xref: /linux/fs/bcachefs/btree_update_interior.c (revision 7f71507851fc7764b36a3221839607d3a45c2025)
1 // SPDX-License-Identifier: GPL-2.0
2 
3 #include "bcachefs.h"
4 #include "alloc_foreground.h"
5 #include "bkey_buf.h"
6 #include "bkey_methods.h"
7 #include "btree_cache.h"
8 #include "btree_gc.h"
9 #include "btree_journal_iter.h"
10 #include "btree_update.h"
11 #include "btree_update_interior.h"
12 #include "btree_io.h"
13 #include "btree_iter.h"
14 #include "btree_locking.h"
15 #include "buckets.h"
16 #include "clock.h"
17 #include "error.h"
18 #include "extents.h"
19 #include "io_write.h"
20 #include "journal.h"
21 #include "journal_reclaim.h"
22 #include "keylist.h"
23 #include "recovery_passes.h"
24 #include "replicas.h"
25 #include "sb-members.h"
26 #include "super-io.h"
27 #include "trace.h"
28 
29 #include <linux/random.h>
30 
31 static const char * const bch2_btree_update_modes[] = {
32 #define x(t) #t,
33 	BTREE_UPDATE_MODES()
34 #undef x
35 	NULL
36 };
37 
38 static int bch2_btree_insert_node(struct btree_update *, struct btree_trans *,
39 				  btree_path_idx_t, struct btree *, struct keylist *);
40 static void bch2_btree_update_add_new_node(struct btree_update *, struct btree *);
41 
42 /*
43  * Verify that child nodes correctly span parent node's range:
44  */
45 int bch2_btree_node_check_topology(struct btree_trans *trans, struct btree *b)
46 {
47 	struct bch_fs *c = trans->c;
48 	struct bpos node_min = b->key.k.type == KEY_TYPE_btree_ptr_v2
49 		? bkey_i_to_btree_ptr_v2(&b->key)->v.min_key
50 		: b->data->min_key;
51 	struct btree_and_journal_iter iter;
52 	struct bkey_s_c k;
53 	struct printbuf buf = PRINTBUF;
54 	struct bkey_buf prev;
55 	int ret = 0;
56 
57 	BUG_ON(b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
58 	       !bpos_eq(bkey_i_to_btree_ptr_v2(&b->key)->v.min_key,
59 			b->data->min_key));
60 
61 	if (b == btree_node_root(c, b)) {
62 		if (!bpos_eq(b->data->min_key, POS_MIN)) {
63 			printbuf_reset(&buf);
64 			bch2_bpos_to_text(&buf, b->data->min_key);
65 			need_fsck_err(trans, btree_root_bad_min_key,
66 				      "btree root with incorrect min_key: %s", buf.buf);
67 			goto topology_repair;
68 		}
69 
70 		if (!bpos_eq(b->data->max_key, SPOS_MAX)) {
71 			printbuf_reset(&buf);
72 			bch2_bpos_to_text(&buf, b->data->max_key);
73 			need_fsck_err(trans, btree_root_bad_max_key,
74 				      "btree root with incorrect max_key: %s", buf.buf);
75 			goto topology_repair;
76 		}
77 	}
78 
79 	if (!b->c.level)
80 		return 0;
81 
82 	bch2_bkey_buf_init(&prev);
83 	bkey_init(&prev.k->k);
84 	bch2_btree_and_journal_iter_init_node_iter(trans, &iter, b);
85 
86 	while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
87 		if (k.k->type != KEY_TYPE_btree_ptr_v2)
88 			goto out;
89 
90 		struct bkey_s_c_btree_ptr_v2 bp = bkey_s_c_to_btree_ptr_v2(k);
91 
92 		struct bpos expected_min = bkey_deleted(&prev.k->k)
93 			? node_min
94 			: bpos_successor(prev.k->k.p);
95 
96 		if (!bpos_eq(expected_min, bp.v->min_key)) {
97 			bch2_topology_error(c);
98 
99 			printbuf_reset(&buf);
100 			prt_str(&buf, "end of prev node doesn't match start of next node\n"),
101 			prt_printf(&buf, "  in btree %s level %u node ",
102 				   bch2_btree_id_str(b->c.btree_id), b->c.level);
103 			bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
104 			prt_str(&buf, "\n  prev ");
105 			bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(prev.k));
106 			prt_str(&buf, "\n  next ");
107 			bch2_bkey_val_to_text(&buf, c, k);
108 
109 			need_fsck_err(trans, btree_node_topology_bad_min_key, "%s", buf.buf);
110 			goto topology_repair;
111 		}
112 
113 		bch2_bkey_buf_reassemble(&prev, c, k);
114 		bch2_btree_and_journal_iter_advance(&iter);
115 	}
116 
117 	if (bkey_deleted(&prev.k->k)) {
118 		bch2_topology_error(c);
119 
120 		printbuf_reset(&buf);
121 		prt_str(&buf, "empty interior node\n");
122 		prt_printf(&buf, "  in btree %s level %u node ",
123 			   bch2_btree_id_str(b->c.btree_id), b->c.level);
124 		bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
125 
126 		need_fsck_err(trans, btree_node_topology_empty_interior_node, "%s", buf.buf);
127 		goto topology_repair;
128 	} else if (!bpos_eq(prev.k->k.p, b->key.k.p)) {
129 		bch2_topology_error(c);
130 
131 		printbuf_reset(&buf);
132 		prt_str(&buf, "last child node doesn't end at end of parent node\n");
133 		prt_printf(&buf, "  in btree %s level %u node ",
134 			   bch2_btree_id_str(b->c.btree_id), b->c.level);
135 		bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
136 		prt_str(&buf, "\n  last key ");
137 		bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(prev.k));
138 
139 		need_fsck_err(trans, btree_node_topology_bad_max_key, "%s", buf.buf);
140 		goto topology_repair;
141 	}
142 out:
143 fsck_err:
144 	bch2_btree_and_journal_iter_exit(&iter);
145 	bch2_bkey_buf_exit(&prev, c);
146 	printbuf_exit(&buf);
147 	return ret;
148 topology_repair:
149 	if ((c->opts.recovery_passes & BIT_ULL(BCH_RECOVERY_PASS_check_topology)) &&
150 	    c->curr_recovery_pass > BCH_RECOVERY_PASS_check_topology) {
151 		bch2_inconsistent_error(c);
152 		ret = -BCH_ERR_btree_need_topology_repair;
153 	} else {
154 		ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
155 	}
156 	goto out;
157 }
158 
159 /* Calculate ideal packed bkey format for new btree nodes: */
160 
161 static void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
162 {
163 	struct bkey_packed *k;
164 	struct bkey uk;
165 
166 	for_each_bset(b, t)
167 		bset_tree_for_each_key(b, t, k)
168 			if (!bkey_deleted(k)) {
169 				uk = bkey_unpack_key(b, k);
170 				bch2_bkey_format_add_key(s, &uk);
171 			}
172 }
173 
174 static struct bkey_format bch2_btree_calc_format(struct btree *b)
175 {
176 	struct bkey_format_state s;
177 
178 	bch2_bkey_format_init(&s);
179 	bch2_bkey_format_add_pos(&s, b->data->min_key);
180 	bch2_bkey_format_add_pos(&s, b->data->max_key);
181 	__bch2_btree_calc_format(&s, b);
182 
183 	return bch2_bkey_format_done(&s);
184 }
185 
186 static size_t btree_node_u64s_with_format(struct btree_nr_keys nr,
187 					  struct bkey_format *old_f,
188 					  struct bkey_format *new_f)
189 {
190 	/* stupid integer promotion rules */
191 	ssize_t delta =
192 	    (((int) new_f->key_u64s - old_f->key_u64s) *
193 	     (int) nr.packed_keys) +
194 	    (((int) new_f->key_u64s - BKEY_U64s) *
195 	     (int) nr.unpacked_keys);
196 
197 	BUG_ON(delta + nr.live_u64s < 0);
198 
199 	return nr.live_u64s + delta;
200 }
201 
202 /**
203  * bch2_btree_node_format_fits - check if we could rewrite node with a new format
204  *
205  * @c:		filesystem handle
206  * @b:		btree node to rewrite
207  * @nr:		number of keys for new node (i.e. b->nr)
208  * @new_f:	bkey format to translate keys to
209  *
210  * Returns: true if all re-packed keys will be able to fit in a new node.
211  *
212  * Assumes all keys will successfully pack with the new format.
213  */
214 static bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
215 				 struct btree_nr_keys nr,
216 				 struct bkey_format *new_f)
217 {
218 	size_t u64s = btree_node_u64s_with_format(nr, &b->format, new_f);
219 
220 	return __vstruct_bytes(struct btree_node, u64s) < btree_buf_bytes(b);
221 }
222 
223 /* Btree node freeing/allocation: */
224 
225 static void __btree_node_free(struct btree_trans *trans, struct btree *b)
226 {
227 	struct bch_fs *c = trans->c;
228 
229 	trace_and_count(c, btree_node_free, trans, b);
230 
231 	BUG_ON(btree_node_write_blocked(b));
232 	BUG_ON(btree_node_dirty(b));
233 	BUG_ON(btree_node_need_write(b));
234 	BUG_ON(b == btree_node_root(c, b));
235 	BUG_ON(b->ob.nr);
236 	BUG_ON(!list_empty(&b->write_blocked));
237 	BUG_ON(b->will_make_reachable);
238 
239 	clear_btree_node_noevict(b);
240 }
241 
242 static void bch2_btree_node_free_inmem(struct btree_trans *trans,
243 				       struct btree_path *path,
244 				       struct btree *b)
245 {
246 	struct bch_fs *c = trans->c;
247 	unsigned i, level = b->c.level;
248 
249 	bch2_btree_node_lock_write_nofail(trans, path, &b->c);
250 
251 	__btree_node_free(trans, b);
252 
253 	mutex_lock(&c->btree_cache.lock);
254 	bch2_btree_node_hash_remove(&c->btree_cache, b);
255 	mutex_unlock(&c->btree_cache.lock);
256 
257 	six_unlock_write(&b->c.lock);
258 	mark_btree_node_locked_noreset(path, level, BTREE_NODE_INTENT_LOCKED);
259 
260 	trans_for_each_path(trans, path, i)
261 		if (path->l[level].b == b) {
262 			btree_node_unlock(trans, path, level);
263 			path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
264 		}
265 }
266 
267 static void bch2_btree_node_free_never_used(struct btree_update *as,
268 					    struct btree_trans *trans,
269 					    struct btree *b)
270 {
271 	struct bch_fs *c = as->c;
272 	struct prealloc_nodes *p = &as->prealloc_nodes[b->c.lock.readers != NULL];
273 	struct btree_path *path;
274 	unsigned i, level = b->c.level;
275 
276 	BUG_ON(!list_empty(&b->write_blocked));
277 	BUG_ON(b->will_make_reachable != (1UL|(unsigned long) as));
278 
279 	b->will_make_reachable = 0;
280 	closure_put(&as->cl);
281 
282 	clear_btree_node_will_make_reachable(b);
283 	clear_btree_node_accessed(b);
284 	clear_btree_node_dirty_acct(c, b);
285 	clear_btree_node_need_write(b);
286 
287 	mutex_lock(&c->btree_cache.lock);
288 	__bch2_btree_node_hash_remove(&c->btree_cache, b);
289 	mutex_unlock(&c->btree_cache.lock);
290 
291 	BUG_ON(p->nr >= ARRAY_SIZE(p->b));
292 	p->b[p->nr++] = b;
293 
294 	six_unlock_intent(&b->c.lock);
295 
296 	trans_for_each_path(trans, path, i)
297 		if (path->l[level].b == b) {
298 			btree_node_unlock(trans, path, level);
299 			path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
300 		}
301 }
302 
303 static struct btree *__bch2_btree_node_alloc(struct btree_trans *trans,
304 					     struct disk_reservation *res,
305 					     struct closure *cl,
306 					     bool interior_node,
307 					     unsigned flags)
308 {
309 	struct bch_fs *c = trans->c;
310 	struct write_point *wp;
311 	struct btree *b;
312 	BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
313 	struct open_buckets obs = { .nr = 0 };
314 	struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
315 	enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
316 	unsigned nr_reserve = watermark < BCH_WATERMARK_reclaim
317 		? BTREE_NODE_RESERVE
318 		: 0;
319 	int ret;
320 
321 	b = bch2_btree_node_mem_alloc(trans, interior_node);
322 	if (IS_ERR(b))
323 		return b;
324 
325 	BUG_ON(b->ob.nr);
326 
327 	mutex_lock(&c->btree_reserve_cache_lock);
328 	if (c->btree_reserve_cache_nr > nr_reserve) {
329 		struct btree_alloc *a =
330 			&c->btree_reserve_cache[--c->btree_reserve_cache_nr];
331 
332 		obs = a->ob;
333 		bkey_copy(&tmp.k, &a->k);
334 		mutex_unlock(&c->btree_reserve_cache_lock);
335 		goto out;
336 	}
337 	mutex_unlock(&c->btree_reserve_cache_lock);
338 retry:
339 	ret = bch2_alloc_sectors_start_trans(trans,
340 				      c->opts.metadata_target ?:
341 				      c->opts.foreground_target,
342 				      0,
343 				      writepoint_ptr(&c->btree_write_point),
344 				      &devs_have,
345 				      res->nr_replicas,
346 				      min(res->nr_replicas,
347 					  c->opts.metadata_replicas_required),
348 				      watermark, 0, cl, &wp);
349 	if (unlikely(ret))
350 		goto err;
351 
352 	if (wp->sectors_free < btree_sectors(c)) {
353 		struct open_bucket *ob;
354 		unsigned i;
355 
356 		open_bucket_for_each(c, &wp->ptrs, ob, i)
357 			if (ob->sectors_free < btree_sectors(c))
358 				ob->sectors_free = 0;
359 
360 		bch2_alloc_sectors_done(c, wp);
361 		goto retry;
362 	}
363 
364 	bkey_btree_ptr_v2_init(&tmp.k);
365 	bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, btree_sectors(c), false);
366 
367 	bch2_open_bucket_get(c, wp, &obs);
368 	bch2_alloc_sectors_done(c, wp);
369 out:
370 	bkey_copy(&b->key, &tmp.k);
371 	b->ob = obs;
372 	six_unlock_write(&b->c.lock);
373 	six_unlock_intent(&b->c.lock);
374 
375 	return b;
376 err:
377 	bch2_btree_node_to_freelist(c, b);
378 	return ERR_PTR(ret);
379 }
380 
381 static struct btree *bch2_btree_node_alloc(struct btree_update *as,
382 					   struct btree_trans *trans,
383 					   unsigned level)
384 {
385 	struct bch_fs *c = as->c;
386 	struct btree *b;
387 	struct prealloc_nodes *p = &as->prealloc_nodes[!!level];
388 	int ret;
389 
390 	BUG_ON(level >= BTREE_MAX_DEPTH);
391 	BUG_ON(!p->nr);
392 
393 	b = p->b[--p->nr];
394 
395 	btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
396 	btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
397 
398 	set_btree_node_accessed(b);
399 	set_btree_node_dirty_acct(c, b);
400 	set_btree_node_need_write(b);
401 
402 	bch2_bset_init_first(b, &b->data->keys);
403 	b->c.level	= level;
404 	b->c.btree_id	= as->btree_id;
405 	b->version_ondisk = c->sb.version;
406 
407 	memset(&b->nr, 0, sizeof(b->nr));
408 	b->data->magic = cpu_to_le64(bset_magic(c));
409 	memset(&b->data->_ptr, 0, sizeof(b->data->_ptr));
410 	b->data->flags = 0;
411 	SET_BTREE_NODE_ID(b->data, as->btree_id);
412 	SET_BTREE_NODE_LEVEL(b->data, level);
413 
414 	if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
415 		struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
416 
417 		bp->v.mem_ptr		= 0;
418 		bp->v.seq		= b->data->keys.seq;
419 		bp->v.sectors_written	= 0;
420 	}
421 
422 	SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
423 
424 	bch2_btree_build_aux_trees(b);
425 
426 	ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
427 	BUG_ON(ret);
428 
429 	trace_and_count(c, btree_node_alloc, trans, b);
430 	bch2_increment_clock(c, btree_sectors(c), WRITE);
431 	return b;
432 }
433 
434 static void btree_set_min(struct btree *b, struct bpos pos)
435 {
436 	if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
437 		bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
438 	b->data->min_key = pos;
439 }
440 
441 static void btree_set_max(struct btree *b, struct bpos pos)
442 {
443 	b->key.k.p = pos;
444 	b->data->max_key = pos;
445 }
446 
447 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
448 						       struct btree_trans *trans,
449 						       struct btree *b)
450 {
451 	struct btree *n = bch2_btree_node_alloc(as, trans, b->c.level);
452 	struct bkey_format format = bch2_btree_calc_format(b);
453 
454 	/*
455 	 * The keys might expand with the new format - if they wouldn't fit in
456 	 * the btree node anymore, use the old format for now:
457 	 */
458 	if (!bch2_btree_node_format_fits(as->c, b, b->nr, &format))
459 		format = b->format;
460 
461 	SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
462 
463 	btree_set_min(n, b->data->min_key);
464 	btree_set_max(n, b->data->max_key);
465 
466 	n->data->format		= format;
467 	btree_node_set_format(n, format);
468 
469 	bch2_btree_sort_into(as->c, n, b);
470 
471 	btree_node_reset_sib_u64s(n);
472 	return n;
473 }
474 
475 static struct btree *__btree_root_alloc(struct btree_update *as,
476 				struct btree_trans *trans, unsigned level)
477 {
478 	struct btree *b = bch2_btree_node_alloc(as, trans, level);
479 
480 	btree_set_min(b, POS_MIN);
481 	btree_set_max(b, SPOS_MAX);
482 	b->data->format = bch2_btree_calc_format(b);
483 
484 	btree_node_set_format(b, b->data->format);
485 	bch2_btree_build_aux_trees(b);
486 
487 	return b;
488 }
489 
490 static void bch2_btree_reserve_put(struct btree_update *as, struct btree_trans *trans)
491 {
492 	struct bch_fs *c = as->c;
493 	struct prealloc_nodes *p;
494 
495 	for (p = as->prealloc_nodes;
496 	     p < as->prealloc_nodes + ARRAY_SIZE(as->prealloc_nodes);
497 	     p++) {
498 		while (p->nr) {
499 			struct btree *b = p->b[--p->nr];
500 
501 			mutex_lock(&c->btree_reserve_cache_lock);
502 
503 			if (c->btree_reserve_cache_nr <
504 			    ARRAY_SIZE(c->btree_reserve_cache)) {
505 				struct btree_alloc *a =
506 					&c->btree_reserve_cache[c->btree_reserve_cache_nr++];
507 
508 				a->ob = b->ob;
509 				b->ob.nr = 0;
510 				bkey_copy(&a->k, &b->key);
511 			} else {
512 				bch2_open_buckets_put(c, &b->ob);
513 			}
514 
515 			mutex_unlock(&c->btree_reserve_cache_lock);
516 
517 			btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
518 			btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
519 			__btree_node_free(trans, b);
520 			bch2_btree_node_to_freelist(c, b);
521 		}
522 	}
523 }
524 
525 static int bch2_btree_reserve_get(struct btree_trans *trans,
526 				  struct btree_update *as,
527 				  unsigned nr_nodes[2],
528 				  unsigned flags,
529 				  struct closure *cl)
530 {
531 	struct btree *b;
532 	unsigned interior;
533 	int ret = 0;
534 
535 	BUG_ON(nr_nodes[0] + nr_nodes[1] > BTREE_RESERVE_MAX);
536 
537 	/*
538 	 * Protects reaping from the btree node cache and using the btree node
539 	 * open bucket reserve:
540 	 */
541 	ret = bch2_btree_cache_cannibalize_lock(trans, cl);
542 	if (ret)
543 		return ret;
544 
545 	for (interior = 0; interior < 2; interior++) {
546 		struct prealloc_nodes *p = as->prealloc_nodes + interior;
547 
548 		while (p->nr < nr_nodes[interior]) {
549 			b = __bch2_btree_node_alloc(trans, &as->disk_res, cl,
550 						    interior, flags);
551 			if (IS_ERR(b)) {
552 				ret = PTR_ERR(b);
553 				goto err;
554 			}
555 
556 			p->b[p->nr++] = b;
557 		}
558 	}
559 err:
560 	bch2_btree_cache_cannibalize_unlock(trans);
561 	return ret;
562 }
563 
564 /* Asynchronous interior node update machinery */
565 
566 static void bch2_btree_update_free(struct btree_update *as, struct btree_trans *trans)
567 {
568 	struct bch_fs *c = as->c;
569 
570 	if (as->took_gc_lock)
571 		up_read(&c->gc_lock);
572 	as->took_gc_lock = false;
573 
574 	bch2_journal_pin_drop(&c->journal, &as->journal);
575 	bch2_journal_pin_flush(&c->journal, &as->journal);
576 	bch2_disk_reservation_put(c, &as->disk_res);
577 	bch2_btree_reserve_put(as, trans);
578 
579 	bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_total],
580 			       as->start_time);
581 
582 	mutex_lock(&c->btree_interior_update_lock);
583 	list_del(&as->unwritten_list);
584 	list_del(&as->list);
585 
586 	closure_debug_destroy(&as->cl);
587 	mempool_free(as, &c->btree_interior_update_pool);
588 
589 	/*
590 	 * Have to do the wakeup with btree_interior_update_lock still held,
591 	 * since being on btree_interior_update_list is our ref on @c:
592 	 */
593 	closure_wake_up(&c->btree_interior_update_wait);
594 
595 	mutex_unlock(&c->btree_interior_update_lock);
596 }
597 
598 static void btree_update_add_key(struct btree_update *as,
599 				 struct keylist *keys, struct btree *b)
600 {
601 	struct bkey_i *k = &b->key;
602 
603 	BUG_ON(bch2_keylist_u64s(keys) + k->k.u64s >
604 	       ARRAY_SIZE(as->_old_keys));
605 
606 	bkey_copy(keys->top, k);
607 	bkey_i_to_btree_ptr_v2(keys->top)->v.mem_ptr = b->c.level + 1;
608 
609 	bch2_keylist_push(keys);
610 }
611 
612 static bool btree_update_new_nodes_marked_sb(struct btree_update *as)
613 {
614 	for_each_keylist_key(&as->new_keys, k)
615 		if (!bch2_dev_btree_bitmap_marked(as->c, bkey_i_to_s_c(k)))
616 			return false;
617 	return true;
618 }
619 
620 static void btree_update_new_nodes_mark_sb(struct btree_update *as)
621 {
622 	struct bch_fs *c = as->c;
623 
624 	mutex_lock(&c->sb_lock);
625 	for_each_keylist_key(&as->new_keys, k)
626 		bch2_dev_btree_bitmap_mark(c, bkey_i_to_s_c(k));
627 
628 	bch2_write_super(c);
629 	mutex_unlock(&c->sb_lock);
630 }
631 
632 /*
633  * The transactional part of an interior btree node update, where we journal the
634  * update we did to the interior node and update alloc info:
635  */
636 static int btree_update_nodes_written_trans(struct btree_trans *trans,
637 					    struct btree_update *as)
638 {
639 	struct jset_entry *e = bch2_trans_jset_entry_alloc(trans, as->journal_u64s);
640 	int ret = PTR_ERR_OR_ZERO(e);
641 	if (ret)
642 		return ret;
643 
644 	memcpy(e, as->journal_entries, as->journal_u64s * sizeof(u64));
645 
646 	trans->journal_pin = &as->journal;
647 
648 	for_each_keylist_key(&as->old_keys, k) {
649 		unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
650 
651 		ret = bch2_key_trigger_old(trans, as->btree_id, level, bkey_i_to_s_c(k),
652 					   BTREE_TRIGGER_transactional);
653 		if (ret)
654 			return ret;
655 	}
656 
657 	for_each_keylist_key(&as->new_keys, k) {
658 		unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
659 
660 		ret = bch2_key_trigger_new(trans, as->btree_id, level, bkey_i_to_s(k),
661 					   BTREE_TRIGGER_transactional);
662 		if (ret)
663 			return ret;
664 	}
665 
666 	return 0;
667 }
668 
669 static void btree_update_nodes_written(struct btree_update *as)
670 {
671 	struct bch_fs *c = as->c;
672 	struct btree *b;
673 	struct btree_trans *trans = bch2_trans_get(c);
674 	u64 journal_seq = 0;
675 	unsigned i;
676 	int ret;
677 
678 	/*
679 	 * If we're already in an error state, it might be because a btree node
680 	 * was never written, and we might be trying to free that same btree
681 	 * node here, but it won't have been marked as allocated and we'll see
682 	 * spurious disk usage inconsistencies in the transactional part below
683 	 * if we don't skip it:
684 	 */
685 	ret = bch2_journal_error(&c->journal);
686 	if (ret)
687 		goto err;
688 
689 	if (!btree_update_new_nodes_marked_sb(as))
690 		btree_update_new_nodes_mark_sb(as);
691 
692 	/*
693 	 * Wait for any in flight writes to finish before we free the old nodes
694 	 * on disk:
695 	 */
696 	for (i = 0; i < as->nr_old_nodes; i++) {
697 		__le64 seq;
698 
699 		b = as->old_nodes[i];
700 
701 		btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
702 		seq = b->data ? b->data->keys.seq : 0;
703 		six_unlock_read(&b->c.lock);
704 
705 		if (seq == as->old_nodes_seq[i])
706 			wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight_inner,
707 				       TASK_UNINTERRUPTIBLE);
708 	}
709 
710 	/*
711 	 * We did an update to a parent node where the pointers we added pointed
712 	 * to child nodes that weren't written yet: now, the child nodes have
713 	 * been written so we can write out the update to the interior node.
714 	 */
715 
716 	/*
717 	 * We can't call into journal reclaim here: we'd block on the journal
718 	 * reclaim lock, but we may need to release the open buckets we have
719 	 * pinned in order for other btree updates to make forward progress, and
720 	 * journal reclaim does btree updates when flushing bkey_cached entries,
721 	 * which may require allocations as well.
722 	 */
723 	ret = commit_do(trans, &as->disk_res, &journal_seq,
724 			BCH_WATERMARK_interior_updates|
725 			BCH_TRANS_COMMIT_no_enospc|
726 			BCH_TRANS_COMMIT_no_check_rw|
727 			BCH_TRANS_COMMIT_journal_reclaim,
728 			btree_update_nodes_written_trans(trans, as));
729 	bch2_trans_unlock(trans);
730 
731 	bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
732 			     "%s", bch2_err_str(ret));
733 err:
734 	/*
735 	 * Ensure transaction is unlocked before using btree_node_lock_nopath()
736 	 * (the use of which is always suspect, we need to work on removing this
737 	 * in the future)
738 	 *
739 	 * It should be, but bch2_path_get_unlocked_mut() -> bch2_path_get()
740 	 * calls bch2_path_upgrade(), before we call path_make_mut(), so we may
741 	 * rarely end up with a locked path besides the one we have here:
742 	 */
743 	bch2_trans_unlock(trans);
744 	bch2_trans_begin(trans);
745 
746 	/*
747 	 * We have to be careful because another thread might be getting ready
748 	 * to free as->b and calling btree_update_reparent() on us - we'll
749 	 * recheck under btree_update_lock below:
750 	 */
751 	b = READ_ONCE(as->b);
752 	if (b) {
753 		/*
754 		 * @b is the node we did the final insert into:
755 		 *
756 		 * On failure to get a journal reservation, we still have to
757 		 * unblock the write and allow most of the write path to happen
758 		 * so that shutdown works, but the i->journal_seq mechanism
759 		 * won't work to prevent the btree write from being visible (we
760 		 * didn't get a journal sequence number) - instead
761 		 * __bch2_btree_node_write() doesn't do the actual write if
762 		 * we're in journal error state:
763 		 */
764 
765 		btree_path_idx_t path_idx = bch2_path_get_unlocked_mut(trans,
766 						as->btree_id, b->c.level, b->key.k.p);
767 		struct btree_path *path = trans->paths + path_idx;
768 		btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
769 		mark_btree_node_locked(trans, path, b->c.level, BTREE_NODE_INTENT_LOCKED);
770 		path->l[b->c.level].lock_seq = six_lock_seq(&b->c.lock);
771 		path->l[b->c.level].b = b;
772 
773 		bch2_btree_node_lock_write_nofail(trans, path, &b->c);
774 
775 		mutex_lock(&c->btree_interior_update_lock);
776 
777 		list_del(&as->write_blocked_list);
778 		if (list_empty(&b->write_blocked))
779 			clear_btree_node_write_blocked(b);
780 
781 		/*
782 		 * Node might have been freed, recheck under
783 		 * btree_interior_update_lock:
784 		 */
785 		if (as->b == b) {
786 			BUG_ON(!b->c.level);
787 			BUG_ON(!btree_node_dirty(b));
788 
789 			if (!ret) {
790 				struct bset *last = btree_bset_last(b);
791 
792 				last->journal_seq = cpu_to_le64(
793 							     max(journal_seq,
794 								 le64_to_cpu(last->journal_seq)));
795 
796 				bch2_btree_add_journal_pin(c, b, journal_seq);
797 			} else {
798 				/*
799 				 * If we didn't get a journal sequence number we
800 				 * can't write this btree node, because recovery
801 				 * won't know to ignore this write:
802 				 */
803 				set_btree_node_never_write(b);
804 			}
805 		}
806 
807 		mutex_unlock(&c->btree_interior_update_lock);
808 
809 		mark_btree_node_locked_noreset(path, b->c.level, BTREE_NODE_INTENT_LOCKED);
810 		six_unlock_write(&b->c.lock);
811 
812 		btree_node_write_if_need(c, b, SIX_LOCK_intent);
813 		btree_node_unlock(trans, path, b->c.level);
814 		bch2_path_put(trans, path_idx, true);
815 	}
816 
817 	bch2_journal_pin_drop(&c->journal, &as->journal);
818 
819 	mutex_lock(&c->btree_interior_update_lock);
820 	for (i = 0; i < as->nr_new_nodes; i++) {
821 		b = as->new_nodes[i];
822 
823 		BUG_ON(b->will_make_reachable != (unsigned long) as);
824 		b->will_make_reachable = 0;
825 		clear_btree_node_will_make_reachable(b);
826 	}
827 	mutex_unlock(&c->btree_interior_update_lock);
828 
829 	for (i = 0; i < as->nr_new_nodes; i++) {
830 		b = as->new_nodes[i];
831 
832 		btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
833 		btree_node_write_if_need(c, b, SIX_LOCK_read);
834 		six_unlock_read(&b->c.lock);
835 	}
836 
837 	for (i = 0; i < as->nr_open_buckets; i++)
838 		bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
839 
840 	bch2_btree_update_free(as, trans);
841 	bch2_trans_put(trans);
842 }
843 
844 static void btree_interior_update_work(struct work_struct *work)
845 {
846 	struct bch_fs *c =
847 		container_of(work, struct bch_fs, btree_interior_update_work);
848 	struct btree_update *as;
849 
850 	while (1) {
851 		mutex_lock(&c->btree_interior_update_lock);
852 		as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
853 					      struct btree_update, unwritten_list);
854 		if (as && !as->nodes_written)
855 			as = NULL;
856 		mutex_unlock(&c->btree_interior_update_lock);
857 
858 		if (!as)
859 			break;
860 
861 		btree_update_nodes_written(as);
862 	}
863 }
864 
865 static CLOSURE_CALLBACK(btree_update_set_nodes_written)
866 {
867 	closure_type(as, struct btree_update, cl);
868 	struct bch_fs *c = as->c;
869 
870 	mutex_lock(&c->btree_interior_update_lock);
871 	as->nodes_written = true;
872 	mutex_unlock(&c->btree_interior_update_lock);
873 
874 	queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
875 }
876 
877 /*
878  * We're updating @b with pointers to nodes that haven't finished writing yet:
879  * block @b from being written until @as completes
880  */
881 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
882 {
883 	struct bch_fs *c = as->c;
884 
885 	BUG_ON(as->mode != BTREE_UPDATE_none);
886 	BUG_ON(as->update_level_end < b->c.level);
887 	BUG_ON(!btree_node_dirty(b));
888 	BUG_ON(!b->c.level);
889 
890 	mutex_lock(&c->btree_interior_update_lock);
891 	list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
892 
893 	as->mode	= BTREE_UPDATE_node;
894 	as->b		= b;
895 	as->update_level_end = b->c.level;
896 
897 	set_btree_node_write_blocked(b);
898 	list_add(&as->write_blocked_list, &b->write_blocked);
899 
900 	mutex_unlock(&c->btree_interior_update_lock);
901 }
902 
903 static int bch2_update_reparent_journal_pin_flush(struct journal *j,
904 				struct journal_entry_pin *_pin, u64 seq)
905 {
906 	return 0;
907 }
908 
909 static void btree_update_reparent(struct btree_update *as,
910 				  struct btree_update *child)
911 {
912 	struct bch_fs *c = as->c;
913 
914 	lockdep_assert_held(&c->btree_interior_update_lock);
915 
916 	child->b = NULL;
917 	child->mode = BTREE_UPDATE_update;
918 
919 	bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal,
920 			      bch2_update_reparent_journal_pin_flush);
921 }
922 
923 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
924 {
925 	struct bkey_i *insert = &b->key;
926 	struct bch_fs *c = as->c;
927 
928 	BUG_ON(as->mode != BTREE_UPDATE_none);
929 
930 	BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
931 	       ARRAY_SIZE(as->journal_entries));
932 
933 	as->journal_u64s +=
934 		journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
935 				  BCH_JSET_ENTRY_btree_root,
936 				  b->c.btree_id, b->c.level,
937 				  insert, insert->k.u64s);
938 
939 	mutex_lock(&c->btree_interior_update_lock);
940 	list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
941 
942 	as->mode	= BTREE_UPDATE_root;
943 	mutex_unlock(&c->btree_interior_update_lock);
944 }
945 
946 /*
947  * bch2_btree_update_add_new_node:
948  *
949  * This causes @as to wait on @b to be written, before it gets to
950  * bch2_btree_update_nodes_written
951  *
952  * Additionally, it sets b->will_make_reachable to prevent any additional writes
953  * to @b from happening besides the first until @b is reachable on disk
954  *
955  * And it adds @b to the list of @as's new nodes, so that we can update sector
956  * counts in bch2_btree_update_nodes_written:
957  */
958 static void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
959 {
960 	struct bch_fs *c = as->c;
961 
962 	closure_get(&as->cl);
963 
964 	mutex_lock(&c->btree_interior_update_lock);
965 	BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
966 	BUG_ON(b->will_make_reachable);
967 
968 	as->new_nodes[as->nr_new_nodes++] = b;
969 	b->will_make_reachable = 1UL|(unsigned long) as;
970 	set_btree_node_will_make_reachable(b);
971 
972 	mutex_unlock(&c->btree_interior_update_lock);
973 
974 	btree_update_add_key(as, &as->new_keys, b);
975 
976 	if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
977 		unsigned bytes = vstruct_end(&b->data->keys) - (void *) b->data;
978 		unsigned sectors = round_up(bytes, block_bytes(c)) >> 9;
979 
980 		bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
981 			cpu_to_le16(sectors);
982 	}
983 }
984 
985 /*
986  * returns true if @b was a new node
987  */
988 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
989 {
990 	struct btree_update *as;
991 	unsigned long v;
992 	unsigned i;
993 
994 	mutex_lock(&c->btree_interior_update_lock);
995 	/*
996 	 * When b->will_make_reachable != 0, it owns a ref on as->cl that's
997 	 * dropped when it gets written by bch2_btree_complete_write - the
998 	 * xchg() is for synchronization with bch2_btree_complete_write:
999 	 */
1000 	v = xchg(&b->will_make_reachable, 0);
1001 	clear_btree_node_will_make_reachable(b);
1002 	as = (struct btree_update *) (v & ~1UL);
1003 
1004 	if (!as) {
1005 		mutex_unlock(&c->btree_interior_update_lock);
1006 		return;
1007 	}
1008 
1009 	for (i = 0; i < as->nr_new_nodes; i++)
1010 		if (as->new_nodes[i] == b)
1011 			goto found;
1012 
1013 	BUG();
1014 found:
1015 	array_remove_item(as->new_nodes, as->nr_new_nodes, i);
1016 	mutex_unlock(&c->btree_interior_update_lock);
1017 
1018 	if (v & 1)
1019 		closure_put(&as->cl);
1020 }
1021 
1022 static void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
1023 {
1024 	while (b->ob.nr)
1025 		as->open_buckets[as->nr_open_buckets++] =
1026 			b->ob.v[--b->ob.nr];
1027 }
1028 
1029 static int bch2_btree_update_will_free_node_journal_pin_flush(struct journal *j,
1030 				struct journal_entry_pin *_pin, u64 seq)
1031 {
1032 	return 0;
1033 }
1034 
1035 /*
1036  * @b is being split/rewritten: it may have pointers to not-yet-written btree
1037  * nodes and thus outstanding btree_updates - redirect @b's
1038  * btree_updates to point to this btree_update:
1039  */
1040 static void bch2_btree_interior_update_will_free_node(struct btree_update *as,
1041 						      struct btree *b)
1042 {
1043 	struct bch_fs *c = as->c;
1044 	struct btree_update *p, *n;
1045 	struct btree_write *w;
1046 
1047 	set_btree_node_dying(b);
1048 
1049 	if (btree_node_fake(b))
1050 		return;
1051 
1052 	mutex_lock(&c->btree_interior_update_lock);
1053 
1054 	/*
1055 	 * Does this node have any btree_update operations preventing
1056 	 * it from being written?
1057 	 *
1058 	 * If so, redirect them to point to this btree_update: we can
1059 	 * write out our new nodes, but we won't make them visible until those
1060 	 * operations complete
1061 	 */
1062 	list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
1063 		list_del_init(&p->write_blocked_list);
1064 		btree_update_reparent(as, p);
1065 
1066 		/*
1067 		 * for flush_held_btree_writes() waiting on updates to flush or
1068 		 * nodes to be writeable:
1069 		 */
1070 		closure_wake_up(&c->btree_interior_update_wait);
1071 	}
1072 
1073 	clear_btree_node_dirty_acct(c, b);
1074 	clear_btree_node_need_write(b);
1075 	clear_btree_node_write_blocked(b);
1076 
1077 	/*
1078 	 * Does this node have unwritten data that has a pin on the journal?
1079 	 *
1080 	 * If so, transfer that pin to the btree_update operation -
1081 	 * note that if we're freeing multiple nodes, we only need to keep the
1082 	 * oldest pin of any of the nodes we're freeing. We'll release the pin
1083 	 * when the new nodes are persistent and reachable on disk:
1084 	 */
1085 	w = btree_current_write(b);
1086 	bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal,
1087 			      bch2_btree_update_will_free_node_journal_pin_flush);
1088 	bch2_journal_pin_drop(&c->journal, &w->journal);
1089 
1090 	w = btree_prev_write(b);
1091 	bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal,
1092 			      bch2_btree_update_will_free_node_journal_pin_flush);
1093 	bch2_journal_pin_drop(&c->journal, &w->journal);
1094 
1095 	mutex_unlock(&c->btree_interior_update_lock);
1096 
1097 	/*
1098 	 * Is this a node that isn't reachable on disk yet?
1099 	 *
1100 	 * Nodes that aren't reachable yet have writes blocked until they're
1101 	 * reachable - now that we've cancelled any pending writes and moved
1102 	 * things waiting on that write to wait on this update, we can drop this
1103 	 * node from the list of nodes that the other update is making
1104 	 * reachable, prior to freeing it:
1105 	 */
1106 	btree_update_drop_new_node(c, b);
1107 
1108 	btree_update_add_key(as, &as->old_keys, b);
1109 
1110 	as->old_nodes[as->nr_old_nodes] = b;
1111 	as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq;
1112 	as->nr_old_nodes++;
1113 }
1114 
1115 static void bch2_btree_update_done(struct btree_update *as, struct btree_trans *trans)
1116 {
1117 	struct bch_fs *c = as->c;
1118 	u64 start_time = as->start_time;
1119 
1120 	BUG_ON(as->mode == BTREE_UPDATE_none);
1121 
1122 	if (as->took_gc_lock)
1123 		up_read(&as->c->gc_lock);
1124 	as->took_gc_lock = false;
1125 
1126 	bch2_btree_reserve_put(as, trans);
1127 
1128 	continue_at(&as->cl, btree_update_set_nodes_written,
1129 		    as->c->btree_interior_update_worker);
1130 
1131 	bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_foreground],
1132 			       start_time);
1133 }
1134 
1135 static struct btree_update *
1136 bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
1137 			unsigned level_start, bool split, unsigned flags)
1138 {
1139 	struct bch_fs *c = trans->c;
1140 	struct btree_update *as;
1141 	u64 start_time = local_clock();
1142 	int disk_res_flags = (flags & BCH_TRANS_COMMIT_no_enospc)
1143 		? BCH_DISK_RESERVATION_NOFAIL : 0;
1144 	unsigned nr_nodes[2] = { 0, 0 };
1145 	unsigned level_end = level_start;
1146 	enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
1147 	int ret = 0;
1148 	u32 restart_count = trans->restart_count;
1149 
1150 	BUG_ON(!path->should_be_locked);
1151 
1152 	if (watermark == BCH_WATERMARK_copygc)
1153 		watermark = BCH_WATERMARK_btree_copygc;
1154 	if (watermark < BCH_WATERMARK_btree)
1155 		watermark = BCH_WATERMARK_btree;
1156 
1157 	flags &= ~BCH_WATERMARK_MASK;
1158 	flags |= watermark;
1159 
1160 	if (watermark < BCH_WATERMARK_reclaim &&
1161 	    test_bit(JOURNAL_space_low, &c->journal.flags)) {
1162 		if (flags & BCH_TRANS_COMMIT_journal_reclaim)
1163 			return ERR_PTR(-BCH_ERR_journal_reclaim_would_deadlock);
1164 
1165 		ret = drop_locks_do(trans,
1166 			({ wait_event(c->journal.wait, !test_bit(JOURNAL_space_low, &c->journal.flags)); 0; }));
1167 		if (ret)
1168 			return ERR_PTR(ret);
1169 	}
1170 
1171 	while (1) {
1172 		nr_nodes[!!level_end] += 1 + split;
1173 		level_end++;
1174 
1175 		ret = bch2_btree_path_upgrade(trans, path, level_end + 1);
1176 		if (ret)
1177 			return ERR_PTR(ret);
1178 
1179 		if (!btree_path_node(path, level_end)) {
1180 			/* Allocating new root? */
1181 			nr_nodes[1] += split;
1182 			level_end = BTREE_MAX_DEPTH;
1183 			break;
1184 		}
1185 
1186 		/*
1187 		 * Always check for space for two keys, even if we won't have to
1188 		 * split at prior level - it might have been a merge instead:
1189 		 */
1190 		if (bch2_btree_node_insert_fits(path->l[level_end].b,
1191 						BKEY_BTREE_PTR_U64s_MAX * 2))
1192 			break;
1193 
1194 		split = path->l[level_end].b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c);
1195 	}
1196 
1197 	if (!down_read_trylock(&c->gc_lock)) {
1198 		ret = drop_locks_do(trans, (down_read(&c->gc_lock), 0));
1199 		if (ret) {
1200 			up_read(&c->gc_lock);
1201 			return ERR_PTR(ret);
1202 		}
1203 	}
1204 
1205 	as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOFS);
1206 	memset(as, 0, sizeof(*as));
1207 	closure_init(&as->cl, NULL);
1208 	as->c			= c;
1209 	as->start_time		= start_time;
1210 	as->ip_started		= _RET_IP_;
1211 	as->mode		= BTREE_UPDATE_none;
1212 	as->flags		= flags;
1213 	as->took_gc_lock	= true;
1214 	as->btree_id		= path->btree_id;
1215 	as->update_level_start	= level_start;
1216 	as->update_level_end	= level_end;
1217 	INIT_LIST_HEAD(&as->list);
1218 	INIT_LIST_HEAD(&as->unwritten_list);
1219 	INIT_LIST_HEAD(&as->write_blocked_list);
1220 	bch2_keylist_init(&as->old_keys, as->_old_keys);
1221 	bch2_keylist_init(&as->new_keys, as->_new_keys);
1222 	bch2_keylist_init(&as->parent_keys, as->inline_keys);
1223 
1224 	mutex_lock(&c->btree_interior_update_lock);
1225 	list_add_tail(&as->list, &c->btree_interior_update_list);
1226 	mutex_unlock(&c->btree_interior_update_lock);
1227 
1228 	/*
1229 	 * We don't want to allocate if we're in an error state, that can cause
1230 	 * deadlock on emergency shutdown due to open buckets getting stuck in
1231 	 * the btree_reserve_cache after allocator shutdown has cleared it out.
1232 	 * This check needs to come after adding us to the btree_interior_update
1233 	 * list but before calling bch2_btree_reserve_get, to synchronize with
1234 	 * __bch2_fs_read_only().
1235 	 */
1236 	ret = bch2_journal_error(&c->journal);
1237 	if (ret)
1238 		goto err;
1239 
1240 	ret = bch2_disk_reservation_get(c, &as->disk_res,
1241 			(nr_nodes[0] + nr_nodes[1]) * btree_sectors(c),
1242 			c->opts.metadata_replicas,
1243 			disk_res_flags);
1244 	if (ret)
1245 		goto err;
1246 
1247 	ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, NULL);
1248 	if (bch2_err_matches(ret, ENOSPC) ||
1249 	    bch2_err_matches(ret, ENOMEM)) {
1250 		struct closure cl;
1251 
1252 		/*
1253 		 * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
1254 		 * flag
1255 		 */
1256 		if (bch2_err_matches(ret, ENOSPC) &&
1257 		    (flags & BCH_TRANS_COMMIT_journal_reclaim) &&
1258 		    watermark < BCH_WATERMARK_reclaim) {
1259 			ret = -BCH_ERR_journal_reclaim_would_deadlock;
1260 			goto err;
1261 		}
1262 
1263 		closure_init_stack(&cl);
1264 
1265 		do {
1266 			ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, &cl);
1267 
1268 			bch2_trans_unlock(trans);
1269 			bch2_wait_on_allocator(c, &cl);
1270 		} while (bch2_err_matches(ret, BCH_ERR_operation_blocked));
1271 	}
1272 
1273 	if (ret) {
1274 		trace_and_count(c, btree_reserve_get_fail, trans->fn,
1275 				_RET_IP_, nr_nodes[0] + nr_nodes[1], ret);
1276 		goto err;
1277 	}
1278 
1279 	ret = bch2_trans_relock(trans);
1280 	if (ret)
1281 		goto err;
1282 
1283 	bch2_trans_verify_not_restarted(trans, restart_count);
1284 	return as;
1285 err:
1286 	bch2_btree_update_free(as, trans);
1287 	if (!bch2_err_matches(ret, ENOSPC) &&
1288 	    !bch2_err_matches(ret, EROFS) &&
1289 	    ret != -BCH_ERR_journal_reclaim_would_deadlock)
1290 		bch_err_fn_ratelimited(c, ret);
1291 	return ERR_PTR(ret);
1292 }
1293 
1294 /* Btree root updates: */
1295 
1296 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1297 {
1298 	/* Root nodes cannot be reaped */
1299 	mutex_lock(&c->btree_cache.lock);
1300 	list_del_init(&b->list);
1301 	mutex_unlock(&c->btree_cache.lock);
1302 
1303 	mutex_lock(&c->btree_root_lock);
1304 	bch2_btree_id_root(c, b->c.btree_id)->b = b;
1305 	mutex_unlock(&c->btree_root_lock);
1306 
1307 	bch2_recalc_btree_reserve(c);
1308 }
1309 
1310 static int bch2_btree_set_root(struct btree_update *as,
1311 			       struct btree_trans *trans,
1312 			       struct btree_path *path,
1313 			       struct btree *b,
1314 			       bool nofail)
1315 {
1316 	struct bch_fs *c = as->c;
1317 
1318 	trace_and_count(c, btree_node_set_root, trans, b);
1319 
1320 	struct btree *old = btree_node_root(c, b);
1321 
1322 	/*
1323 	 * Ensure no one is using the old root while we switch to the
1324 	 * new root:
1325 	 */
1326 	if (nofail) {
1327 		bch2_btree_node_lock_write_nofail(trans, path, &old->c);
1328 	} else {
1329 		int ret = bch2_btree_node_lock_write(trans, path, &old->c);
1330 		if (ret)
1331 			return ret;
1332 	}
1333 
1334 	bch2_btree_set_root_inmem(c, b);
1335 
1336 	btree_update_updated_root(as, b);
1337 
1338 	/*
1339 	 * Unlock old root after new root is visible:
1340 	 *
1341 	 * The new root isn't persistent, but that's ok: we still have
1342 	 * an intent lock on the new root, and any updates that would
1343 	 * depend on the new root would have to update the new root.
1344 	 */
1345 	bch2_btree_node_unlock_write(trans, path, old);
1346 	return 0;
1347 }
1348 
1349 /* Interior node updates: */
1350 
1351 static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
1352 					struct btree_trans *trans,
1353 					struct btree_path *path,
1354 					struct btree *b,
1355 					struct btree_node_iter *node_iter,
1356 					struct bkey_i *insert)
1357 {
1358 	struct bch_fs *c = as->c;
1359 	struct bkey_packed *k;
1360 	struct printbuf buf = PRINTBUF;
1361 	unsigned long old, new;
1362 
1363 	BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1364 	       !btree_ptr_sectors_written(bkey_i_to_s_c(insert)));
1365 
1366 	if (unlikely(!test_bit(JOURNAL_replay_done, &c->journal.flags)))
1367 		bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
1368 
1369 	if (bch2_bkey_validate(c, bkey_i_to_s_c(insert),
1370 			      btree_node_type(b), BCH_VALIDATE_write) ?:
1371 	    bch2_bkey_in_btree_node(c, b, bkey_i_to_s_c(insert), BCH_VALIDATE_write)) {
1372 		bch2_fs_inconsistent(c, "%s: inserting invalid bkey", __func__);
1373 		dump_stack();
1374 	}
1375 
1376 	BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1377 	       ARRAY_SIZE(as->journal_entries));
1378 
1379 	as->journal_u64s +=
1380 		journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1381 				  BCH_JSET_ENTRY_btree_keys,
1382 				  b->c.btree_id, b->c.level,
1383 				  insert, insert->k.u64s);
1384 
1385 	while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1386 	       bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1387 		bch2_btree_node_iter_advance(node_iter, b);
1388 
1389 	bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1390 	set_btree_node_dirty_acct(c, b);
1391 
1392 	old = READ_ONCE(b->flags);
1393 	do {
1394 		new = old;
1395 
1396 		new &= ~BTREE_WRITE_TYPE_MASK;
1397 		new |= BTREE_WRITE_interior;
1398 		new |= 1 << BTREE_NODE_need_write;
1399 	} while (!try_cmpxchg(&b->flags, &old, new));
1400 
1401 	printbuf_exit(&buf);
1402 }
1403 
1404 static void
1405 bch2_btree_insert_keys_interior(struct btree_update *as,
1406 				struct btree_trans *trans,
1407 				struct btree_path *path,
1408 				struct btree *b,
1409 				struct btree_node_iter node_iter,
1410 				struct keylist *keys)
1411 {
1412 	struct bkey_i *insert = bch2_keylist_front(keys);
1413 	struct bkey_packed *k;
1414 
1415 	BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1416 
1417 	while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1418 	       (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1419 		;
1420 
1421 	while (!bch2_keylist_empty(keys)) {
1422 		insert = bch2_keylist_front(keys);
1423 
1424 		if (bpos_gt(insert->k.p, b->key.k.p))
1425 			break;
1426 
1427 		bch2_insert_fixup_btree_ptr(as, trans, path, b, &node_iter, insert);
1428 		bch2_keylist_pop_front(keys);
1429 	}
1430 }
1431 
1432 static bool key_deleted_in_insert(struct keylist *insert_keys, struct bpos pos)
1433 {
1434 	if (insert_keys)
1435 		for_each_keylist_key(insert_keys, k)
1436 			if (bkey_deleted(&k->k) && bpos_eq(k->k.p, pos))
1437 				return true;
1438 	return false;
1439 }
1440 
1441 /*
1442  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1443  * node)
1444  */
1445 static void __btree_split_node(struct btree_update *as,
1446 			       struct btree_trans *trans,
1447 			       struct btree *b,
1448 			       struct btree *n[2],
1449 			       struct keylist *insert_keys)
1450 {
1451 	struct bkey_packed *k;
1452 	struct bpos n1_pos = POS_MIN;
1453 	struct btree_node_iter iter;
1454 	struct bset *bsets[2];
1455 	struct bkey_format_state format[2];
1456 	struct bkey_packed *out[2];
1457 	struct bkey uk;
1458 	unsigned u64s, n1_u64s = (b->nr.live_u64s * 3) / 5;
1459 	struct { unsigned nr_keys, val_u64s; } nr_keys[2];
1460 	int i;
1461 
1462 	memset(&nr_keys, 0, sizeof(nr_keys));
1463 
1464 	for (i = 0; i < 2; i++) {
1465 		BUG_ON(n[i]->nsets != 1);
1466 
1467 		bsets[i] = btree_bset_first(n[i]);
1468 		out[i] = bsets[i]->start;
1469 
1470 		SET_BTREE_NODE_SEQ(n[i]->data, BTREE_NODE_SEQ(b->data) + 1);
1471 		bch2_bkey_format_init(&format[i]);
1472 	}
1473 
1474 	u64s = 0;
1475 	for_each_btree_node_key(b, k, &iter) {
1476 		if (bkey_deleted(k))
1477 			continue;
1478 
1479 		uk = bkey_unpack_key(b, k);
1480 
1481 		if (b->c.level &&
1482 		    u64s < n1_u64s &&
1483 		    u64s + k->u64s >= n1_u64s &&
1484 		    (bch2_key_deleted_in_journal(trans, b->c.btree_id, b->c.level, uk.p) ||
1485 		     key_deleted_in_insert(insert_keys, uk.p)))
1486 			n1_u64s += k->u64s;
1487 
1488 		i = u64s >= n1_u64s;
1489 		u64s += k->u64s;
1490 		if (!i)
1491 			n1_pos = uk.p;
1492 		bch2_bkey_format_add_key(&format[i], &uk);
1493 
1494 		nr_keys[i].nr_keys++;
1495 		nr_keys[i].val_u64s += bkeyp_val_u64s(&b->format, k);
1496 	}
1497 
1498 	btree_set_min(n[0], b->data->min_key);
1499 	btree_set_max(n[0], n1_pos);
1500 	btree_set_min(n[1], bpos_successor(n1_pos));
1501 	btree_set_max(n[1], b->data->max_key);
1502 
1503 	for (i = 0; i < 2; i++) {
1504 		bch2_bkey_format_add_pos(&format[i], n[i]->data->min_key);
1505 		bch2_bkey_format_add_pos(&format[i], n[i]->data->max_key);
1506 
1507 		n[i]->data->format = bch2_bkey_format_done(&format[i]);
1508 
1509 		unsigned u64s = nr_keys[i].nr_keys * n[i]->data->format.key_u64s +
1510 			nr_keys[i].val_u64s;
1511 		if (__vstruct_bytes(struct btree_node, u64s) > btree_buf_bytes(b))
1512 			n[i]->data->format = b->format;
1513 
1514 		btree_node_set_format(n[i], n[i]->data->format);
1515 	}
1516 
1517 	u64s = 0;
1518 	for_each_btree_node_key(b, k, &iter) {
1519 		if (bkey_deleted(k))
1520 			continue;
1521 
1522 		i = u64s >= n1_u64s;
1523 		u64s += k->u64s;
1524 
1525 		if (bch2_bkey_transform(&n[i]->format, out[i], bkey_packed(k)
1526 					? &b->format: &bch2_bkey_format_current, k))
1527 			out[i]->format = KEY_FORMAT_LOCAL_BTREE;
1528 		else
1529 			bch2_bkey_unpack(b, (void *) out[i], k);
1530 
1531 		out[i]->needs_whiteout = false;
1532 
1533 		btree_keys_account_key_add(&n[i]->nr, 0, out[i]);
1534 		out[i] = bkey_p_next(out[i]);
1535 	}
1536 
1537 	for (i = 0; i < 2; i++) {
1538 		bsets[i]->u64s = cpu_to_le16((u64 *) out[i] - bsets[i]->_data);
1539 
1540 		BUG_ON(!bsets[i]->u64s);
1541 
1542 		set_btree_bset_end(n[i], n[i]->set);
1543 
1544 		btree_node_reset_sib_u64s(n[i]);
1545 
1546 		bch2_verify_btree_nr_keys(n[i]);
1547 
1548 		BUG_ON(bch2_btree_node_check_topology(trans, n[i]));
1549 	}
1550 }
1551 
1552 /*
1553  * For updates to interior nodes, we've got to do the insert before we split
1554  * because the stuff we're inserting has to be inserted atomically. Post split,
1555  * the keys might have to go in different nodes and the split would no longer be
1556  * atomic.
1557  *
1558  * Worse, if the insert is from btree node coalescing, if we do the insert after
1559  * we do the split (and pick the pivot) - the pivot we pick might be between
1560  * nodes that were coalesced, and thus in the middle of a child node post
1561  * coalescing:
1562  */
1563 static void btree_split_insert_keys(struct btree_update *as,
1564 				    struct btree_trans *trans,
1565 				    btree_path_idx_t path_idx,
1566 				    struct btree *b,
1567 				    struct keylist *keys)
1568 {
1569 	struct btree_path *path = trans->paths + path_idx;
1570 
1571 	if (!bch2_keylist_empty(keys) &&
1572 	    bpos_le(bch2_keylist_front(keys)->k.p, b->data->max_key)) {
1573 		struct btree_node_iter node_iter;
1574 
1575 		bch2_btree_node_iter_init(&node_iter, b, &bch2_keylist_front(keys)->k.p);
1576 
1577 		bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1578 
1579 		BUG_ON(bch2_btree_node_check_topology(trans, b));
1580 	}
1581 }
1582 
1583 static int btree_split(struct btree_update *as, struct btree_trans *trans,
1584 		       btree_path_idx_t path, struct btree *b,
1585 		       struct keylist *keys)
1586 {
1587 	struct bch_fs *c = as->c;
1588 	struct btree *parent = btree_node_parent(trans->paths + path, b);
1589 	struct btree *n1, *n2 = NULL, *n3 = NULL;
1590 	btree_path_idx_t path1 = 0, path2 = 0;
1591 	u64 start_time = local_clock();
1592 	int ret = 0;
1593 
1594 	bch2_verify_btree_nr_keys(b);
1595 	BUG_ON(!parent && (b != btree_node_root(c, b)));
1596 	BUG_ON(parent && !btree_node_intent_locked(trans->paths + path, b->c.level + 1));
1597 
1598 	ret = bch2_btree_node_check_topology(trans, b);
1599 	if (ret)
1600 		return ret;
1601 
1602 	bch2_btree_interior_update_will_free_node(as, b);
1603 
1604 	if (b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c)) {
1605 		struct btree *n[2];
1606 
1607 		trace_and_count(c, btree_node_split, trans, b);
1608 
1609 		n[0] = n1 = bch2_btree_node_alloc(as, trans, b->c.level);
1610 		n[1] = n2 = bch2_btree_node_alloc(as, trans, b->c.level);
1611 
1612 		__btree_split_node(as, trans, b, n, keys);
1613 
1614 		if (keys) {
1615 			btree_split_insert_keys(as, trans, path, n1, keys);
1616 			btree_split_insert_keys(as, trans, path, n2, keys);
1617 			BUG_ON(!bch2_keylist_empty(keys));
1618 		}
1619 
1620 		bch2_btree_build_aux_trees(n2);
1621 		bch2_btree_build_aux_trees(n1);
1622 
1623 		bch2_btree_update_add_new_node(as, n1);
1624 		bch2_btree_update_add_new_node(as, n2);
1625 		six_unlock_write(&n2->c.lock);
1626 		six_unlock_write(&n1->c.lock);
1627 
1628 		path1 = bch2_path_get_unlocked_mut(trans, as->btree_id, n1->c.level, n1->key.k.p);
1629 		six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1630 		mark_btree_node_locked(trans, trans->paths + path1, n1->c.level, BTREE_NODE_INTENT_LOCKED);
1631 		bch2_btree_path_level_init(trans, trans->paths + path1, n1);
1632 
1633 		path2 = bch2_path_get_unlocked_mut(trans, as->btree_id, n2->c.level, n2->key.k.p);
1634 		six_lock_increment(&n2->c.lock, SIX_LOCK_intent);
1635 		mark_btree_node_locked(trans, trans->paths + path2, n2->c.level, BTREE_NODE_INTENT_LOCKED);
1636 		bch2_btree_path_level_init(trans, trans->paths + path2, n2);
1637 
1638 		/*
1639 		 * Note that on recursive parent_keys == keys, so we
1640 		 * can't start adding new keys to parent_keys before emptying it
1641 		 * out (which we did with btree_split_insert_keys() above)
1642 		 */
1643 		bch2_keylist_add(&as->parent_keys, &n1->key);
1644 		bch2_keylist_add(&as->parent_keys, &n2->key);
1645 
1646 		if (!parent) {
1647 			/* Depth increases, make a new root */
1648 			n3 = __btree_root_alloc(as, trans, b->c.level + 1);
1649 
1650 			bch2_btree_update_add_new_node(as, n3);
1651 			six_unlock_write(&n3->c.lock);
1652 
1653 			trans->paths[path2].locks_want++;
1654 			BUG_ON(btree_node_locked(trans->paths + path2, n3->c.level));
1655 			six_lock_increment(&n3->c.lock, SIX_LOCK_intent);
1656 			mark_btree_node_locked(trans, trans->paths + path2, n3->c.level, BTREE_NODE_INTENT_LOCKED);
1657 			bch2_btree_path_level_init(trans, trans->paths + path2, n3);
1658 
1659 			n3->sib_u64s[0] = U16_MAX;
1660 			n3->sib_u64s[1] = U16_MAX;
1661 
1662 			btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1663 		}
1664 	} else {
1665 		trace_and_count(c, btree_node_compact, trans, b);
1666 
1667 		n1 = bch2_btree_node_alloc_replacement(as, trans, b);
1668 
1669 		if (keys) {
1670 			btree_split_insert_keys(as, trans, path, n1, keys);
1671 			BUG_ON(!bch2_keylist_empty(keys));
1672 		}
1673 
1674 		bch2_btree_build_aux_trees(n1);
1675 		bch2_btree_update_add_new_node(as, n1);
1676 		six_unlock_write(&n1->c.lock);
1677 
1678 		path1 = bch2_path_get_unlocked_mut(trans, as->btree_id, n1->c.level, n1->key.k.p);
1679 		six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1680 		mark_btree_node_locked(trans, trans->paths + path1, n1->c.level, BTREE_NODE_INTENT_LOCKED);
1681 		bch2_btree_path_level_init(trans, trans->paths + path1, n1);
1682 
1683 		if (parent)
1684 			bch2_keylist_add(&as->parent_keys, &n1->key);
1685 	}
1686 
1687 	/* New nodes all written, now make them visible: */
1688 
1689 	if (parent) {
1690 		/* Split a non root node */
1691 		ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys);
1692 	} else if (n3) {
1693 		ret = bch2_btree_set_root(as, trans, trans->paths + path, n3, false);
1694 	} else {
1695 		/* Root filled up but didn't need to be split */
1696 		ret = bch2_btree_set_root(as, trans, trans->paths + path, n1, false);
1697 	}
1698 
1699 	if (ret)
1700 		goto err;
1701 
1702 	if (n3) {
1703 		bch2_btree_update_get_open_buckets(as, n3);
1704 		bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
1705 	}
1706 	if (n2) {
1707 		bch2_btree_update_get_open_buckets(as, n2);
1708 		bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
1709 	}
1710 	bch2_btree_update_get_open_buckets(as, n1);
1711 	bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1712 
1713 	/*
1714 	 * The old node must be freed (in memory) _before_ unlocking the new
1715 	 * nodes - else another thread could re-acquire a read lock on the old
1716 	 * node after another thread has locked and updated the new node, thus
1717 	 * seeing stale data:
1718 	 */
1719 	bch2_btree_node_free_inmem(trans, trans->paths + path, b);
1720 
1721 	if (n3)
1722 		bch2_trans_node_add(trans, trans->paths + path, n3);
1723 	if (n2)
1724 		bch2_trans_node_add(trans, trans->paths + path2, n2);
1725 	bch2_trans_node_add(trans, trans->paths + path1, n1);
1726 
1727 	if (n3)
1728 		six_unlock_intent(&n3->c.lock);
1729 	if (n2)
1730 		six_unlock_intent(&n2->c.lock);
1731 	six_unlock_intent(&n1->c.lock);
1732 out:
1733 	if (path2) {
1734 		__bch2_btree_path_unlock(trans, trans->paths + path2);
1735 		bch2_path_put(trans, path2, true);
1736 	}
1737 	if (path1) {
1738 		__bch2_btree_path_unlock(trans, trans->paths + path1);
1739 		bch2_path_put(trans, path1, true);
1740 	}
1741 
1742 	bch2_trans_verify_locks(trans);
1743 
1744 	bch2_time_stats_update(&c->times[n2
1745 			       ? BCH_TIME_btree_node_split
1746 			       : BCH_TIME_btree_node_compact],
1747 			       start_time);
1748 	return ret;
1749 err:
1750 	if (n3)
1751 		bch2_btree_node_free_never_used(as, trans, n3);
1752 	if (n2)
1753 		bch2_btree_node_free_never_used(as, trans, n2);
1754 	bch2_btree_node_free_never_used(as, trans, n1);
1755 	goto out;
1756 }
1757 
1758 /**
1759  * bch2_btree_insert_node - insert bkeys into a given btree node
1760  *
1761  * @as:			btree_update object
1762  * @trans:		btree_trans object
1763  * @path_idx:		path that points to current node
1764  * @b:			node to insert keys into
1765  * @keys:		list of keys to insert
1766  *
1767  * Returns: 0 on success, typically transaction restart error on failure
1768  *
1769  * Inserts as many keys as it can into a given btree node, splitting it if full.
1770  * If a split occurred, this function will return early. This can only happen
1771  * for leaf nodes -- inserts into interior nodes have to be atomic.
1772  */
1773 static int bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1774 				  btree_path_idx_t path_idx, struct btree *b,
1775 				  struct keylist *keys)
1776 {
1777 	struct bch_fs *c = as->c;
1778 	struct btree_path *path = trans->paths + path_idx, *linked;
1779 	unsigned i;
1780 	int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1781 	int old_live_u64s = b->nr.live_u64s;
1782 	int live_u64s_added, u64s_added;
1783 	int ret;
1784 
1785 	lockdep_assert_held(&c->gc_lock);
1786 	BUG_ON(!btree_node_intent_locked(path, b->c.level));
1787 	BUG_ON(!b->c.level);
1788 	BUG_ON(!as || as->b);
1789 	bch2_verify_keylist_sorted(keys);
1790 
1791 	ret = bch2_btree_node_lock_write(trans, path, &b->c);
1792 	if (ret)
1793 		return ret;
1794 
1795 	bch2_btree_node_prep_for_write(trans, path, b);
1796 
1797 	if (!bch2_btree_node_insert_fits(b, bch2_keylist_u64s(keys))) {
1798 		bch2_btree_node_unlock_write(trans, path, b);
1799 		goto split;
1800 	}
1801 
1802 	ret = bch2_btree_node_check_topology(trans, b);
1803 	if (ret) {
1804 		bch2_btree_node_unlock_write(trans, path, b);
1805 		return ret;
1806 	}
1807 
1808 	bch2_btree_insert_keys_interior(as, trans, path, b,
1809 					path->l[b->c.level].iter, keys);
1810 
1811 	trans_for_each_path_with_node(trans, b, linked, i)
1812 		bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1813 
1814 	bch2_trans_verify_paths(trans);
1815 
1816 	live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1817 	u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1818 
1819 	if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1820 		b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1821 	if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1822 		b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1823 
1824 	if (u64s_added > live_u64s_added &&
1825 	    bch2_maybe_compact_whiteouts(c, b))
1826 		bch2_trans_node_reinit_iter(trans, b);
1827 
1828 	btree_update_updated_node(as, b);
1829 	bch2_btree_node_unlock_write(trans, path, b);
1830 
1831 	BUG_ON(bch2_btree_node_check_topology(trans, b));
1832 	return 0;
1833 split:
1834 	/*
1835 	 * We could attempt to avoid the transaction restart, by calling
1836 	 * bch2_btree_path_upgrade() and allocating more nodes:
1837 	 */
1838 	if (b->c.level >= as->update_level_end) {
1839 		trace_and_count(c, trans_restart_split_race, trans, _THIS_IP_, b);
1840 		return btree_trans_restart(trans, BCH_ERR_transaction_restart_split_race);
1841 	}
1842 
1843 	return btree_split(as, trans, path_idx, b, keys);
1844 }
1845 
1846 int bch2_btree_split_leaf(struct btree_trans *trans,
1847 			  btree_path_idx_t path,
1848 			  unsigned flags)
1849 {
1850 	/* btree_split & merge may both cause paths array to be reallocated */
1851 	struct btree *b = path_l(trans->paths + path)->b;
1852 	struct btree_update *as;
1853 	unsigned l;
1854 	int ret = 0;
1855 
1856 	as = bch2_btree_update_start(trans, trans->paths + path,
1857 				     trans->paths[path].level,
1858 				     true, flags);
1859 	if (IS_ERR(as))
1860 		return PTR_ERR(as);
1861 
1862 	ret = btree_split(as, trans, path, b, NULL);
1863 	if (ret) {
1864 		bch2_btree_update_free(as, trans);
1865 		return ret;
1866 	}
1867 
1868 	bch2_btree_update_done(as, trans);
1869 
1870 	for (l = trans->paths[path].level + 1;
1871 	     btree_node_intent_locked(&trans->paths[path], l) && !ret;
1872 	     l++)
1873 		ret = bch2_foreground_maybe_merge(trans, path, l, flags);
1874 
1875 	return ret;
1876 }
1877 
1878 static void __btree_increase_depth(struct btree_update *as, struct btree_trans *trans,
1879 				   btree_path_idx_t path_idx)
1880 {
1881 	struct bch_fs *c = as->c;
1882 	struct btree_path *path = trans->paths + path_idx;
1883 	struct btree *n, *b = bch2_btree_id_root(c, path->btree_id)->b;
1884 
1885 	BUG_ON(!btree_node_locked(path, b->c.level));
1886 
1887 	n = __btree_root_alloc(as, trans, b->c.level + 1);
1888 
1889 	bch2_btree_update_add_new_node(as, n);
1890 	six_unlock_write(&n->c.lock);
1891 
1892 	path->locks_want++;
1893 	BUG_ON(btree_node_locked(path, n->c.level));
1894 	six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1895 	mark_btree_node_locked(trans, path, n->c.level, BTREE_NODE_INTENT_LOCKED);
1896 	bch2_btree_path_level_init(trans, path, n);
1897 
1898 	n->sib_u64s[0] = U16_MAX;
1899 	n->sib_u64s[1] = U16_MAX;
1900 
1901 	bch2_keylist_add(&as->parent_keys, &b->key);
1902 	btree_split_insert_keys(as, trans, path_idx, n, &as->parent_keys);
1903 
1904 	int ret = bch2_btree_set_root(as, trans, path, n, true);
1905 	BUG_ON(ret);
1906 
1907 	bch2_btree_update_get_open_buckets(as, n);
1908 	bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1909 	bch2_trans_node_add(trans, path, n);
1910 	six_unlock_intent(&n->c.lock);
1911 
1912 	mutex_lock(&c->btree_cache.lock);
1913 	list_add_tail(&b->list, &c->btree_cache.live[btree_node_pinned(b)].list);
1914 	mutex_unlock(&c->btree_cache.lock);
1915 
1916 	bch2_trans_verify_locks(trans);
1917 }
1918 
1919 int bch2_btree_increase_depth(struct btree_trans *trans, btree_path_idx_t path, unsigned flags)
1920 {
1921 	struct bch_fs *c = trans->c;
1922 	struct btree *b = bch2_btree_id_root(c, trans->paths[path].btree_id)->b;
1923 
1924 	if (btree_node_fake(b))
1925 		return bch2_btree_split_leaf(trans, path, flags);
1926 
1927 	struct btree_update *as =
1928 		bch2_btree_update_start(trans, trans->paths + path, b->c.level, true, flags);
1929 	if (IS_ERR(as))
1930 		return PTR_ERR(as);
1931 
1932 	__btree_increase_depth(as, trans, path);
1933 	bch2_btree_update_done(as, trans);
1934 	return 0;
1935 }
1936 
1937 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1938 				  btree_path_idx_t path,
1939 				  unsigned level,
1940 				  unsigned flags,
1941 				  enum btree_node_sibling sib)
1942 {
1943 	struct bch_fs *c = trans->c;
1944 	struct btree_update *as;
1945 	struct bkey_format_state new_s;
1946 	struct bkey_format new_f;
1947 	struct bkey_i delete;
1948 	struct btree *b, *m, *n, *prev, *next, *parent;
1949 	struct bpos sib_pos;
1950 	size_t sib_u64s;
1951 	enum btree_id btree = trans->paths[path].btree_id;
1952 	btree_path_idx_t sib_path = 0, new_path = 0;
1953 	u64 start_time = local_clock();
1954 	int ret = 0;
1955 
1956 	bch2_trans_verify_not_in_restart(trans);
1957 	bch2_trans_verify_not_unlocked(trans);
1958 	BUG_ON(!trans->paths[path].should_be_locked);
1959 	BUG_ON(!btree_node_locked(&trans->paths[path], level));
1960 
1961 	/*
1962 	 * Work around a deadlock caused by the btree write buffer not doing
1963 	 * merges and leaving tons of merges for us to do - we really don't need
1964 	 * to be doing merges at all from the interior update path, and if the
1965 	 * interior update path is generating too many new interior updates we
1966 	 * deadlock:
1967 	 */
1968 	if ((flags & BCH_WATERMARK_MASK) == BCH_WATERMARK_interior_updates)
1969 		return 0;
1970 
1971 	if ((flags & BCH_WATERMARK_MASK) <= BCH_WATERMARK_reclaim) {
1972 		flags &= ~BCH_WATERMARK_MASK;
1973 		flags |= BCH_WATERMARK_btree;
1974 		flags |= BCH_TRANS_COMMIT_journal_reclaim;
1975 	}
1976 
1977 	b = trans->paths[path].l[level].b;
1978 
1979 	if ((sib == btree_prev_sib && bpos_eq(b->data->min_key, POS_MIN)) ||
1980 	    (sib == btree_next_sib && bpos_eq(b->data->max_key, SPOS_MAX))) {
1981 		b->sib_u64s[sib] = U16_MAX;
1982 		return 0;
1983 	}
1984 
1985 	sib_pos = sib == btree_prev_sib
1986 		? bpos_predecessor(b->data->min_key)
1987 		: bpos_successor(b->data->max_key);
1988 
1989 	sib_path = bch2_path_get(trans, btree, sib_pos,
1990 				 U8_MAX, level, BTREE_ITER_intent, _THIS_IP_);
1991 	ret = bch2_btree_path_traverse(trans, sib_path, false);
1992 	if (ret)
1993 		goto err;
1994 
1995 	btree_path_set_should_be_locked(trans, trans->paths + sib_path);
1996 
1997 	m = trans->paths[sib_path].l[level].b;
1998 
1999 	if (btree_node_parent(trans->paths + path, b) !=
2000 	    btree_node_parent(trans->paths + sib_path, m)) {
2001 		b->sib_u64s[sib] = U16_MAX;
2002 		goto out;
2003 	}
2004 
2005 	if (sib == btree_prev_sib) {
2006 		prev = m;
2007 		next = b;
2008 	} else {
2009 		prev = b;
2010 		next = m;
2011 	}
2012 
2013 	if (!bpos_eq(bpos_successor(prev->data->max_key), next->data->min_key)) {
2014 		struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
2015 
2016 		bch2_bpos_to_text(&buf1, prev->data->max_key);
2017 		bch2_bpos_to_text(&buf2, next->data->min_key);
2018 		bch_err(c,
2019 			"%s(): btree topology error:\n"
2020 			"  prev ends at   %s\n"
2021 			"  next starts at %s",
2022 			__func__, buf1.buf, buf2.buf);
2023 		printbuf_exit(&buf1);
2024 		printbuf_exit(&buf2);
2025 		ret = bch2_topology_error(c);
2026 		goto err;
2027 	}
2028 
2029 	bch2_bkey_format_init(&new_s);
2030 	bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
2031 	__bch2_btree_calc_format(&new_s, prev);
2032 	__bch2_btree_calc_format(&new_s, next);
2033 	bch2_bkey_format_add_pos(&new_s, next->data->max_key);
2034 	new_f = bch2_bkey_format_done(&new_s);
2035 
2036 	sib_u64s = btree_node_u64s_with_format(b->nr, &b->format, &new_f) +
2037 		btree_node_u64s_with_format(m->nr, &m->format, &new_f);
2038 
2039 	if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
2040 		sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
2041 		sib_u64s /= 2;
2042 		sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
2043 	}
2044 
2045 	sib_u64s = min(sib_u64s, btree_max_u64s(c));
2046 	sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
2047 	b->sib_u64s[sib] = sib_u64s;
2048 
2049 	if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
2050 		goto out;
2051 
2052 	parent = btree_node_parent(trans->paths + path, b);
2053 	as = bch2_btree_update_start(trans, trans->paths + path, level, false,
2054 				     BCH_TRANS_COMMIT_no_enospc|flags);
2055 	ret = PTR_ERR_OR_ZERO(as);
2056 	if (ret)
2057 		goto err;
2058 
2059 	trace_and_count(c, btree_node_merge, trans, b);
2060 
2061 	bch2_btree_interior_update_will_free_node(as, b);
2062 	bch2_btree_interior_update_will_free_node(as, m);
2063 
2064 	n = bch2_btree_node_alloc(as, trans, b->c.level);
2065 
2066 	SET_BTREE_NODE_SEQ(n->data,
2067 			   max(BTREE_NODE_SEQ(b->data),
2068 			       BTREE_NODE_SEQ(m->data)) + 1);
2069 
2070 	btree_set_min(n, prev->data->min_key);
2071 	btree_set_max(n, next->data->max_key);
2072 
2073 	n->data->format	 = new_f;
2074 	btree_node_set_format(n, new_f);
2075 
2076 	bch2_btree_sort_into(c, n, prev);
2077 	bch2_btree_sort_into(c, n, next);
2078 
2079 	bch2_btree_build_aux_trees(n);
2080 	bch2_btree_update_add_new_node(as, n);
2081 	six_unlock_write(&n->c.lock);
2082 
2083 	new_path = bch2_path_get_unlocked_mut(trans, btree, n->c.level, n->key.k.p);
2084 	six_lock_increment(&n->c.lock, SIX_LOCK_intent);
2085 	mark_btree_node_locked(trans, trans->paths + new_path, n->c.level, BTREE_NODE_INTENT_LOCKED);
2086 	bch2_btree_path_level_init(trans, trans->paths + new_path, n);
2087 
2088 	bkey_init(&delete.k);
2089 	delete.k.p = prev->key.k.p;
2090 	bch2_keylist_add(&as->parent_keys, &delete);
2091 	bch2_keylist_add(&as->parent_keys, &n->key);
2092 
2093 	bch2_trans_verify_paths(trans);
2094 
2095 	ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys);
2096 	if (ret)
2097 		goto err_free_update;
2098 
2099 	bch2_trans_verify_paths(trans);
2100 
2101 	bch2_btree_update_get_open_buckets(as, n);
2102 	bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
2103 
2104 	bch2_btree_node_free_inmem(trans, trans->paths + path, b);
2105 	bch2_btree_node_free_inmem(trans, trans->paths + sib_path, m);
2106 
2107 	bch2_trans_node_add(trans, trans->paths + path, n);
2108 
2109 	bch2_trans_verify_paths(trans);
2110 
2111 	six_unlock_intent(&n->c.lock);
2112 
2113 	bch2_btree_update_done(as, trans);
2114 
2115 	bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
2116 out:
2117 err:
2118 	if (new_path)
2119 		bch2_path_put(trans, new_path, true);
2120 	bch2_path_put(trans, sib_path, true);
2121 	bch2_trans_verify_locks(trans);
2122 	if (ret == -BCH_ERR_journal_reclaim_would_deadlock)
2123 		ret = 0;
2124 	if (!ret)
2125 		ret = bch2_trans_relock(trans);
2126 	return ret;
2127 err_free_update:
2128 	bch2_btree_node_free_never_used(as, trans, n);
2129 	bch2_btree_update_free(as, trans);
2130 	goto out;
2131 }
2132 
2133 int bch2_btree_node_rewrite(struct btree_trans *trans,
2134 			    struct btree_iter *iter,
2135 			    struct btree *b,
2136 			    unsigned flags)
2137 {
2138 	struct bch_fs *c = trans->c;
2139 	struct btree *n, *parent;
2140 	struct btree_update *as;
2141 	btree_path_idx_t new_path = 0;
2142 	int ret;
2143 
2144 	flags |= BCH_TRANS_COMMIT_no_enospc;
2145 
2146 	struct btree_path *path = btree_iter_path(trans, iter);
2147 	parent = btree_node_parent(path, b);
2148 	as = bch2_btree_update_start(trans, path, b->c.level, false, flags);
2149 	ret = PTR_ERR_OR_ZERO(as);
2150 	if (ret)
2151 		goto out;
2152 
2153 	bch2_btree_interior_update_will_free_node(as, b);
2154 
2155 	n = bch2_btree_node_alloc_replacement(as, trans, b);
2156 
2157 	bch2_btree_build_aux_trees(n);
2158 	bch2_btree_update_add_new_node(as, n);
2159 	six_unlock_write(&n->c.lock);
2160 
2161 	new_path = bch2_path_get_unlocked_mut(trans, iter->btree_id, n->c.level, n->key.k.p);
2162 	six_lock_increment(&n->c.lock, SIX_LOCK_intent);
2163 	mark_btree_node_locked(trans, trans->paths + new_path, n->c.level, BTREE_NODE_INTENT_LOCKED);
2164 	bch2_btree_path_level_init(trans, trans->paths + new_path, n);
2165 
2166 	trace_and_count(c, btree_node_rewrite, trans, b);
2167 
2168 	if (parent) {
2169 		bch2_keylist_add(&as->parent_keys, &n->key);
2170 		ret = bch2_btree_insert_node(as, trans, iter->path, parent, &as->parent_keys);
2171 	} else {
2172 		ret = bch2_btree_set_root(as, trans, btree_iter_path(trans, iter), n, false);
2173 	}
2174 
2175 	if (ret)
2176 		goto err;
2177 
2178 	bch2_btree_update_get_open_buckets(as, n);
2179 	bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
2180 
2181 	bch2_btree_node_free_inmem(trans, btree_iter_path(trans, iter), b);
2182 
2183 	bch2_trans_node_add(trans, trans->paths + iter->path, n);
2184 	six_unlock_intent(&n->c.lock);
2185 
2186 	bch2_btree_update_done(as, trans);
2187 out:
2188 	if (new_path)
2189 		bch2_path_put(trans, new_path, true);
2190 	bch2_trans_downgrade(trans);
2191 	return ret;
2192 err:
2193 	bch2_btree_node_free_never_used(as, trans, n);
2194 	bch2_btree_update_free(as, trans);
2195 	goto out;
2196 }
2197 
2198 struct async_btree_rewrite {
2199 	struct bch_fs		*c;
2200 	struct work_struct	work;
2201 	struct list_head	list;
2202 	enum btree_id		btree_id;
2203 	unsigned		level;
2204 	struct bpos		pos;
2205 	__le64			seq;
2206 };
2207 
2208 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
2209 					  struct async_btree_rewrite *a)
2210 {
2211 	struct bch_fs *c = trans->c;
2212 	struct btree_iter iter;
2213 	struct btree *b;
2214 	int ret;
2215 
2216 	bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
2217 				  BTREE_MAX_DEPTH, a->level, 0);
2218 	b = bch2_btree_iter_peek_node(&iter);
2219 	ret = PTR_ERR_OR_ZERO(b);
2220 	if (ret)
2221 		goto out;
2222 
2223 	if (!b || b->data->keys.seq != a->seq) {
2224 		struct printbuf buf = PRINTBUF;
2225 
2226 		if (b)
2227 			bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
2228 		else
2229 			prt_str(&buf, "(null");
2230 		bch_info(c, "%s: node to rewrite not found:, searching for seq %llu, got\n%s",
2231 			 __func__, a->seq, buf.buf);
2232 		printbuf_exit(&buf);
2233 		goto out;
2234 	}
2235 
2236 	ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
2237 out:
2238 	bch2_trans_iter_exit(trans, &iter);
2239 
2240 	return ret;
2241 }
2242 
2243 static void async_btree_node_rewrite_work(struct work_struct *work)
2244 {
2245 	struct async_btree_rewrite *a =
2246 		container_of(work, struct async_btree_rewrite, work);
2247 	struct bch_fs *c = a->c;
2248 
2249 	int ret = bch2_trans_do(c, async_btree_node_rewrite_trans(trans, a));
2250 	bch_err_fn_ratelimited(c, ret);
2251 	bch2_write_ref_put(c, BCH_WRITE_REF_node_rewrite);
2252 	kfree(a);
2253 }
2254 
2255 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
2256 {
2257 	struct async_btree_rewrite *a;
2258 	int ret;
2259 
2260 	a = kmalloc(sizeof(*a), GFP_NOFS);
2261 	if (!a) {
2262 		bch_err(c, "%s: error allocating memory", __func__);
2263 		return;
2264 	}
2265 
2266 	a->c		= c;
2267 	a->btree_id	= b->c.btree_id;
2268 	a->level	= b->c.level;
2269 	a->pos		= b->key.k.p;
2270 	a->seq		= b->data->keys.seq;
2271 	INIT_WORK(&a->work, async_btree_node_rewrite_work);
2272 
2273 	if (unlikely(!test_bit(BCH_FS_may_go_rw, &c->flags))) {
2274 		mutex_lock(&c->pending_node_rewrites_lock);
2275 		list_add(&a->list, &c->pending_node_rewrites);
2276 		mutex_unlock(&c->pending_node_rewrites_lock);
2277 		return;
2278 	}
2279 
2280 	if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_node_rewrite)) {
2281 		if (test_bit(BCH_FS_started, &c->flags)) {
2282 			bch_err(c, "%s: error getting c->writes ref", __func__);
2283 			kfree(a);
2284 			return;
2285 		}
2286 
2287 		ret = bch2_fs_read_write_early(c);
2288 		bch_err_msg(c, ret, "going read-write");
2289 		if (ret) {
2290 			kfree(a);
2291 			return;
2292 		}
2293 
2294 		bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2295 	}
2296 
2297 	queue_work(c->btree_node_rewrite_worker, &a->work);
2298 }
2299 
2300 void bch2_do_pending_node_rewrites(struct bch_fs *c)
2301 {
2302 	struct async_btree_rewrite *a, *n;
2303 
2304 	mutex_lock(&c->pending_node_rewrites_lock);
2305 	list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2306 		list_del(&a->list);
2307 
2308 		bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2309 		queue_work(c->btree_node_rewrite_worker, &a->work);
2310 	}
2311 	mutex_unlock(&c->pending_node_rewrites_lock);
2312 }
2313 
2314 void bch2_free_pending_node_rewrites(struct bch_fs *c)
2315 {
2316 	struct async_btree_rewrite *a, *n;
2317 
2318 	mutex_lock(&c->pending_node_rewrites_lock);
2319 	list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2320 		list_del(&a->list);
2321 
2322 		kfree(a);
2323 	}
2324 	mutex_unlock(&c->pending_node_rewrites_lock);
2325 }
2326 
2327 static int __bch2_btree_node_update_key(struct btree_trans *trans,
2328 					struct btree_iter *iter,
2329 					struct btree *b, struct btree *new_hash,
2330 					struct bkey_i *new_key,
2331 					unsigned commit_flags,
2332 					bool skip_triggers)
2333 {
2334 	struct bch_fs *c = trans->c;
2335 	struct btree_iter iter2 = { NULL };
2336 	struct btree *parent;
2337 	int ret;
2338 
2339 	if (!skip_triggers) {
2340 		ret   = bch2_key_trigger_old(trans, b->c.btree_id, b->c.level + 1,
2341 					     bkey_i_to_s_c(&b->key),
2342 					     BTREE_TRIGGER_transactional) ?:
2343 			bch2_key_trigger_new(trans, b->c.btree_id, b->c.level + 1,
2344 					     bkey_i_to_s(new_key),
2345 					     BTREE_TRIGGER_transactional);
2346 		if (ret)
2347 			return ret;
2348 	}
2349 
2350 	if (new_hash) {
2351 		bkey_copy(&new_hash->key, new_key);
2352 		ret = bch2_btree_node_hash_insert(&c->btree_cache,
2353 				new_hash, b->c.level, b->c.btree_id);
2354 		BUG_ON(ret);
2355 	}
2356 
2357 	parent = btree_node_parent(btree_iter_path(trans, iter), b);
2358 	if (parent) {
2359 		bch2_trans_copy_iter(&iter2, iter);
2360 
2361 		iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
2362 				iter2.flags & BTREE_ITER_intent,
2363 				_THIS_IP_);
2364 
2365 		struct btree_path *path2 = btree_iter_path(trans, &iter2);
2366 		BUG_ON(path2->level != b->c.level);
2367 		BUG_ON(!bpos_eq(path2->pos, new_key->k.p));
2368 
2369 		btree_path_set_level_up(trans, path2);
2370 
2371 		trans->paths_sorted = false;
2372 
2373 		ret   = bch2_btree_iter_traverse(&iter2) ?:
2374 			bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_norun);
2375 		if (ret)
2376 			goto err;
2377 	} else {
2378 		BUG_ON(btree_node_root(c, b) != b);
2379 
2380 		struct jset_entry *e = bch2_trans_jset_entry_alloc(trans,
2381 				       jset_u64s(new_key->k.u64s));
2382 		ret = PTR_ERR_OR_ZERO(e);
2383 		if (ret)
2384 			return ret;
2385 
2386 		journal_entry_set(e,
2387 				  BCH_JSET_ENTRY_btree_root,
2388 				  b->c.btree_id, b->c.level,
2389 				  new_key, new_key->k.u64s);
2390 	}
2391 
2392 	ret = bch2_trans_commit(trans, NULL, NULL, commit_flags);
2393 	if (ret)
2394 		goto err;
2395 
2396 	bch2_btree_node_lock_write_nofail(trans, btree_iter_path(trans, iter), &b->c);
2397 
2398 	if (new_hash) {
2399 		mutex_lock(&c->btree_cache.lock);
2400 		bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
2401 
2402 		__bch2_btree_node_hash_remove(&c->btree_cache, b);
2403 
2404 		bkey_copy(&b->key, new_key);
2405 		ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
2406 		BUG_ON(ret);
2407 		mutex_unlock(&c->btree_cache.lock);
2408 	} else {
2409 		bkey_copy(&b->key, new_key);
2410 	}
2411 
2412 	bch2_btree_node_unlock_write(trans, btree_iter_path(trans, iter), b);
2413 out:
2414 	bch2_trans_iter_exit(trans, &iter2);
2415 	return ret;
2416 err:
2417 	if (new_hash) {
2418 		mutex_lock(&c->btree_cache.lock);
2419 		bch2_btree_node_hash_remove(&c->btree_cache, b);
2420 		mutex_unlock(&c->btree_cache.lock);
2421 	}
2422 	goto out;
2423 }
2424 
2425 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
2426 			       struct btree *b, struct bkey_i *new_key,
2427 			       unsigned commit_flags, bool skip_triggers)
2428 {
2429 	struct bch_fs *c = trans->c;
2430 	struct btree *new_hash = NULL;
2431 	struct btree_path *path = btree_iter_path(trans, iter);
2432 	struct closure cl;
2433 	int ret = 0;
2434 
2435 	ret = bch2_btree_path_upgrade(trans, path, b->c.level + 1);
2436 	if (ret)
2437 		return ret;
2438 
2439 	closure_init_stack(&cl);
2440 
2441 	/*
2442 	 * check btree_ptr_hash_val() after @b is locked by
2443 	 * btree_iter_traverse():
2444 	 */
2445 	if (btree_ptr_hash_val(new_key) != b->hash_val) {
2446 		ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
2447 		if (ret) {
2448 			ret = drop_locks_do(trans, (closure_sync(&cl), 0));
2449 			if (ret)
2450 				return ret;
2451 		}
2452 
2453 		new_hash = bch2_btree_node_mem_alloc(trans, false);
2454 		ret = PTR_ERR_OR_ZERO(new_hash);
2455 		if (ret)
2456 			goto err;
2457 	}
2458 
2459 	path->intent_ref++;
2460 	ret = __bch2_btree_node_update_key(trans, iter, b, new_hash, new_key,
2461 					   commit_flags, skip_triggers);
2462 	--path->intent_ref;
2463 
2464 	if (new_hash)
2465 		bch2_btree_node_to_freelist(c, new_hash);
2466 err:
2467 	closure_sync(&cl);
2468 	bch2_btree_cache_cannibalize_unlock(trans);
2469 	return ret;
2470 }
2471 
2472 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2473 					struct btree *b, struct bkey_i *new_key,
2474 					unsigned commit_flags, bool skip_triggers)
2475 {
2476 	struct btree_iter iter;
2477 	int ret;
2478 
2479 	bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2480 				  BTREE_MAX_DEPTH, b->c.level,
2481 				  BTREE_ITER_intent);
2482 	ret = bch2_btree_iter_traverse(&iter);
2483 	if (ret)
2484 		goto out;
2485 
2486 	/* has node been freed? */
2487 	if (btree_iter_path(trans, &iter)->l[b->c.level].b != b) {
2488 		/* node has been freed: */
2489 		BUG_ON(!btree_node_dying(b));
2490 		goto out;
2491 	}
2492 
2493 	BUG_ON(!btree_node_hashed(b));
2494 
2495 	bch2_bkey_drop_ptrs(bkey_i_to_s(new_key), ptr,
2496 			    !bch2_bkey_has_device(bkey_i_to_s(&b->key), ptr->dev));
2497 
2498 	ret = bch2_btree_node_update_key(trans, &iter, b, new_key,
2499 					 commit_flags, skip_triggers);
2500 out:
2501 	bch2_trans_iter_exit(trans, &iter);
2502 	return ret;
2503 }
2504 
2505 /* Init code: */
2506 
2507 /*
2508  * Only for filesystem bringup, when first reading the btree roots or allocating
2509  * btree roots when initializing a new filesystem:
2510  */
2511 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2512 {
2513 	BUG_ON(btree_node_root(c, b));
2514 
2515 	bch2_btree_set_root_inmem(c, b);
2516 }
2517 
2518 int bch2_btree_root_alloc_fake_trans(struct btree_trans *trans, enum btree_id id, unsigned level)
2519 {
2520 	struct bch_fs *c = trans->c;
2521 	struct closure cl;
2522 	struct btree *b;
2523 	int ret;
2524 
2525 	closure_init_stack(&cl);
2526 
2527 	do {
2528 		ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
2529 		closure_sync(&cl);
2530 	} while (ret);
2531 
2532 	b = bch2_btree_node_mem_alloc(trans, false);
2533 	bch2_btree_cache_cannibalize_unlock(trans);
2534 
2535 	ret = PTR_ERR_OR_ZERO(b);
2536 	if (ret)
2537 		return ret;
2538 
2539 	set_btree_node_fake(b);
2540 	set_btree_node_need_rewrite(b);
2541 	b->c.level	= level;
2542 	b->c.btree_id	= id;
2543 
2544 	bkey_btree_ptr_init(&b->key);
2545 	b->key.k.p = SPOS_MAX;
2546 	*((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2547 
2548 	bch2_bset_init_first(b, &b->data->keys);
2549 	bch2_btree_build_aux_trees(b);
2550 
2551 	b->data->flags = 0;
2552 	btree_set_min(b, POS_MIN);
2553 	btree_set_max(b, SPOS_MAX);
2554 	b->data->format = bch2_btree_calc_format(b);
2555 	btree_node_set_format(b, b->data->format);
2556 
2557 	ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2558 					  b->c.level, b->c.btree_id);
2559 	BUG_ON(ret);
2560 
2561 	bch2_btree_set_root_inmem(c, b);
2562 
2563 	six_unlock_write(&b->c.lock);
2564 	six_unlock_intent(&b->c.lock);
2565 	return 0;
2566 }
2567 
2568 void bch2_btree_root_alloc_fake(struct bch_fs *c, enum btree_id id, unsigned level)
2569 {
2570 	bch2_trans_run(c, lockrestart_do(trans, bch2_btree_root_alloc_fake_trans(trans, id, level)));
2571 }
2572 
2573 static void bch2_btree_update_to_text(struct printbuf *out, struct btree_update *as)
2574 {
2575 	prt_printf(out, "%ps: ", (void *) as->ip_started);
2576 	bch2_trans_commit_flags_to_text(out, as->flags);
2577 
2578 	prt_printf(out, " btree=%s l=%u-%u mode=%s nodes_written=%u cl.remaining=%u journal_seq=%llu\n",
2579 		   bch2_btree_id_str(as->btree_id),
2580 		   as->update_level_start,
2581 		   as->update_level_end,
2582 		   bch2_btree_update_modes[as->mode],
2583 		   as->nodes_written,
2584 		   closure_nr_remaining(&as->cl),
2585 		   as->journal.seq);
2586 }
2587 
2588 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2589 {
2590 	struct btree_update *as;
2591 
2592 	mutex_lock(&c->btree_interior_update_lock);
2593 	list_for_each_entry(as, &c->btree_interior_update_list, list)
2594 		bch2_btree_update_to_text(out, as);
2595 	mutex_unlock(&c->btree_interior_update_lock);
2596 }
2597 
2598 static bool bch2_btree_interior_updates_pending(struct bch_fs *c)
2599 {
2600 	bool ret;
2601 
2602 	mutex_lock(&c->btree_interior_update_lock);
2603 	ret = !list_empty(&c->btree_interior_update_list);
2604 	mutex_unlock(&c->btree_interior_update_lock);
2605 
2606 	return ret;
2607 }
2608 
2609 bool bch2_btree_interior_updates_flush(struct bch_fs *c)
2610 {
2611 	bool ret = bch2_btree_interior_updates_pending(c);
2612 
2613 	if (ret)
2614 		closure_wait_event(&c->btree_interior_update_wait,
2615 				   !bch2_btree_interior_updates_pending(c));
2616 	return ret;
2617 }
2618 
2619 void bch2_journal_entry_to_btree_root(struct bch_fs *c, struct jset_entry *entry)
2620 {
2621 	struct btree_root *r = bch2_btree_id_root(c, entry->btree_id);
2622 
2623 	mutex_lock(&c->btree_root_lock);
2624 
2625 	r->level = entry->level;
2626 	r->alive = true;
2627 	bkey_copy(&r->key, (struct bkey_i *) entry->start);
2628 
2629 	mutex_unlock(&c->btree_root_lock);
2630 }
2631 
2632 struct jset_entry *
2633 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2634 				    struct jset_entry *end,
2635 				    unsigned long skip)
2636 {
2637 	unsigned i;
2638 
2639 	mutex_lock(&c->btree_root_lock);
2640 
2641 	for (i = 0; i < btree_id_nr_alive(c); i++) {
2642 		struct btree_root *r = bch2_btree_id_root(c, i);
2643 
2644 		if (r->alive && !test_bit(i, &skip)) {
2645 			journal_entry_set(end, BCH_JSET_ENTRY_btree_root,
2646 					  i, r->level, &r->key, r->key.k.u64s);
2647 			end = vstruct_next(end);
2648 		}
2649 	}
2650 
2651 	mutex_unlock(&c->btree_root_lock);
2652 
2653 	return end;
2654 }
2655 
2656 static void bch2_btree_alloc_to_text(struct printbuf *out,
2657 				     struct bch_fs *c,
2658 				     struct btree_alloc *a)
2659 {
2660 	printbuf_indent_add(out, 2);
2661 	bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&a->k));
2662 	prt_newline(out);
2663 
2664 	struct open_bucket *ob;
2665 	unsigned i;
2666 	open_bucket_for_each(c, &a->ob, ob, i)
2667 		bch2_open_bucket_to_text(out, c, ob);
2668 
2669 	printbuf_indent_sub(out, 2);
2670 }
2671 
2672 void bch2_btree_reserve_cache_to_text(struct printbuf *out, struct bch_fs *c)
2673 {
2674 	for (unsigned i = 0; i < c->btree_reserve_cache_nr; i++)
2675 		bch2_btree_alloc_to_text(out, c, &c->btree_reserve_cache[i]);
2676 }
2677 
2678 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2679 {
2680 	if (c->btree_node_rewrite_worker)
2681 		destroy_workqueue(c->btree_node_rewrite_worker);
2682 	if (c->btree_interior_update_worker)
2683 		destroy_workqueue(c->btree_interior_update_worker);
2684 	mempool_exit(&c->btree_interior_update_pool);
2685 }
2686 
2687 void bch2_fs_btree_interior_update_init_early(struct bch_fs *c)
2688 {
2689 	mutex_init(&c->btree_reserve_cache_lock);
2690 	INIT_LIST_HEAD(&c->btree_interior_update_list);
2691 	INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2692 	mutex_init(&c->btree_interior_update_lock);
2693 	INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2694 
2695 	INIT_LIST_HEAD(&c->pending_node_rewrites);
2696 	mutex_init(&c->pending_node_rewrites_lock);
2697 }
2698 
2699 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2700 {
2701 	c->btree_interior_update_worker =
2702 		alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 8);
2703 	if (!c->btree_interior_update_worker)
2704 		return -BCH_ERR_ENOMEM_btree_interior_update_worker_init;
2705 
2706 	c->btree_node_rewrite_worker =
2707 		alloc_ordered_workqueue("btree_node_rewrite", WQ_UNBOUND);
2708 	if (!c->btree_node_rewrite_worker)
2709 		return -BCH_ERR_ENOMEM_btree_interior_update_worker_init;
2710 
2711 	if (mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2712 				      sizeof(struct btree_update)))
2713 		return -BCH_ERR_ENOMEM_btree_interior_update_pool_init;
2714 
2715 	return 0;
2716 }
2717