermark > BCH_WATERMARK_reclaim
259 ? BTREE_NODE_RESERVE
260 : 0;
261 int ret;
262
263 mutex_lock(&c->btree_reserve_cache_lock);
264 if (c->btree_reserve_cache_nr > nr_reserve) {
265 struct btree_alloc *a =
266 &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
267
268 obs = a->ob;
269 bkey_copy(&tmp.k, &a->k);
270 mutex_unlock(&c->btree_reserve_cache_lock);
271 goto mem_alloc;
272 }
273 mutex_unlock(&c->btree_reserve_cache_lock);
274
275 retry:
276 ret = bch2_alloc_sectors_start_trans(trans,
277 c->opts.metadata_target ?:
278 c->opts.foreground_target,
279 0,
280 writepoint_ptr(&c->btree_write_point),
281 &devs_have,
282 res->nr_replicas,
283 min(res->nr_replicas,
284 c->opts.metadata_replicas_required),
285 watermark, 0, cl, &wp);
286 if (unlikely(ret))
287 return ERR_PTR(ret);
288
289 if (wp->sectors_free < btree_sectors(c)) {
290 struct open_bucket *ob;
291 unsigned i;
292
293 open_bucket_for_each(c, &wp->ptrs, ob, i)
294 if (ob->sectors_free < btree_sectors(c))
295 ob->sectors_free = 0;
296
297 bch2_alloc_sectors_done(c, wp);
298 goto retry;
299 }
300
301 bkey_btree_ptr_v2_init(&tmp.k);
302 bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, btree_sectors(c), false);
303
304 bch2_open_bucket_get(c, wp, &obs);
305 bch2_alloc_sectors_done(c, wp);
306 mem_alloc:
307 b = bch2_btree_node_mem_alloc(trans, interior_node);
308 six_unlock_write(&b->c.lock);
309 six_unlock_intent(&b->c.lock);
310
311 /* we hold cannibalize_lock: */
312 BUG_ON(IS_ERR(b));
313 BUG_ON(b->ob.nr);
314
315 bkey_copy(&b->key, &tmp.k);
316 b->ob = obs;
317
318 return b;
319 }
320
bch2_btree_node_alloc(struct btree_update * as,struct btree_trans * trans,unsigned level)321 static struct btree *bch2_btree_node_alloc(struct btree_update *as,
322 struct btree_trans *trans,
323 unsigned level)
324 {
325 struct bch_fs *c = as->c;
326 struct btree *b;
327 struct prealloc_nodes *p = &as->prealloc_nodes[!!level];
328 int ret;
329
330 BUG_ON(level >= BTREE_MAX_DEPTH);
331 BUG_ON(!p->nr);
332
333 b = p->b[--p->nr];
334
335 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
336 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
337
338 set_btree_node_accessed(b);
339 set_btree_node_dirty_acct(c, b);
340 set_btree_node_need_write(b);
341
342 bch2_bset_init_first(b, &b->data->keys);
343 b->c.level = level;
344 b->c.btree_id = as->btree_id;
345 b->version_ondisk = c->sb.version;
346
347 memset(&b->nr, 0, sizeof(b->nr));
348 b->data->magic = cpu_to_le64(bset_magic(c));
349 memset(&b->data->_ptr, 0, sizeof(b->data->_ptr));
350 b->data->flags = 0;
351 SET_BTREE_NODE_ID(b->data, as->btree_id);
352 SET_BTREE_NODE_LEVEL(b->data, level);
353
354 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
355 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
356
357 bp->v.mem_ptr = 0;
358 bp->v.seq = b->data->keys.seq;
359 bp->v.sectors_written = 0;
360 }
361
362 SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
363
364 bch2_btree_build_aux_trees(b);
365
366 ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
367 BUG_ON(ret);
368
369 trace_and_count(c, btree_node_alloc, trans, b);
370 bch2_increment_clock(c, btree_sectors(c), WRITE);
371 return b;
372 }
373
btree_set_min(struct btree * b,struct bpos pos)374 static void btree_set_min(struct btree *b, struct bpos pos)
375 {
376 if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
377 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
378 b->data->min_key = pos;
379 }
380
btree_set_max(struct btree * b,struct bpos pos)381 static void btree_set_max(struct btree *b, struct bpos pos)
382 {
383 b->key.k.p = pos;
384 b->data->max_key = pos;
385 }
386
bch2_btree_node_alloc_replacement(struct btree_update * as,struct btree_trans * trans,struct btree * b)387 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
388 struct btree_trans *trans,
389 struct btree *b)
390 {
391 struct btree *n = bch2_btree_node_alloc(as, trans, b->c.level);
392 struct bkey_format format = bch2_btree_calc_format(b);
393
394 /*
395 * The keys might expand with the new format - if they wouldn't fit in
396 * the btree node anymore, use the old format for now:
397 */
398 if (!bch2_btree_node_format_fits(as->c, b, b->nr, &format))
399 format = b->format;
400
401 SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
402
403 btree_set_min(n, b->data->min_key);
404 btree_set_max(n, b->data->max_key);
405
406 n->data->format = format;
407 btree_node_set_format(n, format);
408
409 bch2_btree_sort_into(as->c, n, b);
410
411 btree_node_reset_sib_u64s(n);
412 return n;
413 }
414
__btree_root_alloc(struct btree_update * as,struct btree_trans * trans,unsigned level)415 static struct btree *__btree_root_alloc(struct btree_update *as,
416 struct btree_trans *trans, unsigned level)
417 {
418 struct btree *b = bch2_btree_node_alloc(as, trans, level);
419
420 btree_set_min(b, POS_MIN);
421 btree_set_max(b, SPOS_MAX);
422 b->data->format = bch2_btree_calc_format(b);
423
424 btree_node_set_format(b, b->data->format);
425 bch2_btree_build_aux_trees(b);
426
427 return b;
428 }
429
bch2_btree_reserve_put(struct btree_update * as,struct btree_trans * trans)430 static void bch2_btree_reserve_put(struct btree_update *as, struct btree_trans *trans)
431 {
432 struct bch_fs *c = as->c;
433 struct prealloc_nodes *p;
434
435 for (p = as->prealloc_nodes;
436 p < as->prealloc_nodes + ARRAY_SIZE(as->prealloc_nodes);
437 p++) {
438 while (p->nr) {
439 struct btree *b = p->b[--p->nr];
440
441 mutex_lock(&c->btree_reserve_cache_lock);
442
443 if (c->btree_reserve_cache_nr <
444 ARRAY_SIZE(c->btree_reserve_cache)) {
445 struct btree_alloc *a =
446 &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
447
448 a->ob = b->ob;
449 b->ob.nr = 0;
450 bkey_copy(&a->k, &b->key);
451 } else {
452 bch2_open_buckets_put(c, &b->ob);
453 }
454
455 mutex_unlock(&c->btree_reserve_cache_lock);
456
457 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
458 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
459 __btree_node_free(trans, b);
460 six_unlock_write(&b->c.lock);
461 six_unlock_intent(&b->c.lock);
462 }
463 }
464 }
465
bch2_btree_reserve_get(struct btree_trans * trans,struct btree_update * as,unsigned nr_nodes[2],unsigned flags,struct closure * cl)466 static int bch2_btree_reserve_get(struct btree_trans *trans,
467 struct btree_update *as,
468 unsigned nr_nodes[2],
469 unsigned flags,
470 struct closure *cl)
471 {
472 struct btree *b;
473 unsigned interior;
474 int ret = 0;
475
476 BUG_ON(nr_nodes[0] + nr_nodes[1] > BTREE_RESERVE_MAX);
477
478 /*
479 * Protects reaping from the btree node cache and using the btree node
480 * open bucket reserve:
481 */
482 ret = bch2_btree_cache_cannibalize_lock(trans, cl);
483 if (ret)
484 return ret;
485
486 for (interior = 0; interior < 2; interior++) {
487 struct prealloc_nodes *p = as->prealloc_nodes + interior;
488
489 while (p->nr < nr_nodes[interior]) {
490 b = __bch2_btree_node_alloc(trans, &as->disk_res, cl,
491 interior, flags);
492 if (IS_ERR(b)) {
493 ret = PTR_ERR(b);
494 goto err;
495 }
496
497 p->b[p->nr++] = b;
498 }
499 }
500 err:
501 bch2_btree_cache_cannibalize_unlock(trans);
502 return ret;
503 }
504
505 /* Asynchronous interior node update machinery */
506
bch2_btree_update_free(struct btree_update * as,struct btree_trans * trans)507 static void bch2_btree_update_free(struct btree_update *as, struct btree_trans *trans)
508 {
509 struct bch_fs *c = as->c;
510
511 if (as->took_gc_lock)
512 up_read(&c->gc_lock);
513 as->took_gc_lock = false;
514
515 bch2_journal_pin_drop(&c->journal, &as->journal);
516 bch2_journal_pin_flush(&c->journal, &as->journal);
517 bch2_disk_reservation_put(c, &as->disk_res);
518 bch2_btree_reserve_put(as, trans);
519
520 bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_total],
521 as->start_time);
522
523 mutex_lock(&c->btree_interior_update_lock);
524 list_del(&as->unwritten_list);
525 list_del(&as->list);
526
527 closure_debug_destroy(&as->cl);
528 mempool_free(as, &c->btree_interior_update_pool);
529
530 /*
531 * Have to do the wakeup with btree_interior_update_lock still held,
532 * since being on btree_interior_update_list is our ref on @c:
533 */
534 closure_wake_up(&c->btree_interior_update_wait);
535
536 mutex_unlock(&c->btree_interior_update_lock);
537 }
538
btree_update_add_key(struct btree_update * as,struct keylist * keys,struct btree * b)539 static void btree_update_add_key(struct btree_update *as,
540 struct keylist *keys, struct btree *b)
541 {
542 struct bkey_i *k = &b->key;
543
544 BUG_ON(bch2_keylist_u64s(keys) + k->k.u64s >
545 ARRAY_SIZE(as->_old_keys));
546
547 bkey_copy(keys->top, k);
548 bkey_i_to_btree_ptr_v2(keys->top)->v.mem_ptr = b->c.level + 1;
549
550 bch2_keylist_push(keys);
551 }
552
553 /*
554 * The transactional part of an interior btree node update, where we journal the
555 * update we did to the interior node and update alloc info:
556 */
btree_update_nodes_written_trans(struct btree_trans * trans,struct btree_update * as)557 static int btree_update_nodes_written_trans(struct btree_trans *trans,
558 struct btree_update *as)
559 {
560 struct jset_entry *e = bch2_trans_jset_entry_alloc(trans, as->journal_u64s);
561 int ret = PTR_ERR_OR_ZERO(e);
562 if (ret)
563 return ret;
564
565 memcpy(e, as->journal_entries, as->journal_u64s * sizeof(u64));
566
567 trans->journal_pin = &as->journal;
568
569 for_each_keylist_key(&as->old_keys, k) {
570 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
571
572 ret = bch2_key_trigger_old(trans, as->btree_id, level, bkey_i_to_s_c(k),
573 BTREE_TRIGGER_TRANSACTIONAL);
574 if (ret)
575 return ret;
576 }
577
578 for_each_keylist_key(&as->new_keys, k) {
579 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
580
581 ret = bch2_key_trigger_new(trans, as->btree_id, level, bkey_i_to_s(k),
582 BTREE_TRIGGER_TRANSACTIONAL);
583 if (ret)
584 return ret;
585 }
586
587 return 0;
588 }
589
btree_update_nodes_written(struct btree_update * as)590 static void btree_update_nodes_written(struct btree_update *as)
591 {
592 struct bch_fs *c = as->c;
593 struct btree *b;
594 struct btree_trans *trans = bch2_trans_get(c);
595 u64 journal_seq = 0;
596 unsigned i;
597 int ret;
598
599 /*
600 * If we're already in an error state, it might be because a btree node
601 * was never written, and we might be trying to free that same btree
602 * node here, but it won't have been marked as allocated and we'll see
603 * spurious disk usage inconsistencies in the transactional part below
604 * if we don't skip it:
605 */
606 ret = bch2_journal_error(&c->journal);
607 if (ret)
608 goto err;
609
610 /*
611 * Wait for any in flight writes to finish before we free the old nodes
612 * on disk:
613 */
614 for (i = 0; i < as->nr_old_nodes; i++) {
615 __le64 seq;
616
617 b = as->old_nodes[i];
618
619 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
620 seq = b->data ? b->data->keys.seq : 0;
621 six_unlock_read(&b->c.lock);
622
623 if (seq == as->old_nodes_seq[i])
624 wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight_inner,
625 TASK_UNINTERRUPTIBLE);
626 }
627
628 /*
629 * We did an update to a parent node where the pointers we added pointed
630 * to child nodes that weren't written yet: now, the child nodes have
631 * been written so we can write out the update to the interior node.
632 */
633
634 /*
635 * We can't call into journal reclaim here: we'd block on the journal
636 * reclaim lock, but we may need to release the open buckets we have
637 * pinned in order for other btree updates to make forward progress, and
638 * journal reclaim does btree updates when flushing bkey_cached entries,
639 * which may require allocations as well.
640 */
641 ret = commit_do(trans, &as->disk_res, &journal_seq,
642 BCH_WATERMARK_reclaim|
643 BCH_TRANS_COMMIT_no_enospc|
644 BCH_TRANS_COMMIT_no_check_rw|
645 BCH_TRANS_COMMIT_journal_reclaim,
646 btree_update_nodes_written_trans(trans, as));
647 bch2_trans_unlock(trans);
648
649 bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
650 "%s(): error %s", __func__, bch2_err_str(ret));
651 err:
652 if (as->b) {
653
654 b = as->b;
655 btree_path_idx_t path_idx = get_unlocked_mut_path(trans,
656 as->btree_id, b->c.level, b->key.k.p);
657 struct btree_path *path = trans->paths + path_idx;
658 /*
659 * @b is the node we did the final insert into:
660 *
661 * On failure to get a journal reservation, we still have to
662 * unblock the write and allow most of the write path to happen
663 * so that shutdown works, but the i->journal_seq mechanism
664 * won't work to prevent the btree write from being visible (we
665 * didn't get a journal sequence number) - instead
666 * __bch2_btree_node_write() doesn't do the actual write if
667 * we're in journal error state:
668 */
669
670 /*
671 * Ensure transaction is unlocked before using
672 * btree_node_lock_nopath() (the use of which is always suspect,
673 * we need to work on removing this in the future)
674 *
675 * It should be, but get_unlocked_mut_path() -> bch2_path_get()
676 * calls bch2_path_upgrade(), before we call path_make_mut(), so
677 * we may rarely end up with a locked path besides the one we
678 * have here:
679 */
680 bch2_trans_unlock(trans);
681 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
682 mark_btree_node_locked(trans, path, b->c.level, BTREE_NODE_INTENT_LOCKED);
683 path->l[b->c.level].lock_seq = six_lock_seq(&b->c.lock);
684 path->l[b->c.level].b = b;
685
686 bch2_btree_node_lock_write_nofail(trans, path, &b->c);
687
688 mutex_lock(&c->btree_interior_update_lock);
689
690 list_del(&as->write_blocked_list);
691 if (list_empty(&b->write_blocked))
692 clear_btree_node_write_blocked(b);
693
694 /*
695 * Node might have been freed, recheck under
696 * btree_interior_update_lock:
697 */
698 if (as->b == b) {
699 BUG_ON(!b->c.level);
700 BUG_ON(!btree_node_dirty(b));
701
702 if (!ret) {
703 struct bset *last = btree_bset_last(b);
704
705 last->journal_seq = cpu_to_le64(
706 max(journal_seq,
707 le64_to_cpu(last->journal_seq)));
708
709 bch2_btree_add_journal_pin(c, b, journal_seq);
710 } else {
711 /*
712 * If we didn't get a journal sequence number we
713 * can't write this btree node, because recovery
714 * won't know to ignore this write:
715 */
716 set_btree_node_never_write(b);
717 }
718 }
719
720 mutex_unlock(&c->btree_interior_update_lock);
721
722 mark_btree_node_locked_noreset(path, b->c.level, BTREE_NODE_INTENT_LOCKED);
723 six_unlock_write(&b->c.lock);
724
725 btree_node_write_if_need(c, b, SIX_LOCK_intent);
726 btree_node_unlock(trans, path, b->c.level);
727 bch2_path_put(trans, path_idx, true);
728 }
729
730 bch2_journal_pin_drop(&c->journal, &as->journal);
731
732 mutex_lock(&c->btree_interior_update_lock);
733 for (i = 0; i < as->nr_new_nodes; i++) {
734 b = as->new_nodes[i];
735
736 BUG_ON(b->will_make_reachable != (unsigned long) as);
737 b->will_make_reachable = 0;
738 clear_btree_node_will_make_reachable(b);
739 }
740 mutex_unlock(&c->btree_interior_update_lock);
741
742 for (i = 0; i < as->nr_new_nodes; i++) {
743 b = as->new_nodes[i];
744
745 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
746 btree_node_write_if_need(c, b, SIX_LOCK_read);
747 six_unlock_read(&b->c.lock);
748 }
749
750 for (i = 0; i < as->nr_open_buckets; i++)
751 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
752
753 bch2_btree_update_free(as, trans);
754 bch2_trans_put(trans);
755 }
756
btree_interior_update_work(struct work_struct * work)757 static void btree_interior_update_work(struct work_struct *work)
758 {
759 struct bch_fs *c =
760 container_of(work, struct bch_fs, btree_interior_update_work);
761 struct btree_update *as;
762
763 while (1) {
764 mutex_lock(&c->btree_interior_update_lock);
765 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
766 struct btree_update, unwritten_list);
767 if (as && !as->nodes_written)
768 as = NULL;
769 mutex_unlock(&c->btree_interior_update_lock);
770
771 if (!as)
772 break;
773
774 btree_update_nodes_written(as);
775 }
776 }
777
CLOSURE_CALLBACK(btree_update_set_nodes_written)778 static CLOSURE_CALLBACK(btree_update_set_nodes_written)
779 {
780 closure_type(as, struct btree_update, cl);
781 struct bch_fs *c = as->c;
782
783 mutex_lock(&c->btree_interior_update_lock);
784 as->nodes_written = true;
785 mutex_unlock(&c->btree_interior_update_lock);
786
787 queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
788 }
789
790 /*
791 * We're updating @b with pointers to nodes that haven't finished writing yet:
792 * block @b from being written until @as completes
793 */
btree_update_updated_node(struct btree_update * as,struct btree * b)794 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
795 {
796 struct bch_fs *c = as->c;
797
798 mutex_lock(&c->btree_interior_update_lock);
799 list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
800
801 BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
802 BUG_ON(!btree_node_dirty(b));
803 BUG_ON(!b->c.level);
804
805 as->mode = BTREE_INTERIOR_UPDATING_NODE;
806 as->b = b;
807
808 set_btree_node_write_blocked(b);
809 list_add(&as->write_blocked_list, &b->write_blocked);
810
811 mutex_unlock(&c->btree_interior_update_lock);
812 }
813
bch2_update_reparent_journal_pin_flush(struct journal * j,struct journal_entry_pin * _pin,u64 seq)814 static int bch2_update_reparent_journal_pin_flush(struct journal *j,
815 struct journal_entry_pin *_pin, u64 seq)
816 {
817 return 0;
818 }
819
btree_update_reparent(struct btree_update * as,struct btree_update * child)820 static void btree_update_reparent(struct btree_update *as,
821 struct btree_update *child)
822 {
823 struct bch_fs *c = as->c;
824
825 lockdep_assert_held(&c->btree_interior_update_lock);
826
827 child->b = NULL;
828 child->mode = BTREE_INTERIOR_UPDATING_AS;
829
830 bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal,
831 bch2_update_reparent_journal_pin_flush);
832 }
833
btree_update_updated_root(struct btree_update * as,struct btree * b)834 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
835 {
836 struct bkey_i *insert = &b->key;
837 struct bch_fs *c = as->c;
838
839 BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
840
841 BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
842 ARRAY_SIZE(as->journal_entries));
843
844 as->journal_u64s +=
845 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
846 BCH_JSET_ENTRY_btree_root,
847 b->c.btree_id, b->c.level,
848 insert, insert->k.u64s);
849
850 mutex_lock(&c->btree_interior_update_lock);
851 list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
852
853 as->mode = BTREE_INTERIOR_UPDATING_ROOT;
854 mutex_unlock(&c->btree_interior_update_lock);
855 }
856
857 /*
858 * bch2_btree_update_add_new_node:
859 *
860 * This causes @as to wait on @b to be written, before it gets to
861 * bch2_btree_update_nodes_written
862 *
863 * Additionally, it sets b->will_make_reachable to prevent any additional writes
864 * to @b from happening besides the first until @b is reachable on disk
865 *
866 * And it adds @b to the list of @as's new nodes, so that we can update sector
867 * counts in bch2_btree_update_nodes_written:
868 */
bch2_btree_update_add_new_node(struct btree_update * as,struct btree * b)869 static void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
870 {
871 struct bch_fs *c = as->c;
872
873 closure_get(&as->cl);
874
875 mutex_lock(&c->btree_interior_update_lock);
876 BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
877 BUG_ON(b->will_make_reachable);
878
879 as->new_nodes[as->nr_new_nodes++] = b;
880 b->will_make_reachable = 1UL|(unsigned long) as;
881 set_btree_node_will_make_reachable(b);
882
883 mutex_unlock(&c->btree_interior_update_lock);
884
885 btree_update_add_key(as, &as->new_keys, b);
886
887 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
888 unsigned bytes = vstruct_end(&b->data->keys) - (void *) b->data;
889 unsigned sectors = round_up(bytes, block_bytes(c)) >> 9;
890
891 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
892 cpu_to_le16(sectors);
893 }
894 }
895
896 /*
897 * returns true if @b was a new node
898 */
btree_update_drop_new_node(struct bch_fs * c,struct btree * b)899 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
900 {
901 struct btree_update *as;
902 unsigned long v;
903 unsigned i;
904
905 mutex_lock(&c->btree_interior_update_lock);
906 /*
907 * When b->will_make_reachable != 0, it owns a ref on as->cl that's
908 * dropped when it gets written by bch2_btree_complete_write - the
909 * xchg() is for synchronization with bch2_btree_complete_write:
910 */
911 v = xchg(&b->will_make_reachable, 0);
912 clear_btree_node_will_make_reachable(b);
913 as = (struct btree_update *) (v & ~1UL);
914
915 if (!as) {
916 mutex_unlock(&c->btree_interior_update_lock);
917 return;
918 }
919
920 for (i = 0; i < as->nr_new_nodes; i++)
921 if (as->new_nodes[i] == b)
922 goto found;
923
924 BUG();
925 found:
926 array_remove_item(as->new_nodes, as->nr_new_nodes, i);
927 mutex_unlock(&c->btree_interior_update_lock);
928
929 if (v & 1)
930 closure_put(&as->cl);
931 }
932
bch2_btree_update_get_open_buckets(struct btree_update * as,struct btree * b)933 static void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
934 {
935 while (b->ob.nr)
936 as->open_buckets[as->nr_open_buckets++] =
937 b->ob.v[--b->ob.nr];
938 }
939
bch2_btree_update_will_free_node_journal_pin_flush(struct journal * j,struct journal_entry_pin * _pin,u64 seq)940 static int bch2_btree_update_will_free_node_journal_pin_flush(struct journal *j,
941 struct journal_entry_pin *_pin, u64 seq)
942 {
943 return 0;
944 }
945
946 /*
947 * @b is being split/rewritten: it may have pointers to not-yet-written btree
948 * nodes and thus outstanding btree_updates - redirect @b's
949 * btree_updates to point to this btree_update:
950 */
bch2_btree_interior_update_will_free_node(struct btree_update * as,struct btree * b)951 static void bch2_btree_interior_update_will_free_node(struct btree_update *as,
952 struct btree *b)
953 {
954 struct bch_fs *c = as->c;
955 struct btree_update *p, *n;
956 struct btree_write *w;
957
958 set_btree_node_dying(b);
959
960 if (btree_node_fake(b))
961 return;
962
963 mutex_lock(&c->btree_interior_update_lock);
964
965 /*
966 * Does this node have any btree_update operations preventing
967 * it from being written?
968 *
969 * If so, redirect them to point to this btree_update: we can
970 * write out our new nodes, but we won't make them visible until those
971 * operations complete
972 */
973 list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
974 list_del_init(&p->write_blocked_list);
975 btree_update_reparent(as, p);
976
977 /*
978 * for flush_held_btree_writes() waiting on updates to flush or
979 * nodes to be writeable:
980 */
981 closure_wake_up(&c->btree_interior_update_wait);
982 }
983
984 clear_btree_node_dirty_acct(c, b);
985 clear_btree_node_need_write(b);
986 clear_btree_node_write_blocked(b);
987
988 /*
989 * Does this node have unwritten data that has a pin on the journal?
990 *
991 * If so, transfer that pin to the btree_update operation -
992 * note that if we're freeing multiple nodes, we only need to keep the
993 * oldest pin of any of the nodes we're freeing. We'll release the pin
994 * when the new nodes are persistent and reachable on disk:
995 */
996 w = btree_current_write(b);
997 bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal,
998 bch2_btree_update_will_free_node_journal_pin_flush);
999 bch2_journal_pin_drop(&c->journal, &w->journal);
1000
1001 w = btree_prev_write(b);
1002 bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal,
1003 bch2_btree_update_will_free_node_journal_pin_flush);
1004 bch2_journal_pin_drop(&c->journal, &w->journal);
1005
1006 mutex_unlock(&c->btree_interior_update_lock);
1007
1008 /*
1009 * Is this a node that isn't reachable on disk yet?
1010 *
1011 * Nodes that aren't reachable yet have writes blocked until they're
1012 * reachable - now that we've cancelled any pending writes and moved
1013 * things waiting on that write to wait on this update, we can drop this
1014 * node from the list of nodes that the other update is making
1015 * reachable, prior to freeing it:
1016 */
1017 btree_update_drop_new_node(c, b);
1018
1019 btree_update_add_key(as, &as->old_keys, b);
1020
1021 as->old_nodes[as->nr_old_nodes] = b;
1022 as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq;
1023 as->nr_old_nodes++;
1024 }
1025
bch2_btree_update_done(struct btree_update * as,struct btree_trans * trans)1026 static void bch2_btree_update_done(struct btree_update *as, struct btree_trans *trans)
1027 {
1028 struct bch_fs *c = as->c;
1029 u64 start_time = as->start_time;
1030
1031 BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
1032
1033 if (as->took_gc_lock)
1034 up_read(&as->c->gc_lock);
1035 as->took_gc_lock = false;
1036
1037 bch2_btree_reserve_put(as, trans);
1038
1039 continue_at(&as->cl, btree_update_set_nodes_written,
1040 as->c->btree_interior_update_worker);
1041
1042 bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_foreground],
1043 start_time);
1044 }
1045
1046 static struct btree_update *
bch2_btree_update_start(struct btree_trans * trans,struct btree_path * path,unsigned level,bool split,unsigned flags)1047 bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
1048 unsigned level, bool split, unsigned flags)
1049 {
1050 struct bch_fs *c = trans->c;
1051 struct btree_update *as;
1052 u64 start_time = local_clock();
1053 int disk_res_flags = (flags & BCH_TRANS_COMMIT_no_enospc)
1054 ? BCH_DISK_RESERVATION_NOFAIL : 0;
1055 unsigned nr_nodes[2] = { 0, 0 };
1056 unsigned update_level = level;
1057 enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
1058 int ret = 0;
1059 u32 restart_count = trans->restart_count;
1060
1061 BUG_ON(!path->should_be_locked);
1062
1063 if (watermark == BCH_WATERMARK_copygc)
1064 watermark = BCH_WATERMARK_btree_copygc;
1065 if (watermark < BCH_WATERMARK_btree)
1066 watermark = BCH_WATERMARK_btree;
1067
1068 flags &= ~BCH_WATERMARK_MASK;
1069 flags |= watermark;
1070
1071 if (!(flags & BCH_TRANS_COMMIT_journal_reclaim) &&
1072 watermark < c->journal.watermark) {
1073 struct journal_res res = { 0 };
1074
1075 ret = drop_locks_do(trans,
1076 bch2_journal_res_get(&c->journal, &res, 1,
1077 watermark|JOURNAL_RES_GET_CHECK));
1078 if (ret)
1079 return ERR_PTR(ret);
1080 }
1081
1082 while (1) {
1083 nr_nodes[!!update_level] += 1 + split;
1084 update_level++;
1085
1086 ret = bch2_btree_path_upgrade(trans, path, update_level + 1);
1087 if (ret)
1088 return ERR_PTR(ret);
1089
1090 if (!btree_path_node(path, update_level)) {
1091 /* Allocating new root? */
1092 nr_nodes[1] += split;
1093 update_level = BTREE_MAX_DEPTH;
1094 break;
1095 }
1096
1097 /*
1098 * Always check for space for two keys, even if we won't have to
1099 * split at prior level - it might have been a merge instead:
1100 */
1101 if (bch2_btree_node_insert_fits(path->l[update_level].b,
1102 BKEY_BTREE_PTR_U64s_MAX * 2))
1103 break;
1104
1105 split = path->l[update_level].b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c);
1106 }
1107
1108 if (!down_read_trylock(&c->gc_lock)) {
1109 ret = drop_locks_do(trans, (down_read(&c->gc_lock), 0));
1110 if (ret) {
1111 up_read(&c->gc_lock);
1112 return ERR_PTR(ret);
1113 }
1114 }
1115
1116 as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOFS);
1117 memset(as, 0, sizeof(*as));
1118 closure_init(&as->cl, NULL);
1119 as->c = c;
1120 as->start_time = start_time;
1121 as->mode = BTREE_INTERIOR_NO_UPDATE;
1122 as->took_gc_lock = true;
1123 as->btree_id = path->btree_id;
1124 as->update_level = update_level;
1125 INIT_LIST_HEAD(&as->list);
1126 INIT_LIST_HEAD(&as->unwritten_list);
1127 INIT_LIST_HEAD(&as->write_blocked_list);
1128 bch2_keylist_init(&as->old_keys, as->_old_keys);
1129 bch2_keylist_init(&as->new_keys, as->_new_keys);
1130 bch2_keylist_init(&as->parent_keys, as->inline_keys);
1131
1132 mutex_lock(&c->btree_interior_update_lock);
1133 list_add_tail(&as->list, &c->btree_interior_update_list);
1134 mutex_unlock(&c->btree_interior_update_lock);
1135
1136 /*
1137 * We don't want to allocate if we're in an error state, that can cause
1138 * deadlock on emergency shutdown due to open buckets getting stuck in
1139 * the btree_reserve_cache after allocator shutdown has cleared it out.
1140 * This check needs to come after adding us to the btree_interior_update
1141 * list but before calling bch2_btree_reserve_get, to synchronize with
1142 * __bch2_fs_read_only().
1143 */
1144 ret = bch2_journal_error(&c->journal);
1145 if (ret)
1146 goto err;
1147
1148 ret = bch2_disk_reservation_get(c, &as->disk_res,
1149 (nr_nodes[0] + nr_nodes[1]) * btree_sectors(c),
1150 c->opts.metadata_replicas,
1151 disk_res_flags);
1152 if (ret)
1153 goto err;
1154
1155 ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, NULL);
1156 if (bch2_err_matches(ret, ENOSPC) ||
1157 bch2_err_matches(ret, ENOMEM)) {
1158 struct closure cl;
1159
1160 /*
1161 * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
1162 * flag
1163 */
1164 if (bch2_err_matches(ret, ENOSPC) &&
1165 (flags & BCH_TRANS_COMMIT_journal_reclaim) &&
1166 watermark != BCH_WATERMARK_reclaim) {
1167 ret = -BCH_ERR_journal_reclaim_would_deadlock;
1168 goto err;
1169 }
1170
1171 closure_init_stack(&cl);
1172
1173 do {
1174 ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, &cl);
1175
1176 bch2_trans_unlock(trans);
1177 closure_sync(&cl);
1178 } while (bch2_err_matches(ret, BCH_ERR_operation_blocked));
1179 }
1180
1181 if (ret) {
1182 trace_and_count(c, btree_reserve_get_fail, trans->fn,
1183 _RET_IP_, nr_nodes[0] + nr_nodes[1], ret);
1184 goto err;
1185 }
1186
1187 ret = bch2_trans_relock(trans);
1188 if (ret)
1189 goto err;
1190
1191 bch2_trans_verify_not_restarted(trans, restart_count);
1192 return as;
1193 err:
1194 bch2_btree_update_free(as, trans);
1195 if (!bch2_err_matches(ret, ENOSPC) &&
1196 !bch2_err_matches(ret, EROFS))
1197 bch_err_fn_ratelimited(c, ret);
1198 return ERR_PTR(ret);
1199 }
1200
1201 /* Btree root updates: */
1202
bch2_btree_set_root_inmem(struct bch_fs * c,struct btree * b)1203 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1204 {
1205 /* Root nodes cannot be reaped */
1206 mutex_lock(&c->btree_cache.lock);
1207 list_del_init(&b->list);
1208 mutex_unlock(&c->btree_cache.lock);
1209
1210 mutex_lock(&c->btree_root_lock);
1211 BUG_ON(btree_node_root(c, b) &&
1212 (b->c.level < btree_node_root(c, b)->c.level ||
1213 !btree_node_dying(btree_node_root(c, b))));
1214
1215 bch2_btree_id_root(c, b->c.btree_id)->b = b;
1216 mutex_unlock(&c->btree_root_lock);
1217
1218 bch2_recalc_btree_reserve(c);
1219 }
1220
bch2_btree_set_root(struct btree_update * as,struct btree_trans * trans,struct btree_path * path,struct btree * b)1221 static void bch2_btree_set_root(struct btree_update *as,
1222 struct btree_trans *trans,
1223 struct btree_path *path,
1224 struct btree *b)
1225 {
1226 struct bch_fs *c = as->c;
1227 struct btree *old;
1228
1229 trace_and_count(c, btree_node_set_root, trans, b);
1230
1231 old = btree_node_root(c, b);
1232
1233 /*
1234 * Ensure no one is using the old root while we switch to the
1235 * new root:
1236 */
1237 bch2_btree_node_lock_write_nofail(trans, path, &old->c);
1238
1239 bch2_btree_set_root_inmem(c, b);
1240
1241 btree_update_updated_root(as, b);
1242
1243 /*
1244 * Unlock old root after new root is visible:
1245 *
1246 * The new root isn't persistent, but that's ok: we still have
1247 * an intent lock on the new root, and any updates that would
1248 * depend on the new root would have to update the new root.
1249 */
1250 bch2_btree_node_unlock_write(trans, path, old);
1251 }
1252
1253 /* Interior node updates: */
1254
bch2_insert_fixup_btree_ptr(struct btree_update * as,struct btree_trans * trans,struct btree_path * path,struct btree * b,struct btree_node_iter * node_iter,struct bkey_i * insert)1255 static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
1256 struct btree_trans *trans,
1257 struct btree_path *path,
1258 struct btree *b,
1259 struct btree_node_iter *node_iter,
1260 struct bkey_i *insert)
1261 {
1262 struct bch_fs *c = as->c;
1263 struct bkey_packed *k;
1264 struct printbuf buf = PRINTBUF;
1265 unsigned long old, new, v;
1266
1267 BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1268 !btree_ptr_sectors_written(insert));
1269
1270 if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
1271 bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
1272
1273 if (bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1274 btree_node_type(b), WRITE, &buf) ?:
1275 bch2_bkey_in_btree_node(c, b, bkey_i_to_s_c(insert), &buf)) {
1276 printbuf_reset(&buf);
1277 prt_printf(&buf, "inserting invalid bkey\n ");
1278 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
1279 prt_printf(&buf, "\n ");
1280 bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1281 btree_node_type(b), WRITE, &buf);
1282 bch2_bkey_in_btree_node(c, b, bkey_i_to_s_c(insert), &buf);
1283
1284 bch2_fs_inconsistent(c, "%s", buf.buf);
1285 dump_stack();
1286 }
1287
1288 BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1289 ARRAY_SIZE(as->journal_entries));
1290
1291 as->journal_u64s +=
1292 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1293 BCH_JSET_ENTRY_btree_keys,
1294 b->c.btree_id, b->c.level,
1295 insert, insert->k.u64s);
1296
1297 while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1298 bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1299 bch2_btree_node_iter_advance(node_iter, b);
1300
1301 bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1302 set_btree_node_dirty_acct(c, b);
1303
1304 v = READ_ONCE(b->flags);
1305 do {
1306 old = new = v;
1307
1308 new &= ~BTREE_WRITE_TYPE_MASK;
1309 new |= BTREE_WRITE_interior;
1310 new |= 1 << BTREE_NODE_need_write;
1311 } while ((v = cmpxchg(&b->flags, old, new)) != old);
1312
1313 printbuf_exit(&buf);
1314 }
1315
1316 static void
__bch2_btree_insert_keys_interior(struct btree_update * as,struct btree_trans * trans,struct btree_path * path,struct btree * b,struct btree_node_iter node_iter,struct keylist * keys)1317 __bch2_btree_insert_keys_interior(struct btree_update *as,
1318 struct btree_trans *trans,
1319 struct btree_path *path,
1320 struct btree *b,
1321 struct btree_node_iter node_iter,
1322 struct keylist *keys)
1323 {
1324 struct bkey_i *insert = bch2_keylist_front(keys);
1325 struct bkey_packed *k;
1326
1327 BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1328
1329 while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1330 (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1331 ;
1332
1333 while (!bch2_keylist_empty(keys)) {
1334 insert = bch2_keylist_front(keys);
1335
1336 if (bpos_gt(insert->k.p, b->key.k.p))
1337 break;
1338
1339 bch2_insert_fixup_btree_ptr(as, trans, path, b, &node_iter, insert);
1340 bch2_keylist_pop_front(keys);
1341 }
1342 }
1343
1344 /*
1345 * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1346 * node)
1347 */
__btree_split_node(struct btree_update * as,struct btree_trans * trans,struct btree * b,struct btree * n[2])1348 static void __btree_split_node(struct btree_update *as,
1349 struct btree_trans *trans,
1350 struct btree *b,
1351 struct btree *n[2])
1352 {
1353 struct bkey_packed *k;
1354 struct bpos n1_pos = POS_MIN;
1355 struct btree_node_iter iter;
1356 struct bset *bsets[2];
1357 struct bkey_format_state format[2];
1358 struct bkey_packed *out[2];
1359 struct bkey uk;
1360 unsigned u64s, n1_u64s = (b->nr.live_u64s * 3) / 5;
1361 struct { unsigned nr_keys, val_u64s; } nr_keys[2];
1362 int i;
1363
1364 memset(&nr_keys, 0, sizeof(nr_keys));
1365
1366 for (i = 0; i < 2; i++) {
1367 BUG_ON(n[i]->nsets != 1);
1368
1369 bsets[i] = btree_bset_first(n[i]);
1370 out[i] = bsets[i]->start;
1371
1372 SET_BTREE_NODE_SEQ(n[i]->data, BTREE_NODE_SEQ(b->data) + 1);
1373 bch2_bkey_format_init(&format[i]);
1374 }
1375
1376 u64s = 0;
1377 for_each_btree_node_key(b, k, &iter) {
1378 if (bkey_deleted(k))
1379 continue;
1380
1381 i = u64s >= n1_u64s;
1382 u64s += k->u64s;
1383 uk = bkey_unpack_key(b, k);
1384 if (!i)
1385 n1_pos = uk.p;
1386 bch2_bkey_format_add_key(&format[i], &uk);
1387
1388 nr_keys[i].nr_keys++;
1389 nr_keys[i].val_u64s += bkeyp_val_u64s(&b->format, k);
1390 }
1391
1392 btree_set_min(n[0], b->data->min_key);
1393 btree_set_max(n[0], n1_pos);
1394 btree_set_min(n[1], bpos_successor(n1_pos));
1395 btree_set_max(n[1], b->data->max_key);
1396
1397 for (i = 0; i < 2; i++) {
1398 bch2_bkey_format_add_pos(&format[i], n[i]->data->min_key);
1399 bch2_bkey_format_add_pos(&format[i], n[i]->data->max_key);
1400
1401 n[i]->data->format = bch2_bkey_format_done(&format[i]);
1402
1403 unsigned u64s = nr_keys[i].nr_keys * n[i]->data->format.key_u64s +
1404 nr_keys[i].val_u64s;
1405 if (__vstruct_bytes(struct btree_node, u64s) > btree_buf_bytes(b))
1406 n[i]->data->format = b->format;
1407
1408 btree_node_set_format(n[i], n[i]->data->format);
1409 }
1410
1411 u64s = 0;
1412 for_each_btree_node_key(b, k, &iter) {
1413 if (bkey_deleted(k))
1414 continue;
1415
1416 i = u64s >= n1_u64s;
1417 u64s += k->u64s;
1418
1419 if (bch2_bkey_transform(&n[i]->format, out[i], bkey_packed(k)
1420 ? &b->format: &bch2_bkey_format_current, k))
1421 out[i]->format = KEY_FORMAT_LOCAL_BTREE;
1422 else
1423 bch2_bkey_unpack(b, (void *) out[i], k);
1424
1425 out[i]->needs_whiteout = false;
1426
1427 btree_keys_account_key_add(&n[i]->nr, 0, out[i]);
1428 out[i] = bkey_p_next(out[i]);
1429 }
1430
1431 for (i = 0; i < 2; i++) {
1432 bsets[i]->u64s = cpu_to_le16((u64 *) out[i] - bsets[i]->_data);
1433
1434 BUG_ON(!bsets[i]->u64s);
1435
1436 set_btree_bset_end(n[i], n[i]->set);
1437
1438 btree_node_reset_sib_u64s(n[i]);
1439
1440 bch2_verify_btree_nr_keys(n[i]);
1441
1442 if (b->c.level)
1443 btree_node_interior_verify(as->c, n[i]);
1444 }
1445 }
1446
1447 /*
1448 * For updates to interior nodes, we've got to do the insert before we split
1449 * because the stuff we're inserting has to be inserted atomically. Post split,
1450 * the keys might have to go in different nodes and the split would no longer be
1451 * atomic.
1452 *
1453 * Worse, if the insert is from btree node coalescing, if we do the insert after
1454 * we do the split (and pick the pivot) - the pivot we pick might be between
1455 * nodes that were coalesced, and thus in the middle of a child node post
1456 * coalescing:
1457 */
btree_split_insert_keys(struct btree_update * as,struct btree_trans * trans,btree_path_idx_t path_idx,struct btree * b,struct keylist * keys)1458 static void btree_split_insert_keys(struct btree_update *as,
1459 struct btree_trans *trans,
1460 btree_path_idx_t path_idx,
1461 struct btree *b,
1462 struct keylist *keys)
1463 {
1464 struct btree_path *path = trans->paths + path_idx;
1465
1466 if (!bch2_keylist_empty(keys) &&
1467 bpos_le(bch2_keylist_front(keys)->k.p, b->data->max_key)) {
1468 struct btree_node_iter node_iter;
1469
1470 bch2_btree_node_iter_init(&node_iter, b, &bch2_keylist_front(keys)->k.p);
1471
1472 __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1473
1474 btree_node_interior_verify(as->c, b);
1475 }
1476 }
1477
btree_split(struct btree_update * as,struct btree_trans * trans,btree_path_idx_t path,struct btree * b,struct keylist * keys,unsigned flags)1478 static int btree_split(struct btree_update *as, struct btree_trans *trans,
1479 btree_path_idx_t path, struct btree *b,
1480 struct keylist *keys, unsigned flags)
1481 {
1482 struct bch_fs *c = as->c;
1483 struct btree *parent = btree_node_parent(trans->paths + path, b);
1484 struct btree *n1, *n2 = NULL, *n3 = NULL;
1485 btree_path_idx_t path1 = 0, path2 = 0;
1486 u64 start_time = local_clock();
1487 int ret = 0;
1488
1489 BUG_ON(!parent && (b != btree_node_root(c, b)));
1490 BUG_ON(parent && !btree_node_intent_locked(trans->paths + path, b->c.level + 1));
1491
1492 bch2_btree_interior_update_will_free_node(as, b);
1493
1494 if (b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c)) {
1495 struct btree *n[2];
1496
1497 trace_and_count(c, btree_node_split, trans, b);
1498
1499 n[0] = n1 = bch2_btree_node_alloc(as, trans, b->c.level);
1500 n[1] = n2 = bch2_btree_node_alloc(as, trans, b->c.level);
1501
1502 __btree_split_node(as, trans, b, n);
1503
1504 if (keys) {
1505 btree_split_insert_keys(as, trans, path, n1, keys);
1506 btree_split_insert_keys(as, trans, path, n2, keys);
1507 BUG_ON(!bch2_keylist_empty(keys));
1508 }
1509
1510 bch2_btree_build_aux_trees(n2);
1511 bch2_btree_build_aux_trees(n1);
1512
1513 bch2_btree_update_add_new_node(as, n1);
1514 bch2_btree_update_add_new_node(as, n2);
1515 six_unlock_write(&n2->c.lock);
1516 six_unlock_write(&n1->c.lock);
1517
1518 path1 = get_unlocked_mut_path(trans, as->btree_id, n1->c.level, n1->key.k.p);
1519 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1520 mark_btree_node_locked(trans, trans->paths + path1, n1->c.level, BTREE_NODE_INTENT_LOCKED);
1521 bch2_btree_path_level_init(trans, trans->paths + path1, n1);
1522
1523 path2 = get_unlocked_mut_path(trans, as->btree_id, n2->c.level, n2->key.k.p);
1524 six_lock_increment(&n2->c.lock, SIX_LOCK_intent);
1525 mark_btree_node_locked(trans, trans->paths + path2, n2->c.level, BTREE_NODE_INTENT_LOCKED);
1526 bch2_btree_path_level_init(trans, trans->paths + path2, n2);
1527
1528 /*
1529 * Note that on recursive parent_keys == keys, so we
1530 * can't start adding new keys to parent_keys before emptying it
1531 * out (which we did with btree_split_insert_keys() above)
1532 */
1533 bch2_keylist_add(&as->parent_keys, &n1->key);
1534 bch2_keylist_add(&as->parent_keys, &n2->key);
1535
1536 if (!parent) {
1537 /* Depth increases, make a new root */
1538 n3 = __btree_root_alloc(as, trans, b->c.level + 1);
1539
1540 bch2_btree_update_add_new_node(as, n3);
1541 six_unlock_write(&n3->c.lock);
1542
1543 trans->paths[path2].locks_want++;
1544 BUG_ON(btree_node_locked(trans->paths + path2, n3->c.level));
1545 six_lock_increment(&n3->c.lock, SIX_LOCK_intent);
1546 mark_btree_node_locked(trans, trans->paths + path2, n3->c.level, BTREE_NODE_INTENT_LOCKED);
1547 bch2_btree_path_level_init(trans, trans->paths + path2, n3);
1548
1549 n3->sib_u64s[0] = U16_MAX;
1550 n3->sib_u64s[1] = U16_MAX;
1551
1552 btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1553 }
1554 } else {
1555 trace_and_count(c, btree_node_compact, trans, b);
1556
1557 n1 = bch2_btree_node_alloc_replacement(as, trans, b);
1558
1559 if (keys) {
1560 btree_split_insert_keys(as, trans, path, n1, keys);
1561 BUG_ON(!bch2_keylist_empty(keys));
1562 }
1563
1564 bch2_btree_build_aux_trees(n1);
1565 bch2_btree_update_add_new_node(as, n1);
1566 six_unlock_write(&n1->c.lock);
1567
1568 path1 = get_unlocked_mut_path(trans, as->btree_id, n1->c.level, n1->key.k.p);
1569 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1570 mark_btree_node_locked(trans, trans->paths + path1, n1->c.level, BTREE_NODE_INTENT_LOCKED);
1571 bch2_btree_path_level_init(trans, trans->paths + path1, n1);
1572
1573 if (parent)
1574 bch2_keylist_add(&as->parent_keys, &n1->key);
1575 }
1576
1577 /* New nodes all written, now make them visible: */
1578
1579 if (parent) {
1580 /* Split a non root node */
1581 ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1582 if (ret)
1583 goto err;
1584 } else if (n3) {
1585 bch2_btree_set_root(as, trans, trans->paths + path, n3);
1586 } else {
1587 /* Root filled up but didn't need to be split */
1588 bch2_btree_set_root(as, trans, trans->paths + path, n1);
1589 }
1590
1591 if (n3) {
1592 bch2_btree_update_get_open_buckets(as, n3);
1593 bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
1594 }
1595 if (n2) {
1596 bch2_btree_update_get_open_buckets(as, n2);
1597 bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
1598 }
1599 bch2_btree_update_get_open_buckets(as, n1);
1600 bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1601
1602 /*
1603 * The old node must be freed (in memory) _before_ unlocking the new
1604 * nodes - else another thread could re-acquire a read lock on the old
1605 * node after another thread has locked and updated the new node, thus
1606 * seeing stale data:
1607 */
1608 bch2_btree_node_free_inmem(trans, trans->paths + path, b);
1609
1610 if (n3)
1611 bch2_trans_node_add(trans, trans->paths + path, n3);
1612 if (n2)
1613 bch2_trans_node_add(trans, trans->paths + path2, n2);
1614 bch2_trans_node_add(trans, trans->paths + path1, n1);
1615
1616 if (n3)
1617 six_unlock_intent(&n3->c.lock);
1618 if (n2)
1619 six_unlock_intent(&n2->c.lock);
1620 six_unlock_intent(&n1->c.lock);
1621 out:
1622 if (path2) {
1623 __bch2_btree_path_unlock(trans, trans->paths + path2);
1624 bch2_path_put(trans, path2, true);
1625 }
1626 if (path1) {
1627 __bch2_btree_path_unlock(trans, trans->paths + path1);
1628 bch2_path_put(trans, path1, true);
1629 }
1630
1631 bch2_trans_verify_locks(trans);
1632
1633 bch2_time_stats_update(&c->times[n2
1634 ? BCH_TIME_btree_node_split
1635 : BCH_TIME_btree_node_compact],
1636 start_time);
1637 return ret;
1638 err:
1639 if (n3)
1640 bch2_btree_node_free_never_used(as, trans, n3);
1641 if (n2)
1642 bch2_btree_node_free_never_used(as, trans, n2);
1643 bch2_btree_node_free_never_used(as, trans, n1);
1644 goto out;
1645 }
1646
1647 static void
bch2_btree_insert_keys_interior(struct btree_update * as,struct btree_trans * trans,struct btree_path * path,struct btree * b,struct keylist * keys)1648 bch2_btree_insert_keys_interior(struct btree_update *as,
1649 struct btree_trans *trans,
1650 struct btree_path *path,
1651 struct btree *b,
1652 struct keylist *keys)
1653 {
1654 struct btree_path *linked;
1655 unsigned i;
1656
1657 __bch2_btree_insert_keys_interior(as, trans, path, b,
1658 path->l[b->c.level].iter, keys);
1659
1660 btree_update_updated_node(as, b);
1661
1662 trans_for_each_path_with_node(trans, b, linked, i)
1663 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1664
1665 bch2_trans_verify_paths(trans);
1666 }
1667
1668 /**
1669 * bch2_btree_insert_node - insert bkeys into a given btree node
1670 *
1671 * @as: btree_update object
1672 * @trans: btree_trans object
1673 * @path_idx: path that points to current node
1674 * @b: node to insert keys into
1675 * @keys: list of keys to insert
1676 * @flags: transaction commit flags
1677 *
1678 * Returns: 0 on success, typically transaction restart error on failure
1679 *
1680 * Inserts as many keys as it can into a given btree node, splitting it if full.
1681 * If a split occurred, this function will return early. This can only happen
1682 * for leaf nodes -- inserts into interior nodes have to be atomic.
1683 */
bch2_btree_insert_node(struct btree_update * as,struct btree_trans * trans,btree_path_idx_t path_idx,struct btree * b,struct keylist * keys,unsigned flags)1684 static int bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1685 btree_path_idx_t path_idx, struct btree *b,
1686 struct keylist *keys, unsigned flags)
1687 {
1688 struct bch_fs *c = as->c;
1689 struct btree_path *path = trans->paths + path_idx;
1690 int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1691 int old_live_u64s = b->nr.live_u64s;
1692 int live_u64s_added, u64s_added;
1693 int ret;
1694
1695 lockdep_assert_held(&c->gc_lock);
1696 BUG_ON(!btree_node_intent_locked(path, b->c.level));
1697 BUG_ON(!b->c.level);
1698 BUG_ON(!as || as->b);
1699 bch2_verify_keylist_sorted(keys);
1700
1701 ret = bch2_btree_node_lock_write(trans, path, &b->c);
1702 if (ret)
1703 return ret;
1704
1705 bch2_btree_node_prep_for_write(trans, path, b);
1706
1707 if (!bch2_btree_node_insert_fits(b, bch2_keylist_u64s(keys))) {
1708 bch2_btree_node_unlock_write(trans, path, b);
1709 goto split;
1710 }
1711
1712 btree_node_interior_verify(c, b);
1713
1714 bch2_btree_insert_keys_interior(as, trans, path, b, keys);
1715
1716 live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1717 u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1718
1719 if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1720 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1721 if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1722 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1723
1724 if (u64s_added > live_u64s_added &&
1725 bch2_maybe_compact_whiteouts(c, b))
1726 bch2_trans_node_reinit_iter(trans, b);
1727
1728 bch2_btree_node_unlock_write(trans, path, b);
1729
1730 btree_node_interior_verify(c, b);
1731 return 0;
1732 split:
1733 /*
1734 * We could attempt to avoid the transaction restart, by calling
1735 * bch2_btree_path_upgrade() and allocating more nodes:
1736 */
1737 if (b->c.level >= as->update_level) {
1738 trace_and_count(c, trans_restart_split_race, trans, _THIS_IP_, b);
1739 return btree_trans_restart(trans, BCH_ERR_transaction_restart_split_race);
1740 }
1741
1742 return btree_split(as, trans, path_idx, b, keys, flags);
1743 }
1744
bch2_btree_split_leaf(struct btree_trans * trans,btree_path_idx_t path,unsigned flags)1745 int bch2_btree_split_leaf(struct btree_trans *trans,
1746 btree_path_idx_t path,
1747 unsigned flags)
1748 {
1749 /* btree_split & merge may both cause paths array to be reallocated */
1750
1751 struct btree *b = path_l(trans->paths + path)->b;
1752 struct btree_update *as;
1753 unsigned l;
1754 int ret = 0;
1755
1756 as = bch2_btree_update_start(trans, trans->paths + path,
1757 trans->paths[path].level,
1758 true, flags);
1759 if (IS_ERR(as))
1760 return PTR_ERR(as);
1761
1762 ret = btree_split(as, trans, path, b, NULL, flags);
1763 if (ret) {
1764 bch2_btree_update_free(as, trans);
1765 return ret;
1766 }
1767
1768 bch2_btree_update_done(as, trans);
1769
1770 for (l = trans->paths[path].level + 1;
1771 btree_node_intent_locked(&trans->paths[path], l) && !ret;
1772 l++)
1773 ret = bch2_foreground_maybe_merge(trans, path, l, flags);
1774
1775 return ret;
1776 }
1777
__bch2_foreground_maybe_merge(struct btree_trans * trans,btree_path_idx_t path,unsigned level,unsigned flags,enum btree_node_sibling sib)1778 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1779 btree_path_idx_t path,
1780 unsigned level,
1781 unsigned flags,
1782 enum btree_node_sibling sib)
1783 {
1784 struct bch_fs *c = trans->c;
1785 struct btree_update *as;
1786 struct bkey_format_state new_s;
1787 struct bkey_format new_f;
1788 struct bkey_i delete;
1789 struct btree *b, *m, *n, *prev, *next, *parent;
1790 struct bpos sib_pos;
1791 size_t sib_u64s;
1792 enum btree_id btree = trans->paths[path].btree_id;
1793 btree_path_idx_t sib_path = 0, new_path = 0;
1794 u64 start_time = local_clock();
1795 int ret = 0;
1796
1797 BUG_ON(!trans->paths[path].should_be_locked);
1798 BUG_ON(!btree_node_locked(&trans->paths[path], level));
1799
1800 b = trans->paths[path].l[level].b;
1801
1802 if ((sib == btree_prev_sib && bpos_eq(b->data->min_key, POS_MIN)) ||
1803 (sib == btree_next_sib && bpos_eq(b->data->max_key, SPOS_MAX))) {
1804 b->sib_u64s[sib] = U16_MAX;
1805 return 0;
1806 }
1807
1808 sib_pos = sib == btree_prev_sib
1809 ? bpos_predecessor(b->data->min_key)
1810 : bpos_successor(b->data->max_key);
1811
1812 sib_path = bch2_path_get(trans, btree, sib_pos,
1813 U8_MAX, level, BTREE_ITER_INTENT, _THIS_IP_);
1814 ret = bch2_btree_path_traverse(trans, sib_path, false);
1815 if (ret)
1816 goto err;
1817
1818 btree_path_set_should_be_locked(trans->paths + sib_path);
1819
1820 m = trans->paths[sib_path].l[level].b;
1821
1822 if (btree_node_parent(trans->paths + path, b) !=
1823 btree_node_parent(trans->paths + sib_path, m)) {
1824 b->sib_u64s[sib] = U16_MAX;
1825 goto out;
1826 }
1827
1828 if (sib == btree_prev_sib) {
1829 prev = m;
1830 next = b;
1831 } else {
1832 prev = b;
1833 next = m;
1834 }
1835
1836 if (!bpos_eq(bpos_successor(prev->data->max_key), next->data->min_key)) {
1837 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1838
1839 bch2_bpos_to_text(&buf1, prev->data->max_key);
1840 bch2_bpos_to_text(&buf2, next->data->min_key);
1841 bch_err(c,
1842 "%s(): btree topology error:\n"
1843 " prev ends at %s\n"
1844 " next starts at %s",
1845 __func__, buf1.buf, buf2.buf);
1846 printbuf_exit(&buf1);
1847 printbuf_exit(&buf2);
1848 bch2_topology_error(c);
1849 ret = -EIO;
1850 goto err;
1851 }
1852
1853 bch2_bkey_format_init(&new_s);
1854 bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1855 __bch2_btree_calc_format(&new_s, prev);
1856 __bch2_btree_calc_format(&new_s, next);
1857 bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1858 new_f = bch2_bkey_format_done(&new_s);
1859
1860 sib_u64s = btree_node_u64s_with_format(b->nr, &b->format, &new_f) +
1861 btree_node_u64s_with_format(m->nr, &m->format, &new_f);
1862
1863 if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1864 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1865 sib_u64s /= 2;
1866 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1867 }
1868
1869 sib_u64s = min(sib_u64s, btree_max_u64s(c));
1870 sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1871 b->sib_u64s[sib] = sib_u64s;
1872
1873 if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1874 goto out;
1875
1876 parent = btree_node_parent(trans->paths + path, b);
1877 as = bch2_btree_update_start(trans, trans->paths + path, level, false,
1878 BCH_TRANS_COMMIT_no_enospc|flags);
1879 ret = PTR_ERR_OR_ZERO(as);
1880 if (ret)
1881 goto err;
1882
1883 trace_and_count(c, btree_node_merge, trans, b);
1884
1885 bch2_btree_interior_update_will_free_node(as, b);
1886 bch2_btree_interior_update_will_free_node(as, m);
1887
1888 n = bch2_btree_node_alloc(as, trans, b->c.level);
1889
1890 SET_BTREE_NODE_SEQ(n->data,
1891 max(BTREE_NODE_SEQ(b->data),
1892 BTREE_NODE_SEQ(m->data)) + 1);
1893
1894 btree_set_min(n, prev->data->min_key);
1895 btree_set_max(n, next->data->max_key);
1896
1897 n->data->format = new_f;
1898 btree_node_set_format(n, new_f);
1899
1900 bch2_btree_sort_into(c, n, prev);
1901 bch2_btree_sort_into(c, n, next);
1902
1903 bch2_btree_build_aux_trees(n);
1904 bch2_btree_update_add_new_node(as, n);
1905 six_unlock_write(&n->c.lock);
1906
1907 new_path = get_unlocked_mut_path(trans, btree, n->c.level, n->key.k.p);
1908 six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1909 mark_btree_node_locked(trans, trans->paths + new_path, n->c.level, BTREE_NODE_INTENT_LOCKED);
1910 bch2_btree_path_level_init(trans, trans->paths + new_path, n);
1911
1912 bkey_init(&delete.k);
1913 delete.k.p = prev->key.k.p;
1914 bch2_keylist_add(&as->parent_keys, &delete);
1915 bch2_keylist_add(&as->parent_keys, &n->key);
1916
1917 bch2_trans_verify_paths(trans);
1918
1919 ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1920 if (ret)
1921 goto err_free_update;
1922
1923 bch2_trans_verify_paths(trans);
1924
1925 bch2_btree_update_get_open_buckets(as, n);
1926 bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1927
1928 bch2_btree_node_free_inmem(trans, trans->paths + path, b);
1929 bch2_btree_node_free_inmem(trans, trans->paths + sib_path, m);
1930
1931 bch2_trans_node_add(trans, trans->paths + path, n);
1932
1933 bch2_trans_verify_paths(trans);
1934
1935 six_unlock_intent(&n->c.lock);
1936
1937 bch2_btree_update_done(as, trans);
1938
1939 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
1940 out:
1941 err:
1942 if (new_path)
1943 bch2_path_put(trans, new_path, true);
1944 bch2_path_put(trans, sib_path, true);
1945 bch2_trans_verify_locks(trans);
1946 return ret;
1947 err_free_update:
1948 bch2_btree_node_free_never_used(as, trans, n);
1949 bch2_btree_update_free(as, trans);
1950 goto out;
1951 }
1952
bch2_btree_node_rewrite(struct btree_trans * trans,struct btree_iter * iter,struct btree * b,unsigned flags)1953 int bch2_btree_node_rewrite(struct btree_trans *trans,
1954 struct btree_iter *iter,
1955 struct btree *b,
1956 unsigned flags)
1957 {
1958 struct bch_fs *c = trans->c;
1959 struct btree *n, *parent;
1960 struct btree_update *as;
1961 btree_path_idx_t new_path = 0;
1962 int ret;
1963
1964 flags |= BCH_TRANS_COMMIT_no_enospc;
1965
1966 struct btree_path *path = btree_iter_path(trans, iter);
1967 parent = btree_node_parent(path, b);
1968 as = bch2_btree_update_start(trans, path, b->c.level, false, flags);
1969 ret = PTR_ERR_OR_ZERO(as);
1970 if (ret)
1971 goto out;
1972
1973 bch2_btree_interior_update_will_free_node(as, b);
1974
1975 n = bch2_btree_node_alloc_replacement(as, trans, b);
1976
1977 bch2_btree_build_aux_trees(n);
1978 bch2_btree_update_add_new_node(as, n);
1979 six_unlock_write(&n->c.lock);
1980
1981 new_path = get_unlocked_mut_path(trans, iter->btree_id, n->c.level, n->key.k.p);
1982 six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1983 mark_btree_node_locked(trans, trans->paths + new_path, n->c.level, BTREE_NODE_INTENT_LOCKED);
1984 bch2_btree_path_level_init(trans, trans->paths + new_path, n);
1985
1986 trace_and_count(c, btree_node_rewrite, trans, b);
1987
1988 if (parent) {
1989 bch2_keylist_add(&as->parent_keys, &n->key);
1990 ret = bch2_btree_insert_node(as, trans, iter->path,
1991 parent, &as->parent_keys, flags);
1992 if (ret)
1993 goto err;
1994 } else {
1995 bch2_btree_set_root(as, trans, btree_iter_path(trans, iter), n);
1996 }
1997
1998 bch2_btree_update_get_open_buckets(as, n);
1999 bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
2000
2001 bch2_btree_node_free_inmem(trans, btree_iter_path(trans, iter), b);
2002
2003 bch2_trans_node_add(trans, trans->paths + iter->path, n);
2004 six_unlock_intent(&n->c.lock);
2005
2006 bch2_btree_update_done(as, trans);
2007 out:
2008 if (new_path)
2009 bch2_path_put(trans, new_path, true);
2010 bch2_trans_downgrade(trans);
2011 return ret;
2012 err:
2013 bch2_btree_node_free_never_used(as, trans, n);
2014 bch2_btree_update_free(as, trans);
2015 goto out;
2016 }
2017
2018 struct async_btree_rewrite {
2019 struct bch_fs *c;
2020 struct work_struct work;
2021 struct list_head list;
2022 enum btree_id btree_id;
2023 unsigned level;
2024 struct bpos pos;
2025 __le64 seq;
2026 };
2027
async_btree_node_rewrite_trans(struct btree_trans * trans,struct async_btree_rewrite * a)2028 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
2029 struct async_btree_rewrite *a)
2030 {
2031 struct bch_fs *c = trans->c;
2032 struct btree_iter iter;
2033 struct btree *b;
2034 int ret;
2035
2036 bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
2037 BTREE_MAX_DEPTH, a->level, 0);
2038 b = bch2_btree_iter_peek_node(&iter);
2039 ret = PTR_ERR_OR_ZERO(b);
2040 if (ret)
2041 goto out;
2042
2043 if (!b || b->data->keys.seq != a->seq) {
2044 struct printbuf buf = PRINTBUF;
2045
2046 if (b)
2047 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
2048 else
2049 prt_str(&buf, "(null");
2050 bch_info(c, "%s: node to rewrite not found:, searching for seq %llu, got\n%s",
2051 __func__, a->seq, buf.buf);
2052 printbuf_exit(&buf);
2053 goto out;
2054 }
2055
2056 ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
2057 out:
2058 bch2_trans_iter_exit(trans, &iter);
2059
2060 return ret;
2061 }
2062
async_btree_node_rewrite_work(struct work_struct * work)2063 static void async_btree_node_rewrite_work(struct work_struct *work)
2064 {
2065 struct async_btree_rewrite *a =
2066 container_of(work, struct async_btree_rewrite, work);
2067 struct bch_fs *c = a->c;
2068 int ret;
2069
2070 ret = bch2_trans_do(c, NULL, NULL, 0,
2071 async_btree_node_rewrite_trans(trans, a));
2072 bch_err_fn(c, ret);
2073 bch2_write_ref_put(c, BCH_WRITE_REF_node_rewrite);
2074 kfree(a);
2075 }
2076
bch2_btree_node_rewrite_async(struct bch_fs * c,struct btree * b)2077 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
2078 {
2079 struct async_btree_rewrite *a;
2080 int ret;
2081
2082 a = kmalloc(sizeof(*a), GFP_NOFS);
2083 if (!a) {
2084 bch_err(c, "%s: error allocating memory", __func__);
2085 return;
2086 }
2087
2088 a->c = c;
2089 a->btree_id = b->c.btree_id;
2090 a->level = b->c.level;
2091 a->pos = b->key.k.p;
2092 a->seq = b->data->keys.seq;
2093 INIT_WORK(&a->work, async_btree_node_rewrite_work);
2094
2095 if (unlikely(!test_bit(BCH_FS_may_go_rw, &c->flags))) {
2096 mutex_lock(&c->pending_node_rewrites_lock);
2097 list_add(&a->list, &c->pending_node_rewrites);
2098 mutex_unlock(&c->pending_node_rewrites_lock);
2099 return;
2100 }
2101
2102 if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_node_rewrite)) {
2103 if (test_bit(BCH_FS_started, &c->flags)) {
2104 bch_err(c, "%s: error getting c->writes ref", __func__);
2105 kfree(a);
2106 return;
2107 }
2108
2109 ret = bch2_fs_read_write_early(c);
2110 bch_err_msg(c, ret, "going read-write");
2111 if (ret) {
2112 kfree(a);
2113 return;
2114 }
2115
2116 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2117 }
2118
2119 queue_work(c->btree_interior_update_worker, &a->work);
2120 }
2121
bch2_do_pending_node_rewrites(struct bch_fs * c)2122 void bch2_do_pending_node_rewrites(struct bch_fs *c)
2123 {
2124 struct async_btree_rewrite *a, *n;
2125
2126 mutex_lock(&c->pending_node_rewrites_lock);
2127 list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2128 list_del(&a->list);
2129
2130 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2131 queue_work(c->btree_interior_update_worker, &a->work);
2132 }
2133 mutex_unlock(&c->pending_node_rewrites_lock);
2134 }
2135
bch2_free_pending_node_rewrites(struct bch_fs * c)2136 void bch2_free_pending_node_rewrites(struct bch_fs *c)
2137 {
2138 struct async_btree_rewrite *a, *n;
2139
2140 mutex_lock(&c->pending_node_rewrites_lock);
2141 list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2142 list_del(&a->list);
2143
2144 kfree(a);
2145 }
2146 mutex_unlock(&c->pending_node_rewrites_lock);
2147 }
2148
__bch2_btree_node_update_key(struct btree_trans * trans,struct btree_iter * iter,struct btree * b,struct btree * new_hash,struct bkey_i * new_key,unsigned commit_flags,bool skip_triggers)2149 static int __bch2_btree_node_update_key(struct btree_trans *trans,
2150 struct btree_iter *iter,
2151 struct btree *b, struct btree *new_hash,
2152 struct bkey_i *new_key,
2153 unsigned commit_flags,
2154 bool skip_triggers)
2155 {
2156 struct bch_fs *c = trans->c;
2157 struct btree_iter iter2 = { NULL };
2158 struct btree *parent;
2159 int ret;
2160
2161 if (!skip_triggers) {
2162 ret = bch2_key_trigger_old(trans, b->c.btree_id, b->c.level + 1,
2163 bkey_i_to_s_c(&b->key),
2164 BTREE_TRIGGER_TRANSACTIONAL) ?:
2165 bch2_key_trigger_new(trans, b->c.btree_id, b->c.level + 1,
2166 bkey_i_to_s(new_key),
2167 BTREE_TRIGGER_TRANSACTIONAL);
2168 if (ret)
2169 return ret;
2170 }
2171
2172 if (new_hash) {
2173 bkey_copy(&new_hash->key, new_key);
2174 ret = bch2_btree_node_hash_insert(&c->btree_cache,
2175 new_hash, b->c.level, b->c.btree_id);
2176 BUG_ON(ret);
2177 }
2178
2179 parent = btree_node_parent(btree_iter_path(trans, iter), b);
2180 if (parent) {
2181 bch2_trans_copy_iter(&iter2, iter);
2182
2183 iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
2184 iter2.flags & BTREE_ITER_INTENT,
2185 _THIS_IP_);
2186
2187 struct btree_path *path2 = btree_iter_path(trans, &iter2);
2188 BUG_ON(path2->level != b->c.level);
2189 BUG_ON(!bpos_eq(path2->pos, new_key->k.p));
2190
2191 btree_path_set_level_up(trans, path2);
2192
2193 trans->paths_sorted = false;
2194
2195 ret = bch2_btree_iter_traverse(&iter2) ?:
2196 bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
2197 if (ret)
2198 goto err;
2199 } else {
2200 BUG_ON(btree_node_root(c, b) != b);
2201
2202 struct jset_entry *e = bch2_trans_jset_entry_alloc(trans,
2203 jset_u64s(new_key->k.u64s));
2204 ret = PTR_ERR_OR_ZERO(e);
2205 if (ret)
2206 return ret;
2207
2208 journal_entry_set(e,
2209 BCH_JSET_ENTRY_btree_root,
2210 b->c.btree_id, b->c.level,
2211 new_key, new_key->k.u64s);
2212 }
2213
2214 ret = bch2_trans_commit(trans, NULL, NULL, commit_flags);
2215 if (ret)
2216 goto err;
2217
2218 bch2_btree_node_lock_write_nofail(trans, btree_iter_path(trans, iter), &b->c);
2219
2220 if (new_hash) {
2221 mutex_lock(&c->btree_cache.lock);
2222 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
2223 bch2_btree_node_hash_remove(&c->btree_cache, b);
2224
2225 bkey_copy(&b->key, new_key);
2226 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
2227 BUG_ON(ret);
2228 mutex_unlock(&c->btree_cache.lock);
2229 } else {
2230 bkey_copy(&b->key, new_key);
2231 }
2232
2233 bch2_btree_node_unlock_write(trans, btree_iter_path(trans, iter), b);
2234 out:
2235 bch2_trans_iter_exit(trans, &iter2);
2236 return ret;
2237 err:
2238 if (new_hash) {
2239 mutex_lock(&c->btree_cache.lock);
2240 bch2_btree_node_hash_remove(&c->btree_cache, b);
2241 mutex_unlock(&c->btree_cache.lock);
2242 }
2243 goto out;
2244 }
2245
bch2_btree_node_update_key(struct btree_trans * trans,struct btree_iter * iter,struct btree * b,struct bkey_i * new_key,unsigned commit_flags,bool skip_triggers)2246 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
2247 struct btree *b, struct bkey_i *new_key,
2248 unsigned commit_flags, bool skip_triggers)
2249 {
2250 struct bch_fs *c = trans->c;
2251 struct btree *new_hash = NULL;
2252 struct btree_path *path = btree_iter_path(trans, iter);
2253 struct closure cl;
2254 int ret = 0;
2255
2256 ret = bch2_btree_path_upgrade(trans, path, b->c.level + 1);
2257 if (ret)
2258 return ret;
2259
2260 closure_init_stack(&cl);
2261
2262 /*
2263 * check btree_ptr_hash_val() after @b is locked by
2264 * btree_iter_traverse():
2265 */
2266 if (btree_ptr_hash_val(new_key) != b->hash_val) {
2267 ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
2268 if (ret) {
2269 ret = drop_locks_do(trans, (closure_sync(&cl), 0));
2270 if (ret)
2271 return ret;
2272 }
2273
2274 new_hash = bch2_btree_node_mem_alloc(trans, false);
2275 }
2276
2277 path->intent_ref++;
2278 ret = __bch2_btree_node_update_key(trans, iter, b, new_hash, new_key,
2279 commit_flags, skip_triggers);
2280 --path->intent_ref;
2281
2282 if (new_hash) {
2283 mutex_lock(&c->btree_cache.lock);
2284 list_move(&new_hash->list, &c->btree_cache.freeable);
2285 mutex_unlock(&c->btree_cache.lock);
2286
2287 six_unlock_write(&new_hash->c.lock);
2288 six_unlock_intent(&new_hash->c.lock);
2289 }
2290 closure_sync(&cl);
2291 bch2_btree_cache_cannibalize_unlock(trans);
2292 return ret;
2293 }
2294
bch2_btree_node_update_key_get_iter(struct btree_trans * trans,struct btree * b,struct bkey_i * new_key,unsigned commit_flags,bool skip_triggers)2295 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2296 struct btree *b, struct bkey_i *new_key,
2297 unsigned commit_flags, bool skip_triggers)
2298 {
2299 struct btree_iter iter;
2300 int ret;
2301
2302 bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2303 BTREE_MAX_DEPTH, b->c.level,
2304 BTREE_ITER_INTENT);
2305 ret = bch2_btree_iter_traverse(&iter);
2306 if (ret)
2307 goto out;
2308
2309 /* has node been freed? */
2310 if (btree_iter_path(trans, &iter)->l[b->c.level].b != b) {
2311 /* node has been freed: */
2312 BUG_ON(!btree_node_dying(b));
2313 goto out;
2314 }
2315
2316 BUG_ON(!btree_node_hashed(b));
2317
2318 struct bch_extent_ptr *ptr;
2319 bch2_bkey_drop_ptrs(bkey_i_to_s(new_key), ptr,
2320 !bch2_bkey_has_device(bkey_i_to_s(&b->key), ptr->dev));
2321
2322 ret = bch2_btree_node_update_key(trans, &iter, b, new_key,
2323 commit_flags, skip_triggers);
2324 out:
2325 bch2_trans_iter_exit(trans, &iter);
2326 return ret;
2327 }
2328
2329 /* Init code: */
2330
2331 /*
2332 * Only for filesystem bringup, when first reading the btree roots or allocating
2333 * btree roots when initializing a new filesystem:
2334 */
bch2_btree_set_root_for_read(struct bch_fs * c,struct btree * b)2335 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2336 {
2337 BUG_ON(btree_node_root(c, b));
2338
2339 bch2_btree_set_root_inmem(c, b);
2340 }
2341
__bch2_btree_root_alloc(struct btree_trans * trans,enum btree_id id)2342 static int __bch2_btree_root_alloc(struct btree_trans *trans, enum btree_id id)
2343 {
2344 struct bch_fs *c = trans->c;
2345 struct closure cl;
2346 struct btree *b;
2347 int ret;
2348
2349 closure_init_stack(&cl);
2350
2351 do {
2352 ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
2353 closure_sync(&cl);
2354 } while (ret);
2355
2356 b = bch2_btree_node_mem_alloc(trans, false);
2357 bch2_btree_cache_cannibalize_unlock(trans);
2358
2359 set_btree_node_fake(b);
2360 set_btree_node_need_rewrite(b);
2361 b->c.level = 0;
2362 b->c.btree_id = id;
2363
2364 bkey_btree_ptr_init(&b->key);
2365 b->key.k.p = SPOS_MAX;
2366 *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2367
2368 bch2_bset_init_first(b, &b->data->keys);
2369 bch2_btree_build_aux_trees(b);
2370
2371 b->data->flags = 0;
2372 btree_set_min(b, POS_MIN);
2373 btree_set_max(b, SPOS_MAX);
2374 b->data->format = bch2_btree_calc_format(b);
2375 btree_node_set_format(b, b->data->format);
2376
2377 ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2378 b->c.level, b->c.btree_id);
2379 BUG_ON(ret);
2380
2381 bch2_btree_set_root_inmem(c, b);
2382
2383 six_unlock_write(&b->c.lock);
2384 six_unlock_intent(&b->c.lock);
2385 return 0;
2386 }
2387
bch2_btree_root_alloc(struct bch_fs * c,enum btree_id id)2388 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2389 {
2390 bch2_trans_run(c, __bch2_btree_root_alloc(trans, id));
2391 }
2392
bch2_btree_updates_to_text(struct printbuf * out,struct bch_fs * c)2393 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2394 {
2395 struct btree_update *as;
2396
2397 mutex_lock(&c->btree_interior_update_lock);
2398 list_for_each_entry(as, &c->btree_interior_update_list, list)
2399 prt_printf(out, "%p m %u w %u r %u j %llu\n",
2400 as,
2401 as->mode,
2402 as->nodes_written,
2403 closure_nr_remaining(&as->cl),
2404 as->journal.seq);
2405 mutex_unlock(&c->btree_interior_update_lock);
2406 }
2407
bch2_btree_interior_updates_pending(struct bch_fs * c)2408 static bool bch2_btree_interior_updates_pending(struct bch_fs *c)
2409 {
2410 bool ret;
2411
2412 mutex_lock(&c->btree_interior_update_lock);
2413 ret = !list_empty(&c->btree_interior_update_list);
2414 mutex_unlock(&c->btree_interior_update_lock);
2415
2416 return ret;
2417 }
2418
bch2_btree_interior_updates_flush(struct bch_fs * c)2419 bool bch2_btree_interior_updates_flush(struct bch_fs *c)
2420 {
2421 bool ret = bch2_btree_interior_updates_pending(c);
2422
2423 if (ret)
2424 closure_wait_event(&c->btree_interior_update_wait,
2425 !bch2_btree_interior_updates_pending(c));
2426 return ret;
2427 }
2428
bch2_journal_entry_to_btree_root(struct bch_fs * c,struct jset_entry * entry)2429 void bch2_journal_entry_to_btree_root(struct bch_fs *c, struct jset_entry *entry)
2430 {
2431 struct btree_root *r = bch2_btree_id_root(c, entry->btree_id);
2432
2433 mutex_lock(&c->btree_root_lock);
2434
2435 r->level = entry->level;
2436 r->alive = true;
2437 bkey_copy(&r->key, (struct bkey_i *) entry->start);
2438
2439 mutex_unlock(&c->btree_root_lock);
2440 }
2441
2442 struct jset_entry *
bch2_btree_roots_to_journal_entries(struct bch_fs * c,struct jset_entry * end,unsigned long skip)2443 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2444 struct jset_entry *end,
2445 unsigned long skip)
2446 {
2447 unsigned i;
2448
2449 mutex_lock(&c->btree_root_lock);
2450
2451 for (i = 0; i < btree_id_nr_alive(c); i++) {
2452 struct btree_root *r = bch2_btree_id_root(c, i);
2453
2454 if (r->alive && !test_bit(i, &skip)) {
2455 journal_entry_set(end, BCH_JSET_ENTRY_btree_root,
2456 i, r->level, &r->key, r->key.k.u64s);
2457 end = vstruct_next(end);
2458 }
2459 }
2460
2461 mutex_unlock(&c->btree_root_lock);
2462
2463 return end;
2464 }
2465
bch2_fs_btree_interior_update_exit(struct bch_fs * c)2466 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2467 {
2468 if (c->btree_interior_update_worker)
2469 destroy_workqueue(c->btree_interior_update_worker);
2470 mempool_exit(&c->btree_interior_update_pool);
2471 }
2472
bch2_fs_btree_interior_update_init_early(struct bch_fs * c)2473 void bch2_fs_btree_interior_update_init_early(struct bch_fs *c)
2474 {
2475 mutex_init(&c->btree_reserve_cache_lock);
2476 INIT_LIST_HEAD(&c->btree_interior_update_list);
2477 INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2478 mutex_init(&c->btree_interior_update_lock);
2479 INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2480
2481 INIT_LIST_HEAD(&c->pending_node_rewrites);
2482 mutex_init(&c->pending_node_rewrites_lock);
2483 }
2484
bch2_fs_btree_interior_update_init(struct bch_fs * c)2485 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2486 {
2487 c->btree_interior_update_worker =
2488 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2489 if (!c->btree_interior_update_worker)
2490 return -BCH_ERR_ENOMEM_btree_interior_update_worker_init;
2491
2492 if (mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2493 sizeof(struct btree_update)))
2494 return -BCH_ERR_ENOMEM_btree_interior_update_pool_init;
2495
2496 return 0;
2497 }
2498