xref: /linux/fs/bcachefs/btree_locking.c (revision 4f05e82003d1c20da29fa593420b8d92e2c8d4e6)
1 // SPDX-License-Identifier: GPL-2.0
2 
3 #include "bcachefs.h"
4 #include "btree_locking.h"
5 #include "btree_types.h"
6 
7 static struct lock_class_key bch2_btree_node_lock_key;
8 
9 void bch2_btree_lock_init(struct btree_bkey_cached_common *b,
10 			  enum six_lock_init_flags flags)
11 {
12 	__six_lock_init(&b->lock, "b->c.lock", &bch2_btree_node_lock_key, flags);
13 	lockdep_set_novalidate_class(&b->lock);
14 }
15 
16 #ifdef CONFIG_LOCKDEP
17 void bch2_assert_btree_nodes_not_locked(void)
18 {
19 #if 0
20 	//Re-enable when lock_class_is_held() is merged:
21 	BUG_ON(lock_class_is_held(&bch2_btree_node_lock_key));
22 #endif
23 }
24 #endif
25 
26 /* Btree node locking: */
27 
28 struct six_lock_count bch2_btree_node_lock_counts(struct btree_trans *trans,
29 						  struct btree_path *skip,
30 						  struct btree_bkey_cached_common *b,
31 						  unsigned level)
32 {
33 	struct btree_path *path;
34 	struct six_lock_count ret;
35 	unsigned i;
36 
37 	memset(&ret, 0, sizeof(ret));
38 
39 	if (IS_ERR_OR_NULL(b))
40 		return ret;
41 
42 	trans_for_each_path(trans, path, i)
43 		if (path != skip && &path->l[level].b->c == b) {
44 			int t = btree_node_locked_type(path, level);
45 
46 			if (t != BTREE_NODE_UNLOCKED)
47 				ret.n[t]++;
48 		}
49 
50 	return ret;
51 }
52 
53 /* unlock */
54 
55 void bch2_btree_node_unlock_write(struct btree_trans *trans,
56 			struct btree_path *path, struct btree *b)
57 {
58 	bch2_btree_node_unlock_write_inlined(trans, path, b);
59 }
60 
61 /* lock */
62 
63 /*
64  * @trans wants to lock @b with type @type
65  */
66 struct trans_waiting_for_lock {
67 	struct btree_trans		*trans;
68 	struct btree_bkey_cached_common	*node_want;
69 	enum six_lock_type		lock_want;
70 
71 	/* for iterating over held locks :*/
72 	u8				path_idx;
73 	u8				level;
74 	u64				lock_start_time;
75 };
76 
77 struct lock_graph {
78 	struct trans_waiting_for_lock	g[8];
79 	unsigned			nr;
80 };
81 
82 static noinline void print_cycle(struct printbuf *out, struct lock_graph *g)
83 {
84 	struct trans_waiting_for_lock *i;
85 
86 	prt_printf(out, "Found lock cycle (%u entries):\n", g->nr);
87 
88 	for (i = g->g; i < g->g + g->nr; i++) {
89 		struct task_struct *task = READ_ONCE(i->trans->locking_wait.task);
90 		if (!task)
91 			continue;
92 
93 		bch2_btree_trans_to_text(out, i->trans);
94 		bch2_prt_task_backtrace(out, task, i == g->g ? 5 : 1, GFP_NOWAIT);
95 	}
96 }
97 
98 static noinline void print_chain(struct printbuf *out, struct lock_graph *g)
99 {
100 	struct trans_waiting_for_lock *i;
101 
102 	for (i = g->g; i != g->g + g->nr; i++) {
103 		struct task_struct *task = i->trans->locking_wait.task;
104 		if (i != g->g)
105 			prt_str(out, "<- ");
106 		prt_printf(out, "%u ", task ?task->pid : 0);
107 	}
108 	prt_newline(out);
109 }
110 
111 static void lock_graph_up(struct lock_graph *g)
112 {
113 	closure_put(&g->g[--g->nr].trans->ref);
114 }
115 
116 static noinline void lock_graph_pop_all(struct lock_graph *g)
117 {
118 	while (g->nr)
119 		lock_graph_up(g);
120 }
121 
122 static void __lock_graph_down(struct lock_graph *g, struct btree_trans *trans)
123 {
124 	g->g[g->nr++] = (struct trans_waiting_for_lock) {
125 		.trans		= trans,
126 		.node_want	= trans->locking,
127 		.lock_want	= trans->locking_wait.lock_want,
128 	};
129 }
130 
131 static void lock_graph_down(struct lock_graph *g, struct btree_trans *trans)
132 {
133 	closure_get(&trans->ref);
134 	__lock_graph_down(g, trans);
135 }
136 
137 static bool lock_graph_remove_non_waiters(struct lock_graph *g)
138 {
139 	struct trans_waiting_for_lock *i;
140 
141 	for (i = g->g + 1; i < g->g + g->nr; i++)
142 		if (i->trans->locking != i->node_want ||
143 		    i->trans->locking_wait.start_time != i[-1].lock_start_time) {
144 			while (g->g + g->nr > i)
145 				lock_graph_up(g);
146 			return true;
147 		}
148 
149 	return false;
150 }
151 
152 static void trace_would_deadlock(struct lock_graph *g, struct btree_trans *trans)
153 {
154 	struct bch_fs *c = trans->c;
155 
156 	count_event(c, trans_restart_would_deadlock);
157 
158 	if (trace_trans_restart_would_deadlock_enabled()) {
159 		struct printbuf buf = PRINTBUF;
160 
161 		buf.atomic++;
162 		print_cycle(&buf, g);
163 
164 		trace_trans_restart_would_deadlock(trans, buf.buf);
165 		printbuf_exit(&buf);
166 	}
167 }
168 
169 static int abort_lock(struct lock_graph *g, struct trans_waiting_for_lock *i)
170 {
171 	if (i == g->g) {
172 		trace_would_deadlock(g, i->trans);
173 		return btree_trans_restart(i->trans, BCH_ERR_transaction_restart_would_deadlock);
174 	} else {
175 		i->trans->lock_must_abort = true;
176 		wake_up_process(i->trans->locking_wait.task);
177 		return 0;
178 	}
179 }
180 
181 static int btree_trans_abort_preference(struct btree_trans *trans)
182 {
183 	if (trans->lock_may_not_fail)
184 		return 0;
185 	if (trans->locking_wait.lock_want == SIX_LOCK_write)
186 		return 1;
187 	if (!trans->in_traverse_all)
188 		return 2;
189 	return 3;
190 }
191 
192 static noinline int break_cycle(struct lock_graph *g, struct printbuf *cycle)
193 {
194 	struct trans_waiting_for_lock *i, *abort = NULL;
195 	unsigned best = 0, pref;
196 	int ret;
197 
198 	if (lock_graph_remove_non_waiters(g))
199 		return 0;
200 
201 	/* Only checking, for debugfs: */
202 	if (cycle) {
203 		print_cycle(cycle, g);
204 		ret = -1;
205 		goto out;
206 	}
207 
208 	for (i = g->g; i < g->g + g->nr; i++) {
209 		pref = btree_trans_abort_preference(i->trans);
210 		if (pref > best) {
211 			abort = i;
212 			best = pref;
213 		}
214 	}
215 
216 	if (unlikely(!best)) {
217 		struct printbuf buf = PRINTBUF;
218 
219 		prt_printf(&buf, bch2_fmt(g->g->trans->c, "cycle of nofail locks"));
220 
221 		for (i = g->g; i < g->g + g->nr; i++) {
222 			struct btree_trans *trans = i->trans;
223 
224 			bch2_btree_trans_to_text(&buf, trans);
225 
226 			prt_printf(&buf, "backtrace:\n");
227 			printbuf_indent_add(&buf, 2);
228 			bch2_prt_task_backtrace(&buf, trans->locking_wait.task, 2, GFP_NOWAIT);
229 			printbuf_indent_sub(&buf, 2);
230 			prt_newline(&buf);
231 		}
232 
233 		bch2_print_string_as_lines(KERN_ERR, buf.buf);
234 		printbuf_exit(&buf);
235 		BUG();
236 	}
237 
238 	ret = abort_lock(g, abort);
239 out:
240 	if (ret)
241 		while (g->nr)
242 			lock_graph_up(g);
243 	return ret;
244 }
245 
246 static int lock_graph_descend(struct lock_graph *g, struct btree_trans *trans,
247 			      struct printbuf *cycle)
248 {
249 	struct btree_trans *orig_trans = g->g->trans;
250 	struct trans_waiting_for_lock *i;
251 
252 	for (i = g->g; i < g->g + g->nr; i++)
253 		if (i->trans == trans) {
254 			closure_put(&trans->ref);
255 			return break_cycle(g, cycle);
256 		}
257 
258 	if (g->nr == ARRAY_SIZE(g->g)) {
259 		closure_put(&trans->ref);
260 
261 		if (orig_trans->lock_may_not_fail)
262 			return 0;
263 
264 		while (g->nr)
265 			lock_graph_up(g);
266 
267 		if (cycle)
268 			return 0;
269 
270 		trace_and_count(trans->c, trans_restart_would_deadlock_recursion_limit, trans, _RET_IP_);
271 		return btree_trans_restart(orig_trans, BCH_ERR_transaction_restart_deadlock_recursion_limit);
272 	}
273 
274 	__lock_graph_down(g, trans);
275 	return 0;
276 }
277 
278 static bool lock_type_conflicts(enum six_lock_type t1, enum six_lock_type t2)
279 {
280 	return t1 + t2 > 1;
281 }
282 
283 int bch2_check_for_deadlock(struct btree_trans *trans, struct printbuf *cycle)
284 {
285 	struct lock_graph g;
286 	struct trans_waiting_for_lock *top;
287 	struct btree_bkey_cached_common *b;
288 	btree_path_idx_t path_idx;
289 	int ret = 0;
290 
291 	g.nr = 0;
292 
293 	if (trans->lock_must_abort) {
294 		if (cycle)
295 			return -1;
296 
297 		trace_would_deadlock(&g, trans);
298 		return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock);
299 	}
300 
301 	lock_graph_down(&g, trans);
302 
303 	/* trans->paths is rcu protected vs. freeing */
304 	rcu_read_lock();
305 	if (cycle)
306 		cycle->atomic++;
307 next:
308 	if (!g.nr)
309 		goto out;
310 
311 	top = &g.g[g.nr - 1];
312 
313 	struct btree_path *paths = rcu_dereference(top->trans->paths);
314 	if (!paths)
315 		goto up;
316 
317 	unsigned long *paths_allocated = trans_paths_allocated(paths);
318 
319 	trans_for_each_path_idx_from(paths_allocated, *trans_paths_nr(paths),
320 				     path_idx, top->path_idx) {
321 		struct btree_path *path = paths + path_idx;
322 		if (!path->nodes_locked)
323 			continue;
324 
325 		if (path_idx != top->path_idx) {
326 			top->path_idx		= path_idx;
327 			top->level		= 0;
328 			top->lock_start_time	= 0;
329 		}
330 
331 		for (;
332 		     top->level < BTREE_MAX_DEPTH;
333 		     top->level++, top->lock_start_time = 0) {
334 			int lock_held = btree_node_locked_type(path, top->level);
335 
336 			if (lock_held == BTREE_NODE_UNLOCKED)
337 				continue;
338 
339 			b = &READ_ONCE(path->l[top->level].b)->c;
340 
341 			if (IS_ERR_OR_NULL(b)) {
342 				/*
343 				 * If we get here, it means we raced with the
344 				 * other thread updating its btree_path
345 				 * structures - which means it can't be blocked
346 				 * waiting on a lock:
347 				 */
348 				if (!lock_graph_remove_non_waiters(&g)) {
349 					/*
350 					 * If lock_graph_remove_non_waiters()
351 					 * didn't do anything, it must be
352 					 * because we're being called by debugfs
353 					 * checking for lock cycles, which
354 					 * invokes us on btree_transactions that
355 					 * aren't actually waiting on anything.
356 					 * Just bail out:
357 					 */
358 					lock_graph_pop_all(&g);
359 				}
360 
361 				goto next;
362 			}
363 
364 			if (list_empty_careful(&b->lock.wait_list))
365 				continue;
366 
367 			raw_spin_lock(&b->lock.wait_lock);
368 			list_for_each_entry(trans, &b->lock.wait_list, locking_wait.list) {
369 				BUG_ON(b != trans->locking);
370 
371 				if (top->lock_start_time &&
372 				    time_after_eq64(top->lock_start_time, trans->locking_wait.start_time))
373 					continue;
374 
375 				top->lock_start_time = trans->locking_wait.start_time;
376 
377 				/* Don't check for self deadlock: */
378 				if (trans == top->trans ||
379 				    !lock_type_conflicts(lock_held, trans->locking_wait.lock_want))
380 					continue;
381 
382 				closure_get(&trans->ref);
383 				raw_spin_unlock(&b->lock.wait_lock);
384 
385 				ret = lock_graph_descend(&g, trans, cycle);
386 				if (ret)
387 					goto out;
388 				goto next;
389 
390 			}
391 			raw_spin_unlock(&b->lock.wait_lock);
392 		}
393 	}
394 up:
395 	if (g.nr > 1 && cycle)
396 		print_chain(cycle, &g);
397 	lock_graph_up(&g);
398 	goto next;
399 out:
400 	if (cycle)
401 		--cycle->atomic;
402 	rcu_read_unlock();
403 	return ret;
404 }
405 
406 int bch2_six_check_for_deadlock(struct six_lock *lock, void *p)
407 {
408 	struct btree_trans *trans = p;
409 
410 	return bch2_check_for_deadlock(trans, NULL);
411 }
412 
413 int __bch2_btree_node_lock_write(struct btree_trans *trans, struct btree_path *path,
414 				 struct btree_bkey_cached_common *b,
415 				 bool lock_may_not_fail)
416 {
417 	int readers = bch2_btree_node_lock_counts(trans, NULL, b, b->level).n[SIX_LOCK_read];
418 	int ret;
419 
420 	/*
421 	 * Must drop our read locks before calling six_lock_write() -
422 	 * six_unlock() won't do wakeups until the reader count
423 	 * goes to 0, and it's safe because we have the node intent
424 	 * locked:
425 	 */
426 	six_lock_readers_add(&b->lock, -readers);
427 	ret = __btree_node_lock_nopath(trans, b, SIX_LOCK_write,
428 				       lock_may_not_fail, _RET_IP_);
429 	six_lock_readers_add(&b->lock, readers);
430 
431 	if (ret)
432 		mark_btree_node_locked_noreset(path, b->level, BTREE_NODE_INTENT_LOCKED);
433 
434 	return ret;
435 }
436 
437 void bch2_btree_node_lock_write_nofail(struct btree_trans *trans,
438 				       struct btree_path *path,
439 				       struct btree_bkey_cached_common *b)
440 {
441 	int ret = __btree_node_lock_write(trans, path, b, true);
442 	BUG_ON(ret);
443 }
444 
445 /* relock */
446 
447 static inline bool btree_path_get_locks(struct btree_trans *trans,
448 					struct btree_path *path,
449 					bool upgrade,
450 					struct get_locks_fail *f)
451 {
452 	unsigned l = path->level;
453 	int fail_idx = -1;
454 
455 	do {
456 		if (!btree_path_node(path, l))
457 			break;
458 
459 		if (!(upgrade
460 		      ? bch2_btree_node_upgrade(trans, path, l)
461 		      : bch2_btree_node_relock(trans, path, l))) {
462 			fail_idx	= l;
463 
464 			if (f) {
465 				f->l	= l;
466 				f->b	= path->l[l].b;
467 			}
468 		}
469 
470 		l++;
471 	} while (l < path->locks_want);
472 
473 	/*
474 	 * When we fail to get a lock, we have to ensure that any child nodes
475 	 * can't be relocked so bch2_btree_path_traverse has to walk back up to
476 	 * the node that we failed to relock:
477 	 */
478 	if (fail_idx >= 0) {
479 		__bch2_btree_path_unlock(trans, path);
480 		btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
481 
482 		do {
483 			path->l[fail_idx].b = upgrade
484 				? ERR_PTR(-BCH_ERR_no_btree_node_upgrade)
485 				: ERR_PTR(-BCH_ERR_no_btree_node_relock);
486 			--fail_idx;
487 		} while (fail_idx >= 0);
488 	}
489 
490 	if (path->uptodate == BTREE_ITER_NEED_RELOCK)
491 		path->uptodate = BTREE_ITER_UPTODATE;
492 
493 	return path->uptodate < BTREE_ITER_NEED_RELOCK;
494 }
495 
496 bool __bch2_btree_node_relock(struct btree_trans *trans,
497 			      struct btree_path *path, unsigned level,
498 			      bool trace)
499 {
500 	struct btree *b = btree_path_node(path, level);
501 	int want = __btree_lock_want(path, level);
502 
503 	if (race_fault())
504 		goto fail;
505 
506 	if (six_relock_type(&b->c.lock, want, path->l[level].lock_seq) ||
507 	    (btree_node_lock_seq_matches(path, b, level) &&
508 	     btree_node_lock_increment(trans, &b->c, level, want))) {
509 		mark_btree_node_locked(trans, path, level, want);
510 		return true;
511 	}
512 fail:
513 	if (trace && !trans->notrace_relock_fail)
514 		trace_and_count(trans->c, btree_path_relock_fail, trans, _RET_IP_, path, level);
515 	return false;
516 }
517 
518 /* upgrade */
519 
520 bool bch2_btree_node_upgrade(struct btree_trans *trans,
521 			     struct btree_path *path, unsigned level)
522 {
523 	struct btree *b = path->l[level].b;
524 	struct six_lock_count count = bch2_btree_node_lock_counts(trans, path, &b->c, level);
525 
526 	if (!is_btree_node(path, level))
527 		return false;
528 
529 	switch (btree_lock_want(path, level)) {
530 	case BTREE_NODE_UNLOCKED:
531 		BUG_ON(btree_node_locked(path, level));
532 		return true;
533 	case BTREE_NODE_READ_LOCKED:
534 		BUG_ON(btree_node_intent_locked(path, level));
535 		return bch2_btree_node_relock(trans, path, level);
536 	case BTREE_NODE_INTENT_LOCKED:
537 		break;
538 	case BTREE_NODE_WRITE_LOCKED:
539 		BUG();
540 	}
541 
542 	if (btree_node_intent_locked(path, level))
543 		return true;
544 
545 	if (race_fault())
546 		return false;
547 
548 	if (btree_node_locked(path, level)) {
549 		bool ret;
550 
551 		six_lock_readers_add(&b->c.lock, -count.n[SIX_LOCK_read]);
552 		ret = six_lock_tryupgrade(&b->c.lock);
553 		six_lock_readers_add(&b->c.lock, count.n[SIX_LOCK_read]);
554 
555 		if (ret)
556 			goto success;
557 	} else {
558 		if (six_relock_type(&b->c.lock, SIX_LOCK_intent, path->l[level].lock_seq))
559 			goto success;
560 	}
561 
562 	/*
563 	 * Do we already have an intent lock via another path? If so, just bump
564 	 * lock count:
565 	 */
566 	if (btree_node_lock_seq_matches(path, b, level) &&
567 	    btree_node_lock_increment(trans, &b->c, level, BTREE_NODE_INTENT_LOCKED)) {
568 		btree_node_unlock(trans, path, level);
569 		goto success;
570 	}
571 
572 	trace_and_count(trans->c, btree_path_upgrade_fail, trans, _RET_IP_, path, level);
573 	return false;
574 success:
575 	mark_btree_node_locked_noreset(path, level, BTREE_NODE_INTENT_LOCKED);
576 	return true;
577 }
578 
579 /* Btree path locking: */
580 
581 /*
582  * Only for btree_cache.c - only relocks intent locks
583  */
584 int bch2_btree_path_relock_intent(struct btree_trans *trans,
585 				  struct btree_path *path)
586 {
587 	unsigned l;
588 
589 	for (l = path->level;
590 	     l < path->locks_want && btree_path_node(path, l);
591 	     l++) {
592 		if (!bch2_btree_node_relock(trans, path, l)) {
593 			__bch2_btree_path_unlock(trans, path);
594 			btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
595 			trace_and_count(trans->c, trans_restart_relock_path_intent, trans, _RET_IP_, path);
596 			return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_path_intent);
597 		}
598 	}
599 
600 	return 0;
601 }
602 
603 __flatten
604 bool bch2_btree_path_relock_norestart(struct btree_trans *trans, struct btree_path *path)
605 {
606 	struct get_locks_fail f;
607 
608 	bool ret = btree_path_get_locks(trans, path, false, &f);
609 	bch2_trans_verify_locks(trans);
610 	return ret;
611 }
612 
613 int __bch2_btree_path_relock(struct btree_trans *trans,
614 			struct btree_path *path, unsigned long trace_ip)
615 {
616 	if (!bch2_btree_path_relock_norestart(trans, path)) {
617 		trace_and_count(trans->c, trans_restart_relock_path, trans, trace_ip, path);
618 		return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_path);
619 	}
620 
621 	return 0;
622 }
623 
624 bool bch2_btree_path_upgrade_noupgrade_sibs(struct btree_trans *trans,
625 			       struct btree_path *path,
626 			       unsigned new_locks_want,
627 			       struct get_locks_fail *f)
628 {
629 	EBUG_ON(path->locks_want >= new_locks_want);
630 
631 	path->locks_want = new_locks_want;
632 
633 	bool ret = btree_path_get_locks(trans, path, true, f);
634 	bch2_trans_verify_locks(trans);
635 	return ret;
636 }
637 
638 bool __bch2_btree_path_upgrade(struct btree_trans *trans,
639 			       struct btree_path *path,
640 			       unsigned new_locks_want,
641 			       struct get_locks_fail *f)
642 {
643 	bool ret = bch2_btree_path_upgrade_noupgrade_sibs(trans, path, new_locks_want, f);
644 	if (ret)
645 		goto out;
646 
647 	/*
648 	 * XXX: this is ugly - we'd prefer to not be mucking with other
649 	 * iterators in the btree_trans here.
650 	 *
651 	 * On failure to upgrade the iterator, setting iter->locks_want and
652 	 * calling get_locks() is sufficient to make bch2_btree_path_traverse()
653 	 * get the locks we want on transaction restart.
654 	 *
655 	 * But if this iterator was a clone, on transaction restart what we did
656 	 * to this iterator isn't going to be preserved.
657 	 *
658 	 * Possibly we could add an iterator field for the parent iterator when
659 	 * an iterator is a copy - for now, we'll just upgrade any other
660 	 * iterators with the same btree id.
661 	 *
662 	 * The code below used to be needed to ensure ancestor nodes get locked
663 	 * before interior nodes - now that's handled by
664 	 * bch2_btree_path_traverse_all().
665 	 */
666 	if (!path->cached && !trans->in_traverse_all) {
667 		struct btree_path *linked;
668 		unsigned i;
669 
670 		trans_for_each_path(trans, linked, i)
671 			if (linked != path &&
672 			    linked->cached == path->cached &&
673 			    linked->btree_id == path->btree_id &&
674 			    linked->locks_want < new_locks_want) {
675 				linked->locks_want = new_locks_want;
676 				btree_path_get_locks(trans, linked, true, NULL);
677 			}
678 	}
679 out:
680 	bch2_trans_verify_locks(trans);
681 	return ret;
682 }
683 
684 void __bch2_btree_path_downgrade(struct btree_trans *trans,
685 				 struct btree_path *path,
686 				 unsigned new_locks_want)
687 {
688 	unsigned l, old_locks_want = path->locks_want;
689 
690 	if (trans->restarted)
691 		return;
692 
693 	EBUG_ON(path->locks_want < new_locks_want);
694 
695 	path->locks_want = new_locks_want;
696 
697 	while (path->nodes_locked &&
698 	       (l = btree_path_highest_level_locked(path)) >= path->locks_want) {
699 		if (l > path->level) {
700 			btree_node_unlock(trans, path, l);
701 		} else {
702 			if (btree_node_intent_locked(path, l)) {
703 				six_lock_downgrade(&path->l[l].b->c.lock);
704 				mark_btree_node_locked_noreset(path, l, BTREE_NODE_READ_LOCKED);
705 			}
706 			break;
707 		}
708 	}
709 
710 	bch2_btree_path_verify_locks(path);
711 
712 	trace_path_downgrade(trans, _RET_IP_, path, old_locks_want);
713 }
714 
715 /* Btree transaction locking: */
716 
717 void bch2_trans_downgrade(struct btree_trans *trans)
718 {
719 	struct btree_path *path;
720 	unsigned i;
721 
722 	if (trans->restarted)
723 		return;
724 
725 	trans_for_each_path(trans, path, i)
726 		if (path->ref)
727 			bch2_btree_path_downgrade(trans, path);
728 }
729 
730 static inline void __bch2_trans_unlock(struct btree_trans *trans)
731 {
732 	struct btree_path *path;
733 	unsigned i;
734 
735 	trans_for_each_path(trans, path, i)
736 		__bch2_btree_path_unlock(trans, path);
737 }
738 
739 static noinline __cold int bch2_trans_relock_fail(struct btree_trans *trans, struct btree_path *path,
740 						  struct get_locks_fail *f, bool trace)
741 {
742 	if (!trace)
743 		goto out;
744 
745 	if (trace_trans_restart_relock_enabled()) {
746 		struct printbuf buf = PRINTBUF;
747 
748 		bch2_bpos_to_text(&buf, path->pos);
749 		prt_printf(&buf, " l=%u seq=%u node seq=", f->l, path->l[f->l].lock_seq);
750 		if (IS_ERR_OR_NULL(f->b)) {
751 			prt_str(&buf, bch2_err_str(PTR_ERR(f->b)));
752 		} else {
753 			prt_printf(&buf, "%u", f->b->c.lock.seq);
754 
755 			struct six_lock_count c =
756 				bch2_btree_node_lock_counts(trans, NULL, &f->b->c, f->l);
757 			prt_printf(&buf, " self locked %u.%u.%u", c.n[0], c.n[1], c.n[2]);
758 
759 			c = six_lock_counts(&f->b->c.lock);
760 			prt_printf(&buf, " total locked %u.%u.%u", c.n[0], c.n[1], c.n[2]);
761 		}
762 
763 		trace_trans_restart_relock(trans, _RET_IP_, buf.buf);
764 		printbuf_exit(&buf);
765 	}
766 
767 	count_event(trans->c, trans_restart_relock);
768 out:
769 	__bch2_trans_unlock(trans);
770 	bch2_trans_verify_locks(trans);
771 	return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
772 }
773 
774 static inline int __bch2_trans_relock(struct btree_trans *trans, bool trace)
775 {
776 	bch2_trans_verify_locks(trans);
777 
778 	if (unlikely(trans->restarted))
779 		return -((int) trans->restarted);
780 	if (unlikely(trans->locked))
781 		goto out;
782 
783 	struct btree_path *path;
784 	unsigned i;
785 
786 	trans_for_each_path(trans, path, i) {
787 		struct get_locks_fail f;
788 
789 		if (path->should_be_locked &&
790 		    !btree_path_get_locks(trans, path, false, &f))
791 			return bch2_trans_relock_fail(trans, path, &f, trace);
792 	}
793 
794 	trans->locked = true;
795 out:
796 	bch2_trans_verify_locks(trans);
797 	return 0;
798 }
799 
800 int bch2_trans_relock(struct btree_trans *trans)
801 {
802 	return __bch2_trans_relock(trans, true);
803 }
804 
805 int bch2_trans_relock_notrace(struct btree_trans *trans)
806 {
807 	return __bch2_trans_relock(trans, false);
808 }
809 
810 void bch2_trans_unlock_noassert(struct btree_trans *trans)
811 {
812 	__bch2_trans_unlock(trans);
813 
814 	trans->locked = false;
815 	trans->last_unlock_ip = _RET_IP_;
816 }
817 
818 void bch2_trans_unlock(struct btree_trans *trans)
819 {
820 	__bch2_trans_unlock(trans);
821 
822 	trans->locked = false;
823 	trans->last_unlock_ip = _RET_IP_;
824 }
825 
826 void bch2_trans_unlock_long(struct btree_trans *trans)
827 {
828 	bch2_trans_unlock(trans);
829 	bch2_trans_srcu_unlock(trans);
830 }
831 
832 int __bch2_trans_mutex_lock(struct btree_trans *trans,
833 			    struct mutex *lock)
834 {
835 	int ret = drop_locks_do(trans, (mutex_lock(lock), 0));
836 
837 	if (ret)
838 		mutex_unlock(lock);
839 	return ret;
840 }
841 
842 /* Debug */
843 
844 #ifdef CONFIG_BCACHEFS_DEBUG
845 
846 void bch2_btree_path_verify_locks(struct btree_path *path)
847 {
848 	/*
849 	 * A path may be uptodate and yet have nothing locked if and only if
850 	 * there is no node at path->level, which generally means we were
851 	 * iterating over all nodes and got to the end of the btree
852 	 */
853 	BUG_ON(path->uptodate == BTREE_ITER_UPTODATE &&
854 	       btree_path_node(path, path->level) &&
855 	       !path->nodes_locked);
856 
857 	if (!path->nodes_locked)
858 		return;
859 
860 	for (unsigned l = 0; l < BTREE_MAX_DEPTH; l++) {
861 		int want = btree_lock_want(path, l);
862 		int have = btree_node_locked_type(path, l);
863 
864 		BUG_ON(!is_btree_node(path, l) && have != BTREE_NODE_UNLOCKED);
865 
866 		BUG_ON(is_btree_node(path, l) &&
867 		       (want == BTREE_NODE_UNLOCKED ||
868 			have != BTREE_NODE_WRITE_LOCKED) &&
869 		       want != have);
870 	}
871 }
872 
873 static bool bch2_trans_locked(struct btree_trans *trans)
874 {
875 	struct btree_path *path;
876 	unsigned i;
877 
878 	trans_for_each_path(trans, path, i)
879 		if (path->nodes_locked)
880 			return true;
881 	return false;
882 }
883 
884 void bch2_trans_verify_locks(struct btree_trans *trans)
885 {
886 	if (!trans->locked) {
887 		BUG_ON(bch2_trans_locked(trans));
888 		return;
889 	}
890 
891 	struct btree_path *path;
892 	unsigned i;
893 
894 	trans_for_each_path(trans, path, i)
895 		bch2_btree_path_verify_locks(path);
896 }
897 
898 #endif
899