xref: /linux/fs/bcachefs/btree_trans_commit.c (revision 031fba65fc202abf1f193e321be7a2c274fd88ba)
1 // SPDX-License-Identifier: GPL-2.0
2 
3 #include "bcachefs.h"
4 #include "btree_gc.h"
5 #include "btree_io.h"
6 #include "btree_iter.h"
7 #include "btree_journal_iter.h"
8 #include "btree_key_cache.h"
9 #include "btree_update_interior.h"
10 #include "btree_write_buffer.h"
11 #include "buckets.h"
12 #include "errcode.h"
13 #include "error.h"
14 #include "journal.h"
15 #include "journal_reclaim.h"
16 #include "replicas.h"
17 #include "snapshot.h"
18 
19 #include <linux/prefetch.h>
20 
21 static void verify_update_old_key(struct btree_trans *trans, struct btree_insert_entry *i)
22 {
23 #ifdef CONFIG_BCACHEFS_DEBUG
24 	struct bch_fs *c = trans->c;
25 	struct bkey u;
26 	struct bkey_s_c k = bch2_btree_path_peek_slot_exact(i->path, &u);
27 
28 	if (unlikely(trans->journal_replay_not_finished)) {
29 		struct bkey_i *j_k =
30 			bch2_journal_keys_peek_slot(c, i->btree_id, i->level, i->k->k.p);
31 
32 		if (j_k)
33 			k = bkey_i_to_s_c(j_k);
34 	}
35 
36 	u = *k.k;
37 	u.needs_whiteout = i->old_k.needs_whiteout;
38 
39 	BUG_ON(memcmp(&i->old_k, &u, sizeof(struct bkey)));
40 	BUG_ON(i->old_v != k.v);
41 #endif
42 }
43 
44 static inline struct btree_path_level *insert_l(struct btree_insert_entry *i)
45 {
46 	return i->path->l + i->level;
47 }
48 
49 static inline bool same_leaf_as_prev(struct btree_trans *trans,
50 				     struct btree_insert_entry *i)
51 {
52 	return i != trans->updates &&
53 		insert_l(&i[0])->b == insert_l(&i[-1])->b;
54 }
55 
56 static inline bool same_leaf_as_next(struct btree_trans *trans,
57 				     struct btree_insert_entry *i)
58 {
59 	return i + 1 < trans->updates + trans->nr_updates &&
60 		insert_l(&i[0])->b == insert_l(&i[1])->b;
61 }
62 
63 inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
64 					   struct btree_path *path,
65 					   struct btree *b)
66 {
67 	struct bch_fs *c = trans->c;
68 
69 	if (unlikely(btree_node_just_written(b)) &&
70 	    bch2_btree_post_write_cleanup(c, b))
71 		bch2_trans_node_reinit_iter(trans, b);
72 
73 	/*
74 	 * If the last bset has been written, or if it's gotten too big - start
75 	 * a new bset to insert into:
76 	 */
77 	if (want_new_bset(c, b))
78 		bch2_btree_init_next(trans, b);
79 }
80 
81 /* Inserting into a given leaf node (last stage of insert): */
82 
83 /* Handle overwrites and do insert, for non extents: */
84 bool bch2_btree_bset_insert_key(struct btree_trans *trans,
85 				struct btree_path *path,
86 				struct btree *b,
87 				struct btree_node_iter *node_iter,
88 				struct bkey_i *insert)
89 {
90 	struct bkey_packed *k;
91 	unsigned clobber_u64s = 0, new_u64s = 0;
92 
93 	EBUG_ON(btree_node_just_written(b));
94 	EBUG_ON(bset_written(b, btree_bset_last(b)));
95 	EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
96 	EBUG_ON(bpos_lt(insert->k.p, b->data->min_key));
97 	EBUG_ON(bpos_gt(insert->k.p, b->data->max_key));
98 	EBUG_ON(insert->k.u64s >
99 		bch_btree_keys_u64s_remaining(trans->c, b));
100 	EBUG_ON(!b->c.level && !bpos_eq(insert->k.p, path->pos));
101 
102 	k = bch2_btree_node_iter_peek_all(node_iter, b);
103 	if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
104 		k = NULL;
105 
106 	/* @k is the key being overwritten/deleted, if any: */
107 	EBUG_ON(k && bkey_deleted(k));
108 
109 	/* Deleting, but not found? nothing to do: */
110 	if (bkey_deleted(&insert->k) && !k)
111 		return false;
112 
113 	if (bkey_deleted(&insert->k)) {
114 		/* Deleting: */
115 		btree_account_key_drop(b, k);
116 		k->type = KEY_TYPE_deleted;
117 
118 		if (k->needs_whiteout)
119 			push_whiteout(trans->c, b, insert->k.p);
120 		k->needs_whiteout = false;
121 
122 		if (k >= btree_bset_last(b)->start) {
123 			clobber_u64s = k->u64s;
124 			bch2_bset_delete(b, k, clobber_u64s);
125 			goto fix_iter;
126 		} else {
127 			bch2_btree_path_fix_key_modified(trans, b, k);
128 		}
129 
130 		return true;
131 	}
132 
133 	if (k) {
134 		/* Overwriting: */
135 		btree_account_key_drop(b, k);
136 		k->type = KEY_TYPE_deleted;
137 
138 		insert->k.needs_whiteout = k->needs_whiteout;
139 		k->needs_whiteout = false;
140 
141 		if (k >= btree_bset_last(b)->start) {
142 			clobber_u64s = k->u64s;
143 			goto overwrite;
144 		} else {
145 			bch2_btree_path_fix_key_modified(trans, b, k);
146 		}
147 	}
148 
149 	k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
150 overwrite:
151 	bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
152 	new_u64s = k->u64s;
153 fix_iter:
154 	if (clobber_u64s != new_u64s)
155 		bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
156 					 clobber_u64s, new_u64s);
157 	return true;
158 }
159 
160 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
161 			       unsigned i, u64 seq)
162 {
163 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
164 	struct btree_write *w = container_of(pin, struct btree_write, journal);
165 	struct btree *b = container_of(w, struct btree, writes[i]);
166 	struct btree_trans *trans = bch2_trans_get(c);
167 	unsigned long old, new, v;
168 	unsigned idx = w - b->writes;
169 
170 	btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
171 	v = READ_ONCE(b->flags);
172 
173 	do {
174 		old = new = v;
175 
176 		if (!(old & (1 << BTREE_NODE_dirty)) ||
177 		    !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
178 		    w->journal.seq != seq)
179 			break;
180 
181 		new &= ~BTREE_WRITE_TYPE_MASK;
182 		new |= BTREE_WRITE_journal_reclaim;
183 		new |= 1 << BTREE_NODE_need_write;
184 	} while ((v = cmpxchg(&b->flags, old, new)) != old);
185 
186 	btree_node_write_if_need(c, b, SIX_LOCK_read);
187 	six_unlock_read(&b->c.lock);
188 
189 	bch2_trans_put(trans);
190 	return 0;
191 }
192 
193 int bch2_btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
194 {
195 	return __btree_node_flush(j, pin, 0, seq);
196 }
197 
198 int bch2_btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
199 {
200 	return __btree_node_flush(j, pin, 1, seq);
201 }
202 
203 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
204 				       struct btree *b, u64 seq)
205 {
206 	struct btree_write *w = btree_current_write(b);
207 
208 	bch2_journal_pin_add(&c->journal, seq, &w->journal,
209 			     btree_node_write_idx(b) == 0
210 			     ? bch2_btree_node_flush0
211 			     : bch2_btree_node_flush1);
212 }
213 
214 /**
215  * bch2_btree_insert_key_leaf() - insert a key one key into a leaf node
216  * @trans:		btree transaction object
217  * @path:		path pointing to @insert's pos
218  * @insert:		key to insert
219  * @journal_seq:	sequence number of journal reservation
220  */
221 inline void bch2_btree_insert_key_leaf(struct btree_trans *trans,
222 				       struct btree_path *path,
223 				       struct bkey_i *insert,
224 				       u64 journal_seq)
225 {
226 	struct bch_fs *c = trans->c;
227 	struct btree *b = path_l(path)->b;
228 	struct bset_tree *t = bset_tree_last(b);
229 	struct bset *i = bset(b, t);
230 	int old_u64s = bset_u64s(t);
231 	int old_live_u64s = b->nr.live_u64s;
232 	int live_u64s_added, u64s_added;
233 
234 	if (unlikely(!bch2_btree_bset_insert_key(trans, path, b,
235 					&path_l(path)->iter, insert)))
236 		return;
237 
238 	i->journal_seq = cpu_to_le64(max(journal_seq, le64_to_cpu(i->journal_seq)));
239 
240 	bch2_btree_add_journal_pin(c, b, journal_seq);
241 
242 	if (unlikely(!btree_node_dirty(b))) {
243 		EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
244 		set_btree_node_dirty_acct(c, b);
245 	}
246 
247 	live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
248 	u64s_added = (int) bset_u64s(t) - old_u64s;
249 
250 	if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
251 		b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
252 	if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
253 		b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
254 
255 	if (u64s_added > live_u64s_added &&
256 	    bch2_maybe_compact_whiteouts(c, b))
257 		bch2_trans_node_reinit_iter(trans, b);
258 }
259 
260 /* Cached btree updates: */
261 
262 /* Normal update interface: */
263 
264 static inline void btree_insert_entry_checks(struct btree_trans *trans,
265 					     struct btree_insert_entry *i)
266 {
267 	BUG_ON(!bpos_eq(i->k->k.p, i->path->pos));
268 	BUG_ON(i->cached	!= i->path->cached);
269 	BUG_ON(i->level		!= i->path->level);
270 	BUG_ON(i->btree_id	!= i->path->btree_id);
271 	EBUG_ON(!i->level &&
272 		!(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
273 		test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
274 		i->k->k.p.snapshot &&
275 		bch2_snapshot_is_internal_node(trans->c, i->k->k.p.snapshot));
276 }
277 
278 static noinline int
279 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned flags,
280 				   unsigned long trace_ip)
281 {
282 	return drop_locks_do(trans,
283 		bch2_journal_preres_get(&trans->c->journal,
284 			&trans->journal_preres,
285 			trans->journal_preres_u64s,
286 			(flags & BCH_WATERMARK_MASK)));
287 }
288 
289 static __always_inline int bch2_trans_journal_res_get(struct btree_trans *trans,
290 						      unsigned flags)
291 {
292 	return bch2_journal_res_get(&trans->c->journal, &trans->journal_res,
293 				    trans->journal_u64s, flags);
294 }
295 
296 #define JSET_ENTRY_LOG_U64s		4
297 
298 static noinline void journal_transaction_name(struct btree_trans *trans)
299 {
300 	struct bch_fs *c = trans->c;
301 	struct journal *j = &c->journal;
302 	struct jset_entry *entry =
303 		bch2_journal_add_entry(j, &trans->journal_res,
304 				       BCH_JSET_ENTRY_log, 0, 0,
305 				       JSET_ENTRY_LOG_U64s);
306 	struct jset_entry_log *l =
307 		container_of(entry, struct jset_entry_log, entry);
308 
309 	strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
310 }
311 
312 static inline int btree_key_can_insert(struct btree_trans *trans,
313 				       struct btree *b, unsigned u64s)
314 {
315 	struct bch_fs *c = trans->c;
316 
317 	if (!bch2_btree_node_insert_fits(c, b, u64s))
318 		return -BCH_ERR_btree_insert_btree_node_full;
319 
320 	return 0;
321 }
322 
323 static int btree_key_can_insert_cached(struct btree_trans *trans, unsigned flags,
324 				       struct btree_path *path, unsigned u64s)
325 {
326 	struct bch_fs *c = trans->c;
327 	struct bkey_cached *ck = (void *) path->l[0].b;
328 	struct btree_insert_entry *i;
329 	unsigned new_u64s;
330 	struct bkey_i *new_k;
331 
332 	EBUG_ON(path->level);
333 
334 	if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
335 	    bch2_btree_key_cache_must_wait(c) &&
336 	    !(flags & BTREE_INSERT_JOURNAL_RECLAIM))
337 		return -BCH_ERR_btree_insert_need_journal_reclaim;
338 
339 	/*
340 	 * bch2_varint_decode can read past the end of the buffer by at most 7
341 	 * bytes (it won't be used):
342 	 */
343 	u64s += 1;
344 
345 	if (u64s <= ck->u64s)
346 		return 0;
347 
348 	new_u64s	= roundup_pow_of_two(u64s);
349 	new_k		= krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
350 	if (!new_k) {
351 		bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
352 			bch2_btree_ids[path->btree_id], new_u64s);
353 		return -BCH_ERR_ENOMEM_btree_key_cache_insert;
354 	}
355 
356 	trans_for_each_update(trans, i)
357 		if (i->old_v == &ck->k->v)
358 			i->old_v = &new_k->v;
359 
360 	ck->u64s	= new_u64s;
361 	ck->k		= new_k;
362 	return 0;
363 }
364 
365 /* Triggers: */
366 
367 static int run_one_mem_trigger(struct btree_trans *trans,
368 			       struct btree_insert_entry *i,
369 			       unsigned flags)
370 {
371 	struct bkey_s_c old = { &i->old_k, i->old_v };
372 	struct bkey_i *new = i->k;
373 	const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
374 	const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
375 	int ret;
376 
377 	verify_update_old_key(trans, i);
378 
379 	if (unlikely(flags & BTREE_TRIGGER_NORUN))
380 		return 0;
381 
382 	if (!btree_node_type_needs_gc((enum btree_node_type) i->btree_id))
383 		return 0;
384 
385 	if (old_ops->atomic_trigger == new_ops->atomic_trigger &&
386 	    ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
387 		ret   = bch2_mark_key(trans, i->btree_id, i->level,
388 				old, bkey_i_to_s_c(new),
389 				BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
390 	} else {
391 		struct bkey		_deleted = KEY(0, 0, 0);
392 		struct bkey_s_c		deleted = (struct bkey_s_c) { &_deleted, NULL };
393 
394 		_deleted.p = i->path->pos;
395 
396 		ret   = bch2_mark_key(trans, i->btree_id, i->level,
397 				deleted, bkey_i_to_s_c(new),
398 				BTREE_TRIGGER_INSERT|flags) ?:
399 			bch2_mark_key(trans, i->btree_id, i->level,
400 				old, deleted,
401 				BTREE_TRIGGER_OVERWRITE|flags);
402 	}
403 
404 	return ret;
405 }
406 
407 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
408 				 bool overwrite)
409 {
410 	/*
411 	 * Transactional triggers create new btree_insert_entries, so we can't
412 	 * pass them a pointer to a btree_insert_entry, that memory is going to
413 	 * move:
414 	 */
415 	struct bkey old_k = i->old_k;
416 	struct bkey_s_c old = { &old_k, i->old_v };
417 	const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
418 	const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
419 
420 	verify_update_old_key(trans, i);
421 
422 	if ((i->flags & BTREE_TRIGGER_NORUN) ||
423 	    !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
424 		return 0;
425 
426 	if (!i->insert_trigger_run &&
427 	    !i->overwrite_trigger_run &&
428 	    old_ops->trans_trigger == new_ops->trans_trigger &&
429 	    ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
430 		i->overwrite_trigger_run = true;
431 		i->insert_trigger_run = true;
432 		return bch2_trans_mark_key(trans, i->btree_id, i->level, old, i->k,
433 					   BTREE_TRIGGER_INSERT|
434 					   BTREE_TRIGGER_OVERWRITE|
435 					   i->flags) ?: 1;
436 	} else if (overwrite && !i->overwrite_trigger_run) {
437 		i->overwrite_trigger_run = true;
438 		return bch2_trans_mark_old(trans, i->btree_id, i->level, old, i->flags) ?: 1;
439 	} else if (!overwrite && !i->insert_trigger_run) {
440 		i->insert_trigger_run = true;
441 		return bch2_trans_mark_new(trans, i->btree_id, i->level, i->k, i->flags) ?: 1;
442 	} else {
443 		return 0;
444 	}
445 }
446 
447 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
448 			      struct btree_insert_entry *btree_id_start)
449 {
450 	struct btree_insert_entry *i;
451 	bool trans_trigger_run;
452 	int ret, overwrite;
453 
454 	for (overwrite = 1; overwrite >= 0; --overwrite) {
455 
456 		/*
457 		 * Running triggers will append more updates to the list of updates as
458 		 * we're walking it:
459 		 */
460 		do {
461 			trans_trigger_run = false;
462 
463 			for (i = btree_id_start;
464 			     i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
465 			     i++) {
466 				if (i->btree_id != btree_id)
467 					continue;
468 
469 				ret = run_one_trans_trigger(trans, i, overwrite);
470 				if (ret < 0)
471 					return ret;
472 				if (ret)
473 					trans_trigger_run = true;
474 			}
475 		} while (trans_trigger_run);
476 	}
477 
478 	return 0;
479 }
480 
481 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
482 {
483 	struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
484 	unsigned btree_id = 0;
485 	int ret = 0;
486 
487 	/*
488 	 *
489 	 * For a given btree, this algorithm runs insert triggers before
490 	 * overwrite triggers: this is so that when extents are being moved
491 	 * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
492 	 * they are re-added.
493 	 */
494 	for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
495 		if (btree_id == BTREE_ID_alloc)
496 			continue;
497 
498 		while (btree_id_start < trans->updates + trans->nr_updates &&
499 		       btree_id_start->btree_id < btree_id)
500 			btree_id_start++;
501 
502 		ret = run_btree_triggers(trans, btree_id, btree_id_start);
503 		if (ret)
504 			return ret;
505 	}
506 
507 	trans_for_each_update(trans, i) {
508 		if (i->btree_id > BTREE_ID_alloc)
509 			break;
510 		if (i->btree_id == BTREE_ID_alloc) {
511 			ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
512 			if (ret)
513 				return ret;
514 			break;
515 		}
516 	}
517 
518 #ifdef CONFIG_BCACHEFS_DEBUG
519 	trans_for_each_update(trans, i)
520 		BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
521 		       (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
522 		       (!i->insert_trigger_run || !i->overwrite_trigger_run));
523 #endif
524 	return 0;
525 }
526 
527 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
528 {
529 	struct bch_fs *c = trans->c;
530 	struct btree_insert_entry *i;
531 	int ret = 0;
532 
533 	trans_for_each_update(trans, i) {
534 		/*
535 		 * XXX: synchronization of cached update triggers with gc
536 		 * XXX: synchronization of interior node updates with gc
537 		 */
538 		BUG_ON(i->cached || i->level);
539 
540 		if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
541 			ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
542 			if (ret)
543 				break;
544 		}
545 	}
546 
547 	return ret;
548 }
549 
550 static inline int
551 bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
552 			       struct btree_insert_entry **stopped_at,
553 			       unsigned long trace_ip)
554 {
555 	struct bch_fs *c = trans->c;
556 	struct btree_insert_entry *i;
557 	struct btree_write_buffered_key *wb;
558 	struct btree_trans_commit_hook *h;
559 	unsigned u64s = 0;
560 	int ret;
561 
562 	if (race_fault()) {
563 		trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
564 		return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
565 	}
566 
567 	/*
568 	 * Check if the insert will fit in the leaf node with the write lock
569 	 * held, otherwise another thread could write the node changing the
570 	 * amount of space available:
571 	 */
572 
573 	prefetch(&trans->c->journal.flags);
574 
575 	trans_for_each_update(trans, i) {
576 		/* Multiple inserts might go to same leaf: */
577 		if (!same_leaf_as_prev(trans, i))
578 			u64s = 0;
579 
580 		u64s += i->k->k.u64s;
581 		ret = !i->cached
582 			? btree_key_can_insert(trans, insert_l(i)->b, u64s)
583 			: btree_key_can_insert_cached(trans, flags, i->path, u64s);
584 		if (ret) {
585 			*stopped_at = i;
586 			return ret;
587 		}
588 	}
589 
590 	if (trans->nr_wb_updates &&
591 	    trans->nr_wb_updates + c->btree_write_buffer.state.nr > c->btree_write_buffer.size)
592 		return -BCH_ERR_btree_insert_need_flush_buffer;
593 
594 	/*
595 	 * Don't get journal reservation until after we know insert will
596 	 * succeed:
597 	 */
598 	if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
599 		ret = bch2_trans_journal_res_get(trans,
600 				(flags & BCH_WATERMARK_MASK)|
601 				JOURNAL_RES_GET_NONBLOCK);
602 		if (ret)
603 			return ret;
604 
605 		if (unlikely(trans->journal_transaction_names))
606 			journal_transaction_name(trans);
607 	} else {
608 		trans->journal_res.seq = c->journal.replay_journal_seq;
609 	}
610 
611 	/*
612 	 * Not allowed to fail after we've gotten our journal reservation - we
613 	 * have to use it:
614 	 */
615 
616 	if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
617 	    !(flags & BTREE_INSERT_JOURNAL_REPLAY)) {
618 		if (bch2_journal_seq_verify)
619 			trans_for_each_update(trans, i)
620 				i->k->k.version.lo = trans->journal_res.seq;
621 		else if (bch2_inject_invalid_keys)
622 			trans_for_each_update(trans, i)
623 				i->k->k.version = MAX_VERSION;
624 	}
625 
626 	if (trans->fs_usage_deltas &&
627 	    bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
628 		return -BCH_ERR_btree_insert_need_mark_replicas;
629 
630 	if (trans->nr_wb_updates) {
631 		EBUG_ON(flags & BTREE_INSERT_JOURNAL_REPLAY);
632 
633 		ret = bch2_btree_insert_keys_write_buffer(trans);
634 		if (ret)
635 			goto revert_fs_usage;
636 	}
637 
638 	h = trans->hooks;
639 	while (h) {
640 		ret = h->fn(trans, h);
641 		if (ret)
642 			goto revert_fs_usage;
643 		h = h->next;
644 	}
645 
646 	trans_for_each_update(trans, i)
647 		if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
648 			ret = run_one_mem_trigger(trans, i, i->flags);
649 			if (ret)
650 				goto fatal_err;
651 		}
652 
653 	if (unlikely(c->gc_pos.phase)) {
654 		ret = bch2_trans_commit_run_gc_triggers(trans);
655 		if  (ret)
656 			goto fatal_err;
657 	}
658 
659 	if (unlikely(trans->extra_journal_entries.nr)) {
660 		memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
661 				  trans->extra_journal_entries.data,
662 				  trans->extra_journal_entries.nr);
663 
664 		trans->journal_res.offset	+= trans->extra_journal_entries.nr;
665 		trans->journal_res.u64s		-= trans->extra_journal_entries.nr;
666 	}
667 
668 	if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
669 		struct journal *j = &c->journal;
670 		struct jset_entry *entry;
671 
672 		trans_for_each_update(trans, i) {
673 			if (i->key_cache_already_flushed)
674 				continue;
675 
676 			if (i->flags & BTREE_UPDATE_NOJOURNAL)
677 				continue;
678 
679 			verify_update_old_key(trans, i);
680 
681 			if (trans->journal_transaction_names) {
682 				entry = bch2_journal_add_entry(j, &trans->journal_res,
683 						       BCH_JSET_ENTRY_overwrite,
684 						       i->btree_id, i->level,
685 						       i->old_k.u64s);
686 				bkey_reassemble(&entry->start[0],
687 						(struct bkey_s_c) { &i->old_k, i->old_v });
688 			}
689 
690 			entry = bch2_journal_add_entry(j, &trans->journal_res,
691 					       BCH_JSET_ENTRY_btree_keys,
692 					       i->btree_id, i->level,
693 					       i->k->k.u64s);
694 			bkey_copy(&entry->start[0], i->k);
695 		}
696 
697 		trans_for_each_wb_update(trans, wb) {
698 			entry = bch2_journal_add_entry(j, &trans->journal_res,
699 					       BCH_JSET_ENTRY_btree_keys,
700 					       wb->btree, 0,
701 					       wb->k.k.u64s);
702 			bkey_copy(&entry->start[0], &wb->k);
703 		}
704 
705 		if (trans->journal_seq)
706 			*trans->journal_seq = trans->journal_res.seq;
707 	}
708 
709 	trans_for_each_update(trans, i) {
710 		i->k->k.needs_whiteout = false;
711 
712 		if (!i->cached) {
713 			u64 seq = trans->journal_res.seq;
714 
715 			if (i->flags & BTREE_UPDATE_PREJOURNAL)
716 				seq = i->seq;
717 
718 			bch2_btree_insert_key_leaf(trans, i->path, i->k, seq);
719 		} else if (!i->key_cache_already_flushed)
720 			bch2_btree_insert_key_cached(trans, flags, i);
721 		else {
722 			bch2_btree_key_cache_drop(trans, i->path);
723 			btree_path_set_dirty(i->path, BTREE_ITER_NEED_TRAVERSE);
724 		}
725 	}
726 
727 	return 0;
728 fatal_err:
729 	bch2_fatal_error(c);
730 revert_fs_usage:
731 	if (trans->fs_usage_deltas)
732 		bch2_trans_fs_usage_revert(trans, trans->fs_usage_deltas);
733 	return ret;
734 }
735 
736 static noinline int trans_lock_write_fail(struct btree_trans *trans, struct btree_insert_entry *i)
737 {
738 	while (--i >= trans->updates) {
739 		if (same_leaf_as_prev(trans, i))
740 			continue;
741 
742 		bch2_btree_node_unlock_write(trans, i->path, insert_l(i)->b);
743 	}
744 
745 	trace_and_count(trans->c, trans_restart_would_deadlock_write, trans);
746 	return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
747 }
748 
749 static inline int trans_lock_write(struct btree_trans *trans)
750 {
751 	struct btree_insert_entry *i;
752 
753 	trans_for_each_update(trans, i) {
754 		if (same_leaf_as_prev(trans, i))
755 			continue;
756 
757 		if (bch2_btree_node_lock_write(trans, i->path, &insert_l(i)->b->c))
758 			return trans_lock_write_fail(trans, i);
759 
760 		if (!i->cached)
761 			bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
762 	}
763 
764 	return 0;
765 }
766 
767 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
768 {
769 	struct btree_insert_entry *i;
770 	struct btree_write_buffered_key *wb;
771 
772 	trans_for_each_update(trans, i)
773 		bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
774 
775 	trans_for_each_wb_update(trans, wb)
776 		bch2_journal_key_overwritten(trans->c, wb->btree, 0, wb->k.k.p);
777 }
778 
779 static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans, unsigned flags,
780 						   struct btree_insert_entry *i,
781 						   struct printbuf *err)
782 {
783 	struct bch_fs *c = trans->c;
784 	int rw = (flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
785 
786 	printbuf_reset(err);
787 	prt_printf(err, "invalid bkey on insert from %s -> %ps",
788 		   trans->fn, (void *) i->ip_allocated);
789 	prt_newline(err);
790 	printbuf_indent_add(err, 2);
791 
792 	bch2_bkey_val_to_text(err, c, bkey_i_to_s_c(i->k));
793 	prt_newline(err);
794 
795 	bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
796 			  i->bkey_type, rw, err);
797 	bch2_print_string_as_lines(KERN_ERR, err->buf);
798 
799 	bch2_inconsistent_error(c);
800 	bch2_dump_trans_updates(trans);
801 
802 	return -EINVAL;
803 }
804 
805 /*
806  * Get journal reservation, take write locks, and attempt to do btree update(s):
807  */
808 static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags,
809 				       struct btree_insert_entry **stopped_at,
810 				       unsigned long trace_ip)
811 {
812 	struct bch_fs *c = trans->c;
813 	struct btree_insert_entry *i;
814 	int ret = 0, u64s_delta = 0;
815 
816 	trans_for_each_update(trans, i) {
817 		if (i->cached)
818 			continue;
819 
820 		u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
821 		u64s_delta -= i->old_btree_u64s;
822 
823 		if (!same_leaf_as_next(trans, i)) {
824 			if (u64s_delta <= 0) {
825 				ret = bch2_foreground_maybe_merge(trans, i->path,
826 							i->level, flags);
827 				if (unlikely(ret))
828 					return ret;
829 			}
830 
831 			u64s_delta = 0;
832 		}
833 	}
834 
835 	ret = bch2_journal_preres_get(&c->journal,
836 			&trans->journal_preres, trans->journal_preres_u64s,
837 			(flags & BCH_WATERMARK_MASK)|JOURNAL_RES_GET_NONBLOCK);
838 	if (unlikely(ret == -BCH_ERR_journal_preres_get_blocked))
839 		ret = bch2_trans_journal_preres_get_cold(trans, flags, trace_ip);
840 	if (unlikely(ret))
841 		return ret;
842 
843 	ret = trans_lock_write(trans);
844 	if (unlikely(ret))
845 		return ret;
846 
847 	ret = bch2_trans_commit_write_locked(trans, flags, stopped_at, trace_ip);
848 
849 	if (!ret && unlikely(trans->journal_replay_not_finished))
850 		bch2_drop_overwrites_from_journal(trans);
851 
852 	trans_for_each_update(trans, i)
853 		if (!same_leaf_as_prev(trans, i))
854 			bch2_btree_node_unlock_write_inlined(trans, i->path,
855 							insert_l(i)->b);
856 
857 	if (!ret && trans->journal_pin)
858 		bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
859 				     trans->journal_pin, NULL);
860 
861 	/*
862 	 * Drop journal reservation after dropping write locks, since dropping
863 	 * the journal reservation may kick off a journal write:
864 	 */
865 	bch2_journal_res_put(&c->journal, &trans->journal_res);
866 
867 	if (unlikely(ret))
868 		return ret;
869 
870 	bch2_trans_downgrade(trans);
871 
872 	return 0;
873 }
874 
875 static int journal_reclaim_wait_done(struct bch_fs *c)
876 {
877 	int ret = bch2_journal_error(&c->journal) ?:
878 		!bch2_btree_key_cache_must_wait(c);
879 
880 	if (!ret)
881 		journal_reclaim_kick(&c->journal);
882 	return ret;
883 }
884 
885 static noinline
886 int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
887 			    struct btree_insert_entry *i,
888 			    int ret, unsigned long trace_ip)
889 {
890 	struct bch_fs *c = trans->c;
891 
892 	switch (ret) {
893 	case -BCH_ERR_btree_insert_btree_node_full:
894 		ret = bch2_btree_split_leaf(trans, i->path, flags);
895 		if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
896 			trace_and_count(c, trans_restart_btree_node_split, trans, trace_ip, i->path);
897 		break;
898 	case -BCH_ERR_btree_insert_need_mark_replicas:
899 		ret = drop_locks_do(trans,
900 			bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas));
901 		break;
902 	case -BCH_ERR_journal_res_get_blocked:
903 		/*
904 		 * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
905 		 * flag
906 		 */
907 		if ((flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
908 		    (flags & BCH_WATERMARK_MASK) != BCH_WATERMARK_reclaim) {
909 			ret = -BCH_ERR_journal_reclaim_would_deadlock;
910 			break;
911 		}
912 
913 		ret = drop_locks_do(trans,
914 			bch2_trans_journal_res_get(trans,
915 					(flags & BCH_WATERMARK_MASK)|
916 					JOURNAL_RES_GET_CHECK));
917 		break;
918 	case -BCH_ERR_btree_insert_need_journal_reclaim:
919 		bch2_trans_unlock(trans);
920 
921 		trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
922 
923 		wait_event_freezable(c->journal.reclaim_wait,
924 				     (ret = journal_reclaim_wait_done(c)));
925 		if (ret < 0)
926 			break;
927 
928 		ret = bch2_trans_relock(trans);
929 		break;
930 	case -BCH_ERR_btree_insert_need_flush_buffer: {
931 		struct btree_write_buffer *wb = &c->btree_write_buffer;
932 
933 		ret = 0;
934 
935 		if (wb->state.nr > wb->size * 3 / 4) {
936 			bch2_trans_unlock(trans);
937 			mutex_lock(&wb->flush_lock);
938 
939 			if (wb->state.nr > wb->size * 3 / 4) {
940 				bch2_trans_begin(trans);
941 				ret = __bch2_btree_write_buffer_flush(trans,
942 						flags|BTREE_INSERT_NOCHECK_RW, true);
943 				if (!ret) {
944 					trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_);
945 					ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush);
946 				}
947 			} else {
948 				mutex_unlock(&wb->flush_lock);
949 				ret = bch2_trans_relock(trans);
950 			}
951 		}
952 		break;
953 	}
954 	default:
955 		BUG_ON(ret >= 0);
956 		break;
957 	}
958 
959 	BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
960 
961 	bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
962 				!(flags & BTREE_INSERT_NOWAIT) &&
963 				(flags & BTREE_INSERT_NOFAIL), c,
964 		"%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
965 
966 	return ret;
967 }
968 
969 static noinline int
970 bch2_trans_commit_get_rw_cold(struct btree_trans *trans, unsigned flags)
971 {
972 	struct bch_fs *c = trans->c;
973 	int ret;
974 
975 	if (likely(!(flags & BTREE_INSERT_LAZY_RW)) ||
976 	    test_bit(BCH_FS_STARTED, &c->flags))
977 		return -BCH_ERR_erofs_trans_commit;
978 
979 	ret = drop_locks_do(trans, bch2_fs_read_write_early(c));
980 	if (ret)
981 		return ret;
982 
983 	bch2_write_ref_get(c, BCH_WRITE_REF_trans);
984 	return 0;
985 }
986 
987 /*
988  * This is for updates done in the early part of fsck - btree_gc - before we've
989  * gone RW. we only add the new key to the list of keys for journal replay to
990  * do.
991  */
992 static noinline int
993 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
994 {
995 	struct bch_fs *c = trans->c;
996 	struct btree_insert_entry *i;
997 	int ret = 0;
998 
999 	trans_for_each_update(trans, i) {
1000 		ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
1001 		if (ret)
1002 			break;
1003 	}
1004 
1005 	return ret;
1006 }
1007 
1008 int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
1009 {
1010 	struct bch_fs *c = trans->c;
1011 	struct btree_insert_entry *i = NULL;
1012 	struct btree_write_buffered_key *wb;
1013 	unsigned u64s;
1014 	int ret = 0;
1015 
1016 	if (!trans->nr_updates &&
1017 	    !trans->nr_wb_updates &&
1018 	    !trans->extra_journal_entries.nr)
1019 		goto out_reset;
1020 
1021 	if (flags & BTREE_INSERT_GC_LOCK_HELD)
1022 		lockdep_assert_held(&c->gc_lock);
1023 
1024 	ret = bch2_trans_commit_run_triggers(trans);
1025 	if (ret)
1026 		goto out_reset;
1027 
1028 	trans_for_each_update(trans, i) {
1029 		struct printbuf buf = PRINTBUF;
1030 		enum bkey_invalid_flags invalid_flags = 0;
1031 
1032 		if (!(flags & BTREE_INSERT_JOURNAL_REPLAY))
1033 			invalid_flags |= BKEY_INVALID_WRITE|BKEY_INVALID_COMMIT;
1034 
1035 		if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
1036 					       i->bkey_type, invalid_flags, &buf)))
1037 			ret = bch2_trans_commit_bkey_invalid(trans, flags, i, &buf);
1038 		btree_insert_entry_checks(trans, i);
1039 		printbuf_exit(&buf);
1040 
1041 		if (ret)
1042 			return ret;
1043 	}
1044 
1045 	if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
1046 		ret = do_bch2_trans_commit_to_journal_replay(trans);
1047 		goto out_reset;
1048 	}
1049 
1050 	if (!(flags & BTREE_INSERT_NOCHECK_RW) &&
1051 	    unlikely(!bch2_write_ref_tryget(c, BCH_WRITE_REF_trans))) {
1052 		ret = bch2_trans_commit_get_rw_cold(trans, flags);
1053 		if (ret)
1054 			goto out_reset;
1055 	}
1056 
1057 	if (c->btree_write_buffer.state.nr > c->btree_write_buffer.size / 2 &&
1058 	    mutex_trylock(&c->btree_write_buffer.flush_lock)) {
1059 		bch2_trans_begin(trans);
1060 		bch2_trans_unlock(trans);
1061 
1062 		ret = __bch2_btree_write_buffer_flush(trans,
1063 					flags|BTREE_INSERT_NOCHECK_RW, true);
1064 		if (!ret) {
1065 			trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_);
1066 			ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush);
1067 		}
1068 		goto out;
1069 	}
1070 
1071 	EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
1072 
1073 	memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
1074 
1075 	trans->journal_u64s		= trans->extra_journal_entries.nr;
1076 	trans->journal_preres_u64s	= 0;
1077 
1078 	trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
1079 
1080 	if (trans->journal_transaction_names)
1081 		trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1082 
1083 	trans_for_each_update(trans, i) {
1084 		EBUG_ON(!i->path->should_be_locked);
1085 
1086 		ret = bch2_btree_path_upgrade(trans, i->path, i->level + 1);
1087 		if (unlikely(ret))
1088 			goto out;
1089 
1090 		EBUG_ON(!btree_node_intent_locked(i->path, i->level));
1091 
1092 		if (i->key_cache_already_flushed)
1093 			continue;
1094 
1095 		/* we're going to journal the key being updated: */
1096 		u64s = jset_u64s(i->k->k.u64s);
1097 		if (i->cached &&
1098 		    likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY)))
1099 			trans->journal_preres_u64s += u64s;
1100 
1101 		if (i->flags & BTREE_UPDATE_NOJOURNAL)
1102 			continue;
1103 
1104 		trans->journal_u64s += u64s;
1105 
1106 		/* and we're also going to log the overwrite: */
1107 		if (trans->journal_transaction_names)
1108 			trans->journal_u64s += jset_u64s(i->old_k.u64s);
1109 	}
1110 
1111 	trans_for_each_wb_update(trans, wb)
1112 		trans->journal_u64s += jset_u64s(wb->k.k.u64s);
1113 
1114 	if (trans->extra_journal_res) {
1115 		ret = bch2_disk_reservation_add(c, trans->disk_res,
1116 				trans->extra_journal_res,
1117 				(flags & BTREE_INSERT_NOFAIL)
1118 				? BCH_DISK_RESERVATION_NOFAIL : 0);
1119 		if (ret)
1120 			goto err;
1121 	}
1122 retry:
1123 	bch2_trans_verify_not_in_restart(trans);
1124 	memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1125 
1126 	ret = do_bch2_trans_commit(trans, flags, &i, _RET_IP_);
1127 
1128 	/* make sure we didn't drop or screw up locks: */
1129 	bch2_trans_verify_locks(trans);
1130 
1131 	if (ret)
1132 		goto err;
1133 
1134 	trace_and_count(c, transaction_commit, trans, _RET_IP_);
1135 out:
1136 	bch2_journal_preres_put(&c->journal, &trans->journal_preres);
1137 
1138 	if (likely(!(flags & BTREE_INSERT_NOCHECK_RW)))
1139 		bch2_write_ref_put(c, BCH_WRITE_REF_trans);
1140 out_reset:
1141 	bch2_trans_reset_updates(trans);
1142 
1143 	return ret;
1144 err:
1145 	ret = bch2_trans_commit_error(trans, flags, i, ret, _RET_IP_);
1146 	if (ret)
1147 		goto out;
1148 
1149 	goto retry;
1150 }
1151