xref: /linux/fs/bcachefs/btree_iter.c (revision e7d759f31ca295d589f7420719c311870bb3166f)
1 // SPDX-License-Identifier: GPL-2.0
2 
3 #include "bcachefs.h"
4 #include "bkey_methods.h"
5 #include "bkey_buf.h"
6 #include "btree_cache.h"
7 #include "btree_iter.h"
8 #include "btree_journal_iter.h"
9 #include "btree_key_cache.h"
10 #include "btree_locking.h"
11 #include "btree_update.h"
12 #include "debug.h"
13 #include "error.h"
14 #include "extents.h"
15 #include "journal.h"
16 #include "journal_io.h"
17 #include "replicas.h"
18 #include "snapshot.h"
19 #include "trace.h"
20 
21 #include <linux/random.h>
22 #include <linux/prefetch.h>
23 
24 static inline void btree_path_list_remove(struct btree_trans *, struct btree_path *);
25 static inline void btree_path_list_add(struct btree_trans *,
26 			btree_path_idx_t, btree_path_idx_t);
27 
28 static inline unsigned long btree_iter_ip_allocated(struct btree_iter *iter)
29 {
30 #ifdef TRACK_PATH_ALLOCATED
31 	return iter->ip_allocated;
32 #else
33 	return 0;
34 #endif
35 }
36 
37 static btree_path_idx_t btree_path_alloc(struct btree_trans *, btree_path_idx_t);
38 static void bch2_trans_srcu_lock(struct btree_trans *);
39 
40 static inline int __btree_path_cmp(const struct btree_path *l,
41 				   enum btree_id	r_btree_id,
42 				   bool			r_cached,
43 				   struct bpos		r_pos,
44 				   unsigned		r_level)
45 {
46 	/*
47 	 * Must match lock ordering as defined by __bch2_btree_node_lock:
48 	 */
49 	return   cmp_int(l->btree_id,	r_btree_id) ?:
50 		 cmp_int((int) l->cached,	(int) r_cached) ?:
51 		 bpos_cmp(l->pos,	r_pos) ?:
52 		-cmp_int(l->level,	r_level);
53 }
54 
55 static inline int btree_path_cmp(const struct btree_path *l,
56 				 const struct btree_path *r)
57 {
58 	return __btree_path_cmp(l, r->btree_id, r->cached, r->pos, r->level);
59 }
60 
61 static inline struct bpos bkey_successor(struct btree_iter *iter, struct bpos p)
62 {
63 	/* Are we iterating over keys in all snapshots? */
64 	if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
65 		p = bpos_successor(p);
66 	} else {
67 		p = bpos_nosnap_successor(p);
68 		p.snapshot = iter->snapshot;
69 	}
70 
71 	return p;
72 }
73 
74 static inline struct bpos bkey_predecessor(struct btree_iter *iter, struct bpos p)
75 {
76 	/* Are we iterating over keys in all snapshots? */
77 	if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
78 		p = bpos_predecessor(p);
79 	} else {
80 		p = bpos_nosnap_predecessor(p);
81 		p.snapshot = iter->snapshot;
82 	}
83 
84 	return p;
85 }
86 
87 static inline struct bpos btree_iter_search_key(struct btree_iter *iter)
88 {
89 	struct bpos pos = iter->pos;
90 
91 	if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
92 	    !bkey_eq(pos, POS_MAX))
93 		pos = bkey_successor(iter, pos);
94 	return pos;
95 }
96 
97 static inline bool btree_path_pos_before_node(struct btree_path *path,
98 					      struct btree *b)
99 {
100 	return bpos_lt(path->pos, b->data->min_key);
101 }
102 
103 static inline bool btree_path_pos_after_node(struct btree_path *path,
104 					     struct btree *b)
105 {
106 	return bpos_gt(path->pos, b->key.k.p);
107 }
108 
109 static inline bool btree_path_pos_in_node(struct btree_path *path,
110 					  struct btree *b)
111 {
112 	return path->btree_id == b->c.btree_id &&
113 		!btree_path_pos_before_node(path, b) &&
114 		!btree_path_pos_after_node(path, b);
115 }
116 
117 /* Btree iterator: */
118 
119 #ifdef CONFIG_BCACHEFS_DEBUG
120 
121 static void bch2_btree_path_verify_cached(struct btree_trans *trans,
122 					  struct btree_path *path)
123 {
124 	struct bkey_cached *ck;
125 	bool locked = btree_node_locked(path, 0);
126 
127 	if (!bch2_btree_node_relock(trans, path, 0))
128 		return;
129 
130 	ck = (void *) path->l[0].b;
131 	BUG_ON(ck->key.btree_id != path->btree_id ||
132 	       !bkey_eq(ck->key.pos, path->pos));
133 
134 	if (!locked)
135 		btree_node_unlock(trans, path, 0);
136 }
137 
138 static void bch2_btree_path_verify_level(struct btree_trans *trans,
139 				struct btree_path *path, unsigned level)
140 {
141 	struct btree_path_level *l;
142 	struct btree_node_iter tmp;
143 	bool locked;
144 	struct bkey_packed *p, *k;
145 	struct printbuf buf1 = PRINTBUF;
146 	struct printbuf buf2 = PRINTBUF;
147 	struct printbuf buf3 = PRINTBUF;
148 	const char *msg;
149 
150 	if (!bch2_debug_check_iterators)
151 		return;
152 
153 	l	= &path->l[level];
154 	tmp	= l->iter;
155 	locked	= btree_node_locked(path, level);
156 
157 	if (path->cached) {
158 		if (!level)
159 			bch2_btree_path_verify_cached(trans, path);
160 		return;
161 	}
162 
163 	if (!btree_path_node(path, level))
164 		return;
165 
166 	if (!bch2_btree_node_relock_notrace(trans, path, level))
167 		return;
168 
169 	BUG_ON(!btree_path_pos_in_node(path, l->b));
170 
171 	bch2_btree_node_iter_verify(&l->iter, l->b);
172 
173 	/*
174 	 * For interior nodes, the iterator will have skipped past deleted keys:
175 	 */
176 	p = level
177 		? bch2_btree_node_iter_prev(&tmp, l->b)
178 		: bch2_btree_node_iter_prev_all(&tmp, l->b);
179 	k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
180 
181 	if (p && bkey_iter_pos_cmp(l->b, p, &path->pos) >= 0) {
182 		msg = "before";
183 		goto err;
184 	}
185 
186 	if (k && bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
187 		msg = "after";
188 		goto err;
189 	}
190 
191 	if (!locked)
192 		btree_node_unlock(trans, path, level);
193 	return;
194 err:
195 	bch2_bpos_to_text(&buf1, path->pos);
196 
197 	if (p) {
198 		struct bkey uk = bkey_unpack_key(l->b, p);
199 
200 		bch2_bkey_to_text(&buf2, &uk);
201 	} else {
202 		prt_printf(&buf2, "(none)");
203 	}
204 
205 	if (k) {
206 		struct bkey uk = bkey_unpack_key(l->b, k);
207 
208 		bch2_bkey_to_text(&buf3, &uk);
209 	} else {
210 		prt_printf(&buf3, "(none)");
211 	}
212 
213 	panic("path should be %s key at level %u:\n"
214 	      "path pos %s\n"
215 	      "prev key %s\n"
216 	      "cur  key %s\n",
217 	      msg, level, buf1.buf, buf2.buf, buf3.buf);
218 }
219 
220 static void bch2_btree_path_verify(struct btree_trans *trans,
221 				   struct btree_path *path)
222 {
223 	struct bch_fs *c = trans->c;
224 	unsigned i;
225 
226 	EBUG_ON(path->btree_id >= BTREE_ID_NR);
227 
228 	for (i = 0; i < (!path->cached ? BTREE_MAX_DEPTH : 1); i++) {
229 		if (!path->l[i].b) {
230 			BUG_ON(!path->cached &&
231 			       bch2_btree_id_root(c, path->btree_id)->b->c.level > i);
232 			break;
233 		}
234 
235 		bch2_btree_path_verify_level(trans, path, i);
236 	}
237 
238 	bch2_btree_path_verify_locks(path);
239 }
240 
241 void bch2_trans_verify_paths(struct btree_trans *trans)
242 {
243 	struct btree_path *path;
244 	unsigned iter;
245 
246 	trans_for_each_path(trans, path, iter)
247 		bch2_btree_path_verify(trans, path);
248 }
249 
250 static void bch2_btree_iter_verify(struct btree_iter *iter)
251 {
252 	struct btree_trans *trans = iter->trans;
253 
254 	BUG_ON(iter->btree_id >= BTREE_ID_NR);
255 
256 	BUG_ON(!!(iter->flags & BTREE_ITER_CACHED) != btree_iter_path(trans, iter)->cached);
257 
258 	BUG_ON((iter->flags & BTREE_ITER_IS_EXTENTS) &&
259 	       (iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
260 
261 	BUG_ON(!(iter->flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
262 	       (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
263 	       !btree_type_has_snapshot_field(iter->btree_id));
264 
265 	if (iter->update_path)
266 		bch2_btree_path_verify(trans, &trans->paths[iter->update_path]);
267 	bch2_btree_path_verify(trans, btree_iter_path(trans, iter));
268 }
269 
270 static void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter)
271 {
272 	BUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
273 	       !iter->pos.snapshot);
274 
275 	BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
276 	       iter->pos.snapshot != iter->snapshot);
277 
278 	BUG_ON(bkey_lt(iter->pos, bkey_start_pos(&iter->k)) ||
279 	       bkey_gt(iter->pos, iter->k.p));
280 }
281 
282 static int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k)
283 {
284 	struct btree_trans *trans = iter->trans;
285 	struct btree_iter copy;
286 	struct bkey_s_c prev;
287 	int ret = 0;
288 
289 	if (!bch2_debug_check_iterators)
290 		return 0;
291 
292 	if (!(iter->flags & BTREE_ITER_FILTER_SNAPSHOTS))
293 		return 0;
294 
295 	if (bkey_err(k) || !k.k)
296 		return 0;
297 
298 	BUG_ON(!bch2_snapshot_is_ancestor(trans->c,
299 					  iter->snapshot,
300 					  k.k->p.snapshot));
301 
302 	bch2_trans_iter_init(trans, &copy, iter->btree_id, iter->pos,
303 			     BTREE_ITER_NOPRESERVE|
304 			     BTREE_ITER_ALL_SNAPSHOTS);
305 	prev = bch2_btree_iter_prev(&copy);
306 	if (!prev.k)
307 		goto out;
308 
309 	ret = bkey_err(prev);
310 	if (ret)
311 		goto out;
312 
313 	if (bkey_eq(prev.k->p, k.k->p) &&
314 	    bch2_snapshot_is_ancestor(trans->c, iter->snapshot,
315 				      prev.k->p.snapshot) > 0) {
316 		struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
317 
318 		bch2_bkey_to_text(&buf1, k.k);
319 		bch2_bkey_to_text(&buf2, prev.k);
320 
321 		panic("iter snap %u\n"
322 		      "k    %s\n"
323 		      "prev %s\n",
324 		      iter->snapshot,
325 		      buf1.buf, buf2.buf);
326 	}
327 out:
328 	bch2_trans_iter_exit(trans, &copy);
329 	return ret;
330 }
331 
332 void bch2_assert_pos_locked(struct btree_trans *trans, enum btree_id id,
333 			    struct bpos pos, bool key_cache)
334 {
335 	struct btree_path *path;
336 	struct trans_for_each_path_inorder_iter iter;
337 	struct printbuf buf = PRINTBUF;
338 
339 	btree_trans_sort_paths(trans);
340 
341 	trans_for_each_path_inorder(trans, path, iter) {
342 		int cmp = cmp_int(path->btree_id, id) ?:
343 			cmp_int(path->cached, key_cache);
344 
345 		if (cmp > 0)
346 			break;
347 		if (cmp < 0)
348 			continue;
349 
350 		if (!btree_node_locked(path, 0) ||
351 		    !path->should_be_locked)
352 			continue;
353 
354 		if (!key_cache) {
355 			if (bkey_ge(pos, path->l[0].b->data->min_key) &&
356 			    bkey_le(pos, path->l[0].b->key.k.p))
357 				return;
358 		} else {
359 			if (bkey_eq(pos, path->pos))
360 				return;
361 		}
362 	}
363 
364 	bch2_dump_trans_paths_updates(trans);
365 	bch2_bpos_to_text(&buf, pos);
366 
367 	panic("not locked: %s %s%s\n",
368 	      bch2_btree_id_str(id), buf.buf,
369 	      key_cache ? " cached" : "");
370 }
371 
372 #else
373 
374 static inline void bch2_btree_path_verify_level(struct btree_trans *trans,
375 						struct btree_path *path, unsigned l) {}
376 static inline void bch2_btree_path_verify(struct btree_trans *trans,
377 					  struct btree_path *path) {}
378 static inline void bch2_btree_iter_verify(struct btree_iter *iter) {}
379 static inline void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) {}
380 static inline int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k) { return 0; }
381 
382 #endif
383 
384 /* Btree path: fixups after btree updates */
385 
386 static void btree_node_iter_set_set_pos(struct btree_node_iter *iter,
387 					struct btree *b,
388 					struct bset_tree *t,
389 					struct bkey_packed *k)
390 {
391 	struct btree_node_iter_set *set;
392 
393 	btree_node_iter_for_each(iter, set)
394 		if (set->end == t->end_offset) {
395 			set->k = __btree_node_key_to_offset(b, k);
396 			bch2_btree_node_iter_sort(iter, b);
397 			return;
398 		}
399 
400 	bch2_btree_node_iter_push(iter, b, k, btree_bkey_last(b, t));
401 }
402 
403 static void __bch2_btree_path_fix_key_modified(struct btree_path *path,
404 					       struct btree *b,
405 					       struct bkey_packed *where)
406 {
407 	struct btree_path_level *l = &path->l[b->c.level];
408 
409 	if (where != bch2_btree_node_iter_peek_all(&l->iter, l->b))
410 		return;
411 
412 	if (bkey_iter_pos_cmp(l->b, where, &path->pos) < 0)
413 		bch2_btree_node_iter_advance(&l->iter, l->b);
414 }
415 
416 void bch2_btree_path_fix_key_modified(struct btree_trans *trans,
417 				      struct btree *b,
418 				      struct bkey_packed *where)
419 {
420 	struct btree_path *path;
421 	unsigned i;
422 
423 	trans_for_each_path_with_node(trans, b, path, i) {
424 		__bch2_btree_path_fix_key_modified(path, b, where);
425 		bch2_btree_path_verify_level(trans, path, b->c.level);
426 	}
427 }
428 
429 static void __bch2_btree_node_iter_fix(struct btree_path *path,
430 				       struct btree *b,
431 				       struct btree_node_iter *node_iter,
432 				       struct bset_tree *t,
433 				       struct bkey_packed *where,
434 				       unsigned clobber_u64s,
435 				       unsigned new_u64s)
436 {
437 	const struct bkey_packed *end = btree_bkey_last(b, t);
438 	struct btree_node_iter_set *set;
439 	unsigned offset = __btree_node_key_to_offset(b, where);
440 	int shift = new_u64s - clobber_u64s;
441 	unsigned old_end = t->end_offset - shift;
442 	unsigned orig_iter_pos = node_iter->data[0].k;
443 	bool iter_current_key_modified =
444 		orig_iter_pos >= offset &&
445 		orig_iter_pos <= offset + clobber_u64s;
446 
447 	btree_node_iter_for_each(node_iter, set)
448 		if (set->end == old_end)
449 			goto found;
450 
451 	/* didn't find the bset in the iterator - might have to readd it: */
452 	if (new_u64s &&
453 	    bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
454 		bch2_btree_node_iter_push(node_iter, b, where, end);
455 		goto fixup_done;
456 	} else {
457 		/* Iterator is after key that changed */
458 		return;
459 	}
460 found:
461 	set->end = t->end_offset;
462 
463 	/* Iterator hasn't gotten to the key that changed yet: */
464 	if (set->k < offset)
465 		return;
466 
467 	if (new_u64s &&
468 	    bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
469 		set->k = offset;
470 	} else if (set->k < offset + clobber_u64s) {
471 		set->k = offset + new_u64s;
472 		if (set->k == set->end)
473 			bch2_btree_node_iter_set_drop(node_iter, set);
474 	} else {
475 		/* Iterator is after key that changed */
476 		set->k = (int) set->k + shift;
477 		return;
478 	}
479 
480 	bch2_btree_node_iter_sort(node_iter, b);
481 fixup_done:
482 	if (node_iter->data[0].k != orig_iter_pos)
483 		iter_current_key_modified = true;
484 
485 	/*
486 	 * When a new key is added, and the node iterator now points to that
487 	 * key, the iterator might have skipped past deleted keys that should
488 	 * come after the key the iterator now points to. We have to rewind to
489 	 * before those deleted keys - otherwise
490 	 * bch2_btree_node_iter_prev_all() breaks:
491 	 */
492 	if (!bch2_btree_node_iter_end(node_iter) &&
493 	    iter_current_key_modified &&
494 	    b->c.level) {
495 		struct bkey_packed *k, *k2, *p;
496 
497 		k = bch2_btree_node_iter_peek_all(node_iter, b);
498 
499 		for_each_bset(b, t) {
500 			bool set_pos = false;
501 
502 			if (node_iter->data[0].end == t->end_offset)
503 				continue;
504 
505 			k2 = bch2_btree_node_iter_bset_pos(node_iter, b, t);
506 
507 			while ((p = bch2_bkey_prev_all(b, t, k2)) &&
508 			       bkey_iter_cmp(b, k, p) < 0) {
509 				k2 = p;
510 				set_pos = true;
511 			}
512 
513 			if (set_pos)
514 				btree_node_iter_set_set_pos(node_iter,
515 							    b, t, k2);
516 		}
517 	}
518 }
519 
520 void bch2_btree_node_iter_fix(struct btree_trans *trans,
521 			      struct btree_path *path,
522 			      struct btree *b,
523 			      struct btree_node_iter *node_iter,
524 			      struct bkey_packed *where,
525 			      unsigned clobber_u64s,
526 			      unsigned new_u64s)
527 {
528 	struct bset_tree *t = bch2_bkey_to_bset_inlined(b, where);
529 	struct btree_path *linked;
530 	unsigned i;
531 
532 	if (node_iter != &path->l[b->c.level].iter) {
533 		__bch2_btree_node_iter_fix(path, b, node_iter, t,
534 					   where, clobber_u64s, new_u64s);
535 
536 		if (bch2_debug_check_iterators)
537 			bch2_btree_node_iter_verify(node_iter, b);
538 	}
539 
540 	trans_for_each_path_with_node(trans, b, linked, i) {
541 		__bch2_btree_node_iter_fix(linked, b,
542 					   &linked->l[b->c.level].iter, t,
543 					   where, clobber_u64s, new_u64s);
544 		bch2_btree_path_verify_level(trans, linked, b->c.level);
545 	}
546 }
547 
548 /* Btree path level: pointer to a particular btree node and node iter */
549 
550 static inline struct bkey_s_c __btree_iter_unpack(struct bch_fs *c,
551 						  struct btree_path_level *l,
552 						  struct bkey *u,
553 						  struct bkey_packed *k)
554 {
555 	if (unlikely(!k)) {
556 		/*
557 		 * signal to bch2_btree_iter_peek_slot() that we're currently at
558 		 * a hole
559 		 */
560 		u->type = KEY_TYPE_deleted;
561 		return bkey_s_c_null;
562 	}
563 
564 	return bkey_disassemble(l->b, k, u);
565 }
566 
567 static inline struct bkey_s_c btree_path_level_peek_all(struct bch_fs *c,
568 							struct btree_path_level *l,
569 							struct bkey *u)
570 {
571 	return __btree_iter_unpack(c, l, u,
572 			bch2_btree_node_iter_peek_all(&l->iter, l->b));
573 }
574 
575 static inline struct bkey_s_c btree_path_level_peek(struct btree_trans *trans,
576 						    struct btree_path *path,
577 						    struct btree_path_level *l,
578 						    struct bkey *u)
579 {
580 	struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u,
581 			bch2_btree_node_iter_peek(&l->iter, l->b));
582 
583 	path->pos = k.k ? k.k->p : l->b->key.k.p;
584 	trans->paths_sorted = false;
585 	bch2_btree_path_verify_level(trans, path, l - path->l);
586 	return k;
587 }
588 
589 static inline struct bkey_s_c btree_path_level_prev(struct btree_trans *trans,
590 						    struct btree_path *path,
591 						    struct btree_path_level *l,
592 						    struct bkey *u)
593 {
594 	struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u,
595 			bch2_btree_node_iter_prev(&l->iter, l->b));
596 
597 	path->pos = k.k ? k.k->p : l->b->data->min_key;
598 	trans->paths_sorted = false;
599 	bch2_btree_path_verify_level(trans, path, l - path->l);
600 	return k;
601 }
602 
603 static inline bool btree_path_advance_to_pos(struct btree_path *path,
604 					     struct btree_path_level *l,
605 					     int max_advance)
606 {
607 	struct bkey_packed *k;
608 	int nr_advanced = 0;
609 
610 	while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
611 	       bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
612 		if (max_advance > 0 && nr_advanced >= max_advance)
613 			return false;
614 
615 		bch2_btree_node_iter_advance(&l->iter, l->b);
616 		nr_advanced++;
617 	}
618 
619 	return true;
620 }
621 
622 static inline void __btree_path_level_init(struct btree_path *path,
623 					   unsigned level)
624 {
625 	struct btree_path_level *l = &path->l[level];
626 
627 	bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
628 
629 	/*
630 	 * Iterators to interior nodes should always be pointed at the first non
631 	 * whiteout:
632 	 */
633 	if (level)
634 		bch2_btree_node_iter_peek(&l->iter, l->b);
635 }
636 
637 void bch2_btree_path_level_init(struct btree_trans *trans,
638 				struct btree_path *path,
639 				struct btree *b)
640 {
641 	BUG_ON(path->cached);
642 
643 	EBUG_ON(!btree_path_pos_in_node(path, b));
644 
645 	path->l[b->c.level].lock_seq = six_lock_seq(&b->c.lock);
646 	path->l[b->c.level].b = b;
647 	__btree_path_level_init(path, b->c.level);
648 }
649 
650 /* Btree path: fixups after btree node updates: */
651 
652 static void bch2_trans_revalidate_updates_in_node(struct btree_trans *trans, struct btree *b)
653 {
654 	struct bch_fs *c = trans->c;
655 
656 	trans_for_each_update(trans, i)
657 		if (!i->cached &&
658 		    i->level	== b->c.level &&
659 		    i->btree_id	== b->c.btree_id &&
660 		    bpos_cmp(i->k->k.p, b->data->min_key) >= 0 &&
661 		    bpos_cmp(i->k->k.p, b->data->max_key) <= 0) {
662 			i->old_v = bch2_btree_path_peek_slot(trans->paths + i->path, &i->old_k).v;
663 
664 			if (unlikely(trans->journal_replay_not_finished)) {
665 				struct bkey_i *j_k =
666 					bch2_journal_keys_peek_slot(c, i->btree_id, i->level,
667 								    i->k->k.p);
668 
669 				if (j_k) {
670 					i->old_k = j_k->k;
671 					i->old_v = &j_k->v;
672 				}
673 			}
674 		}
675 }
676 
677 /*
678  * A btree node is being replaced - update the iterator to point to the new
679  * node:
680  */
681 void bch2_trans_node_add(struct btree_trans *trans,
682 			 struct btree_path *path,
683 			 struct btree *b)
684 {
685 	struct btree_path *prev;
686 
687 	BUG_ON(!btree_path_pos_in_node(path, b));
688 
689 	while ((prev = prev_btree_path(trans, path)) &&
690 	       btree_path_pos_in_node(prev, b))
691 		path = prev;
692 
693 	for (;
694 	     path && btree_path_pos_in_node(path, b);
695 	     path = next_btree_path(trans, path))
696 		if (path->uptodate == BTREE_ITER_UPTODATE && !path->cached) {
697 			enum btree_node_locked_type t =
698 				btree_lock_want(path, b->c.level);
699 
700 			if (t != BTREE_NODE_UNLOCKED) {
701 				btree_node_unlock(trans, path, b->c.level);
702 				six_lock_increment(&b->c.lock, (enum six_lock_type) t);
703 				mark_btree_node_locked(trans, path, b->c.level, t);
704 			}
705 
706 			bch2_btree_path_level_init(trans, path, b);
707 		}
708 
709 	bch2_trans_revalidate_updates_in_node(trans, b);
710 }
711 
712 /*
713  * A btree node has been modified in such a way as to invalidate iterators - fix
714  * them:
715  */
716 void bch2_trans_node_reinit_iter(struct btree_trans *trans, struct btree *b)
717 {
718 	struct btree_path *path;
719 	unsigned i;
720 
721 	trans_for_each_path_with_node(trans, b, path, i)
722 		__btree_path_level_init(path, b->c.level);
723 
724 	bch2_trans_revalidate_updates_in_node(trans, b);
725 }
726 
727 /* Btree path: traverse, set_pos: */
728 
729 static inline int btree_path_lock_root(struct btree_trans *trans,
730 				       struct btree_path *path,
731 				       unsigned depth_want,
732 				       unsigned long trace_ip)
733 {
734 	struct bch_fs *c = trans->c;
735 	struct btree *b, **rootp = &bch2_btree_id_root(c, path->btree_id)->b;
736 	enum six_lock_type lock_type;
737 	unsigned i;
738 	int ret;
739 
740 	EBUG_ON(path->nodes_locked);
741 
742 	while (1) {
743 		b = READ_ONCE(*rootp);
744 		path->level = READ_ONCE(b->c.level);
745 
746 		if (unlikely(path->level < depth_want)) {
747 			/*
748 			 * the root is at a lower depth than the depth we want:
749 			 * got to the end of the btree, or we're walking nodes
750 			 * greater than some depth and there are no nodes >=
751 			 * that depth
752 			 */
753 			path->level = depth_want;
754 			for (i = path->level; i < BTREE_MAX_DEPTH; i++)
755 				path->l[i].b = NULL;
756 			return 1;
757 		}
758 
759 		lock_type = __btree_lock_want(path, path->level);
760 		ret = btree_node_lock(trans, path, &b->c,
761 				      path->level, lock_type, trace_ip);
762 		if (unlikely(ret)) {
763 			if (bch2_err_matches(ret, BCH_ERR_lock_fail_root_changed))
764 				continue;
765 			if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
766 				return ret;
767 			BUG();
768 		}
769 
770 		if (likely(b == READ_ONCE(*rootp) &&
771 			   b->c.level == path->level &&
772 			   !race_fault())) {
773 			for (i = 0; i < path->level; i++)
774 				path->l[i].b = ERR_PTR(-BCH_ERR_no_btree_node_lock_root);
775 			path->l[path->level].b = b;
776 			for (i = path->level + 1; i < BTREE_MAX_DEPTH; i++)
777 				path->l[i].b = NULL;
778 
779 			mark_btree_node_locked(trans, path, path->level,
780 					       (enum btree_node_locked_type) lock_type);
781 			bch2_btree_path_level_init(trans, path, b);
782 			return 0;
783 		}
784 
785 		six_unlock_type(&b->c.lock, lock_type);
786 	}
787 }
788 
789 noinline
790 static int btree_path_prefetch(struct btree_trans *trans, struct btree_path *path)
791 {
792 	struct bch_fs *c = trans->c;
793 	struct btree_path_level *l = path_l(path);
794 	struct btree_node_iter node_iter = l->iter;
795 	struct bkey_packed *k;
796 	struct bkey_buf tmp;
797 	unsigned nr = test_bit(BCH_FS_started, &c->flags)
798 		? (path->level > 1 ? 0 :  2)
799 		: (path->level > 1 ? 1 : 16);
800 	bool was_locked = btree_node_locked(path, path->level);
801 	int ret = 0;
802 
803 	bch2_bkey_buf_init(&tmp);
804 
805 	while (nr-- && !ret) {
806 		if (!bch2_btree_node_relock(trans, path, path->level))
807 			break;
808 
809 		bch2_btree_node_iter_advance(&node_iter, l->b);
810 		k = bch2_btree_node_iter_peek(&node_iter, l->b);
811 		if (!k)
812 			break;
813 
814 		bch2_bkey_buf_unpack(&tmp, c, l->b, k);
815 		ret = bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id,
816 					       path->level - 1);
817 	}
818 
819 	if (!was_locked)
820 		btree_node_unlock(trans, path, path->level);
821 
822 	bch2_bkey_buf_exit(&tmp, c);
823 	return ret;
824 }
825 
826 static int btree_path_prefetch_j(struct btree_trans *trans, struct btree_path *path,
827 				 struct btree_and_journal_iter *jiter)
828 {
829 	struct bch_fs *c = trans->c;
830 	struct bkey_s_c k;
831 	struct bkey_buf tmp;
832 	unsigned nr = test_bit(BCH_FS_started, &c->flags)
833 		? (path->level > 1 ? 0 :  2)
834 		: (path->level > 1 ? 1 : 16);
835 	bool was_locked = btree_node_locked(path, path->level);
836 	int ret = 0;
837 
838 	bch2_bkey_buf_init(&tmp);
839 
840 	while (nr-- && !ret) {
841 		if (!bch2_btree_node_relock(trans, path, path->level))
842 			break;
843 
844 		bch2_btree_and_journal_iter_advance(jiter);
845 		k = bch2_btree_and_journal_iter_peek(jiter);
846 		if (!k.k)
847 			break;
848 
849 		bch2_bkey_buf_reassemble(&tmp, c, k);
850 		ret = bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id,
851 					       path->level - 1);
852 	}
853 
854 	if (!was_locked)
855 		btree_node_unlock(trans, path, path->level);
856 
857 	bch2_bkey_buf_exit(&tmp, c);
858 	return ret;
859 }
860 
861 static noinline void btree_node_mem_ptr_set(struct btree_trans *trans,
862 					    struct btree_path *path,
863 					    unsigned plevel, struct btree *b)
864 {
865 	struct btree_path_level *l = &path->l[plevel];
866 	bool locked = btree_node_locked(path, plevel);
867 	struct bkey_packed *k;
868 	struct bch_btree_ptr_v2 *bp;
869 
870 	if (!bch2_btree_node_relock(trans, path, plevel))
871 		return;
872 
873 	k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
874 	BUG_ON(k->type != KEY_TYPE_btree_ptr_v2);
875 
876 	bp = (void *) bkeyp_val(&l->b->format, k);
877 	bp->mem_ptr = (unsigned long)b;
878 
879 	if (!locked)
880 		btree_node_unlock(trans, path, plevel);
881 }
882 
883 static noinline int btree_node_iter_and_journal_peek(struct btree_trans *trans,
884 						     struct btree_path *path,
885 						     unsigned flags,
886 						     struct bkey_buf *out)
887 {
888 	struct bch_fs *c = trans->c;
889 	struct btree_path_level *l = path_l(path);
890 	struct btree_and_journal_iter jiter;
891 	struct bkey_s_c k;
892 	int ret = 0;
893 
894 	__bch2_btree_and_journal_iter_init_node_iter(&jiter, c, l->b, l->iter, path->pos);
895 
896 	k = bch2_btree_and_journal_iter_peek(&jiter);
897 
898 	bch2_bkey_buf_reassemble(out, c, k);
899 
900 	if ((flags & BTREE_ITER_PREFETCH) &&
901 	    c->opts.btree_node_prefetch)
902 		ret = btree_path_prefetch_j(trans, path, &jiter);
903 
904 	bch2_btree_and_journal_iter_exit(&jiter);
905 	return ret;
906 }
907 
908 static __always_inline int btree_path_down(struct btree_trans *trans,
909 					   struct btree_path *path,
910 					   unsigned flags,
911 					   unsigned long trace_ip)
912 {
913 	struct bch_fs *c = trans->c;
914 	struct btree_path_level *l = path_l(path);
915 	struct btree *b;
916 	unsigned level = path->level - 1;
917 	enum six_lock_type lock_type = __btree_lock_want(path, level);
918 	struct bkey_buf tmp;
919 	int ret;
920 
921 	EBUG_ON(!btree_node_locked(path, path->level));
922 
923 	bch2_bkey_buf_init(&tmp);
924 
925 	if (unlikely(trans->journal_replay_not_finished)) {
926 		ret = btree_node_iter_and_journal_peek(trans, path, flags, &tmp);
927 		if (ret)
928 			goto err;
929 	} else {
930 		bch2_bkey_buf_unpack(&tmp, c, l->b,
931 				 bch2_btree_node_iter_peek(&l->iter, l->b));
932 
933 		if ((flags & BTREE_ITER_PREFETCH) &&
934 		    c->opts.btree_node_prefetch) {
935 			ret = btree_path_prefetch(trans, path);
936 			if (ret)
937 				goto err;
938 		}
939 	}
940 
941 	b = bch2_btree_node_get(trans, path, tmp.k, level, lock_type, trace_ip);
942 	ret = PTR_ERR_OR_ZERO(b);
943 	if (unlikely(ret))
944 		goto err;
945 
946 	if (likely(!trans->journal_replay_not_finished &&
947 		   tmp.k->k.type == KEY_TYPE_btree_ptr_v2) &&
948 	    unlikely(b != btree_node_mem_ptr(tmp.k)))
949 		btree_node_mem_ptr_set(trans, path, level + 1, b);
950 
951 	if (btree_node_read_locked(path, level + 1))
952 		btree_node_unlock(trans, path, level + 1);
953 
954 	mark_btree_node_locked(trans, path, level,
955 			       (enum btree_node_locked_type) lock_type);
956 	path->level = level;
957 	bch2_btree_path_level_init(trans, path, b);
958 
959 	bch2_btree_path_verify_locks(path);
960 err:
961 	bch2_bkey_buf_exit(&tmp, c);
962 	return ret;
963 }
964 
965 
966 static int bch2_btree_path_traverse_all(struct btree_trans *trans)
967 {
968 	struct bch_fs *c = trans->c;
969 	struct btree_path *path;
970 	unsigned long trace_ip = _RET_IP_;
971 	unsigned i;
972 	int ret = 0;
973 
974 	if (trans->in_traverse_all)
975 		return -BCH_ERR_transaction_restart_in_traverse_all;
976 
977 	trans->in_traverse_all = true;
978 retry_all:
979 	trans->restarted = 0;
980 	trans->last_restarted_ip = 0;
981 
982 	trans_for_each_path(trans, path, i)
983 		path->should_be_locked = false;
984 
985 	btree_trans_sort_paths(trans);
986 
987 	bch2_trans_unlock(trans);
988 	cond_resched();
989 
990 	if (unlikely(trans->memory_allocation_failure)) {
991 		struct closure cl;
992 
993 		closure_init_stack(&cl);
994 
995 		do {
996 			ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
997 			closure_sync(&cl);
998 		} while (ret);
999 	}
1000 
1001 	/* Now, redo traversals in correct order: */
1002 	i = 0;
1003 	while (i < trans->nr_sorted) {
1004 		btree_path_idx_t idx = trans->sorted[i];
1005 
1006 		/*
1007 		 * Traversing a path can cause another path to be added at about
1008 		 * the same position:
1009 		 */
1010 		if (trans->paths[idx].uptodate) {
1011 			__btree_path_get(&trans->paths[idx], false);
1012 			ret = bch2_btree_path_traverse_one(trans, idx, 0, _THIS_IP_);
1013 			__btree_path_put(&trans->paths[idx], false);
1014 
1015 			if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1016 			    bch2_err_matches(ret, ENOMEM))
1017 				goto retry_all;
1018 			if (ret)
1019 				goto err;
1020 		} else {
1021 			i++;
1022 		}
1023 	}
1024 
1025 	/*
1026 	 * We used to assert that all paths had been traversed here
1027 	 * (path->uptodate < BTREE_ITER_NEED_TRAVERSE); however, since
1028 	 * path->should_be_locked is not set yet, we might have unlocked and
1029 	 * then failed to relock a path - that's fine.
1030 	 */
1031 err:
1032 	bch2_btree_cache_cannibalize_unlock(trans);
1033 
1034 	trans->in_traverse_all = false;
1035 
1036 	trace_and_count(c, trans_traverse_all, trans, trace_ip);
1037 	return ret;
1038 }
1039 
1040 static inline bool btree_path_check_pos_in_node(struct btree_path *path,
1041 						unsigned l, int check_pos)
1042 {
1043 	if (check_pos < 0 && btree_path_pos_before_node(path, path->l[l].b))
1044 		return false;
1045 	if (check_pos > 0 && btree_path_pos_after_node(path, path->l[l].b))
1046 		return false;
1047 	return true;
1048 }
1049 
1050 static inline bool btree_path_good_node(struct btree_trans *trans,
1051 					struct btree_path *path,
1052 					unsigned l, int check_pos)
1053 {
1054 	return is_btree_node(path, l) &&
1055 		bch2_btree_node_relock(trans, path, l) &&
1056 		btree_path_check_pos_in_node(path, l, check_pos);
1057 }
1058 
1059 static void btree_path_set_level_down(struct btree_trans *trans,
1060 				      struct btree_path *path,
1061 				      unsigned new_level)
1062 {
1063 	unsigned l;
1064 
1065 	path->level = new_level;
1066 
1067 	for (l = path->level + 1; l < BTREE_MAX_DEPTH; l++)
1068 		if (btree_lock_want(path, l) == BTREE_NODE_UNLOCKED)
1069 			btree_node_unlock(trans, path, l);
1070 
1071 	btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1072 	bch2_btree_path_verify(trans, path);
1073 }
1074 
1075 static noinline unsigned __btree_path_up_until_good_node(struct btree_trans *trans,
1076 							 struct btree_path *path,
1077 							 int check_pos)
1078 {
1079 	unsigned i, l = path->level;
1080 again:
1081 	while (btree_path_node(path, l) &&
1082 	       !btree_path_good_node(trans, path, l, check_pos))
1083 		__btree_path_set_level_up(trans, path, l++);
1084 
1085 	/* If we need intent locks, take them too: */
1086 	for (i = l + 1;
1087 	     i < path->locks_want && btree_path_node(path, i);
1088 	     i++)
1089 		if (!bch2_btree_node_relock(trans, path, i)) {
1090 			while (l <= i)
1091 				__btree_path_set_level_up(trans, path, l++);
1092 			goto again;
1093 		}
1094 
1095 	return l;
1096 }
1097 
1098 static inline unsigned btree_path_up_until_good_node(struct btree_trans *trans,
1099 						     struct btree_path *path,
1100 						     int check_pos)
1101 {
1102 	return likely(btree_node_locked(path, path->level) &&
1103 		      btree_path_check_pos_in_node(path, path->level, check_pos))
1104 		? path->level
1105 		: __btree_path_up_until_good_node(trans, path, check_pos);
1106 }
1107 
1108 /*
1109  * This is the main state machine for walking down the btree - walks down to a
1110  * specified depth
1111  *
1112  * Returns 0 on success, -EIO on error (error reading in a btree node).
1113  *
1114  * On error, caller (peek_node()/peek_key()) must return NULL; the error is
1115  * stashed in the iterator and returned from bch2_trans_exit().
1116  */
1117 int bch2_btree_path_traverse_one(struct btree_trans *trans,
1118 				 btree_path_idx_t path_idx,
1119 				 unsigned flags,
1120 				 unsigned long trace_ip)
1121 {
1122 	struct btree_path *path = &trans->paths[path_idx];
1123 	unsigned depth_want = path->level;
1124 	int ret = -((int) trans->restarted);
1125 
1126 	if (unlikely(ret))
1127 		goto out;
1128 
1129 	if (unlikely(!trans->srcu_held))
1130 		bch2_trans_srcu_lock(trans);
1131 
1132 	/*
1133 	 * Ensure we obey path->should_be_locked: if it's set, we can't unlock
1134 	 * and re-traverse the path without a transaction restart:
1135 	 */
1136 	if (path->should_be_locked) {
1137 		ret = bch2_btree_path_relock(trans, path, trace_ip);
1138 		goto out;
1139 	}
1140 
1141 	if (path->cached) {
1142 		ret = bch2_btree_path_traverse_cached(trans, path, flags);
1143 		goto out;
1144 	}
1145 
1146 	path = &trans->paths[path_idx];
1147 
1148 	if (unlikely(path->level >= BTREE_MAX_DEPTH))
1149 		goto out;
1150 
1151 	path->level = btree_path_up_until_good_node(trans, path, 0);
1152 
1153 	EBUG_ON(btree_path_node(path, path->level) &&
1154 		!btree_node_locked(path, path->level));
1155 
1156 	/*
1157 	 * Note: path->nodes[path->level] may be temporarily NULL here - that
1158 	 * would indicate to other code that we got to the end of the btree,
1159 	 * here it indicates that relocking the root failed - it's critical that
1160 	 * btree_path_lock_root() comes next and that it can't fail
1161 	 */
1162 	while (path->level > depth_want) {
1163 		ret = btree_path_node(path, path->level)
1164 			? btree_path_down(trans, path, flags, trace_ip)
1165 			: btree_path_lock_root(trans, path, depth_want, trace_ip);
1166 		if (unlikely(ret)) {
1167 			if (ret == 1) {
1168 				/*
1169 				 * No nodes at this level - got to the end of
1170 				 * the btree:
1171 				 */
1172 				ret = 0;
1173 				goto out;
1174 			}
1175 
1176 			__bch2_btree_path_unlock(trans, path);
1177 			path->level = depth_want;
1178 			path->l[path->level].b = ERR_PTR(ret);
1179 			goto out;
1180 		}
1181 	}
1182 
1183 	path->uptodate = BTREE_ITER_UPTODATE;
1184 out:
1185 	if (bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted)
1186 		panic("ret %s (%i) trans->restarted %s (%i)\n",
1187 		      bch2_err_str(ret), ret,
1188 		      bch2_err_str(trans->restarted), trans->restarted);
1189 	bch2_btree_path_verify(trans, path);
1190 	return ret;
1191 }
1192 
1193 static inline void btree_path_copy(struct btree_trans *trans, struct btree_path *dst,
1194 			    struct btree_path *src)
1195 {
1196 	unsigned i, offset = offsetof(struct btree_path, pos);
1197 
1198 	memcpy((void *) dst + offset,
1199 	       (void *) src + offset,
1200 	       sizeof(struct btree_path) - offset);
1201 
1202 	for (i = 0; i < BTREE_MAX_DEPTH; i++) {
1203 		unsigned t = btree_node_locked_type(dst, i);
1204 
1205 		if (t != BTREE_NODE_UNLOCKED)
1206 			six_lock_increment(&dst->l[i].b->c.lock, t);
1207 	}
1208 }
1209 
1210 static btree_path_idx_t btree_path_clone(struct btree_trans *trans, btree_path_idx_t src,
1211 					 bool intent)
1212 {
1213 	btree_path_idx_t new = btree_path_alloc(trans, src);
1214 	btree_path_copy(trans, trans->paths + new, trans->paths + src);
1215 	__btree_path_get(trans->paths + new, intent);
1216 	return new;
1217 }
1218 
1219 __flatten
1220 btree_path_idx_t __bch2_btree_path_make_mut(struct btree_trans *trans,
1221 			btree_path_idx_t path, bool intent, unsigned long ip)
1222 {
1223 	__btree_path_put(trans->paths + path, intent);
1224 	path = btree_path_clone(trans, path, intent);
1225 	trans->paths[path].preserve = false;
1226 	return path;
1227 }
1228 
1229 btree_path_idx_t __must_check
1230 __bch2_btree_path_set_pos(struct btree_trans *trans,
1231 			  btree_path_idx_t path_idx, struct bpos new_pos,
1232 			  bool intent, unsigned long ip)
1233 {
1234 	int cmp = bpos_cmp(new_pos, trans->paths[path_idx].pos);
1235 
1236 	bch2_trans_verify_not_in_restart(trans);
1237 	EBUG_ON(!trans->paths[path_idx].ref);
1238 
1239 	path_idx = bch2_btree_path_make_mut(trans, path_idx, intent, ip);
1240 
1241 	struct btree_path *path = trans->paths + path_idx;
1242 	path->pos		= new_pos;
1243 	trans->paths_sorted	= false;
1244 
1245 	if (unlikely(path->cached)) {
1246 		btree_node_unlock(trans, path, 0);
1247 		path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_up);
1248 		btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1249 		goto out;
1250 	}
1251 
1252 	unsigned level = btree_path_up_until_good_node(trans, path, cmp);
1253 
1254 	if (btree_path_node(path, level)) {
1255 		struct btree_path_level *l = &path->l[level];
1256 
1257 		BUG_ON(!btree_node_locked(path, level));
1258 		/*
1259 		 * We might have to skip over many keys, or just a few: try
1260 		 * advancing the node iterator, and if we have to skip over too
1261 		 * many keys just reinit it (or if we're rewinding, since that
1262 		 * is expensive).
1263 		 */
1264 		if (cmp < 0 ||
1265 		    !btree_path_advance_to_pos(path, l, 8))
1266 			bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
1267 
1268 		/*
1269 		 * Iterators to interior nodes should always be pointed at the first non
1270 		 * whiteout:
1271 		 */
1272 		if (unlikely(level))
1273 			bch2_btree_node_iter_peek(&l->iter, l->b);
1274 	}
1275 
1276 	if (unlikely(level != path->level)) {
1277 		btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1278 		__bch2_btree_path_unlock(trans, path);
1279 	}
1280 out:
1281 	bch2_btree_path_verify(trans, path);
1282 	return path_idx;
1283 }
1284 
1285 /* Btree path: main interface: */
1286 
1287 static struct btree_path *have_path_at_pos(struct btree_trans *trans, struct btree_path *path)
1288 {
1289 	struct btree_path *sib;
1290 
1291 	sib = prev_btree_path(trans, path);
1292 	if (sib && !btree_path_cmp(sib, path))
1293 		return sib;
1294 
1295 	sib = next_btree_path(trans, path);
1296 	if (sib && !btree_path_cmp(sib, path))
1297 		return sib;
1298 
1299 	return NULL;
1300 }
1301 
1302 static struct btree_path *have_node_at_pos(struct btree_trans *trans, struct btree_path *path)
1303 {
1304 	struct btree_path *sib;
1305 
1306 	sib = prev_btree_path(trans, path);
1307 	if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1308 		return sib;
1309 
1310 	sib = next_btree_path(trans, path);
1311 	if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1312 		return sib;
1313 
1314 	return NULL;
1315 }
1316 
1317 static inline void __bch2_path_free(struct btree_trans *trans, btree_path_idx_t path)
1318 {
1319 	__bch2_btree_path_unlock(trans, trans->paths + path);
1320 	btree_path_list_remove(trans, trans->paths + path);
1321 	__clear_bit(path, trans->paths_allocated);
1322 }
1323 
1324 void bch2_path_put(struct btree_trans *trans, btree_path_idx_t path_idx, bool intent)
1325 {
1326 	struct btree_path *path = trans->paths + path_idx, *dup;
1327 
1328 	if (!__btree_path_put(path, intent))
1329 		return;
1330 
1331 	dup = path->preserve
1332 		? have_path_at_pos(trans, path)
1333 		: have_node_at_pos(trans, path);
1334 
1335 	if (!dup && !(!path->preserve && !is_btree_node(path, path->level)))
1336 		return;
1337 
1338 	if (path->should_be_locked &&
1339 	    !trans->restarted &&
1340 	    (!dup || !bch2_btree_path_relock_norestart(trans, dup)))
1341 		return;
1342 
1343 	if (dup) {
1344 		dup->preserve		|= path->preserve;
1345 		dup->should_be_locked	|= path->should_be_locked;
1346 	}
1347 
1348 	__bch2_path_free(trans, path_idx);
1349 }
1350 
1351 static void bch2_path_put_nokeep(struct btree_trans *trans, btree_path_idx_t path,
1352 				 bool intent)
1353 {
1354 	if (!__btree_path_put(trans->paths + path, intent))
1355 		return;
1356 
1357 	__bch2_path_free(trans, path);
1358 }
1359 
1360 void __noreturn bch2_trans_restart_error(struct btree_trans *trans, u32 restart_count)
1361 {
1362 	panic("trans->restart_count %u, should be %u, last restarted by %pS\n",
1363 	      trans->restart_count, restart_count,
1364 	      (void *) trans->last_begin_ip);
1365 }
1366 
1367 void __noreturn bch2_trans_in_restart_error(struct btree_trans *trans)
1368 {
1369 	panic("in transaction restart: %s, last restarted by %pS\n",
1370 	      bch2_err_str(trans->restarted),
1371 	      (void *) trans->last_restarted_ip);
1372 }
1373 
1374 noinline __cold
1375 void bch2_trans_updates_to_text(struct printbuf *buf, struct btree_trans *trans)
1376 {
1377 	prt_printf(buf, "transaction updates for %s journal seq %llu",
1378 	       trans->fn, trans->journal_res.seq);
1379 	prt_newline(buf);
1380 	printbuf_indent_add(buf, 2);
1381 
1382 	trans_for_each_update(trans, i) {
1383 		struct bkey_s_c old = { &i->old_k, i->old_v };
1384 
1385 		prt_printf(buf, "update: btree=%s cached=%u %pS",
1386 		       bch2_btree_id_str(i->btree_id),
1387 		       i->cached,
1388 		       (void *) i->ip_allocated);
1389 		prt_newline(buf);
1390 
1391 		prt_printf(buf, "  old ");
1392 		bch2_bkey_val_to_text(buf, trans->c, old);
1393 		prt_newline(buf);
1394 
1395 		prt_printf(buf, "  new ");
1396 		bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(i->k));
1397 		prt_newline(buf);
1398 	}
1399 
1400 	for (struct jset_entry *e = trans->journal_entries;
1401 	     e != btree_trans_journal_entries_top(trans);
1402 	     e = vstruct_next(e))
1403 		bch2_journal_entry_to_text(buf, trans->c, e);
1404 
1405 	printbuf_indent_sub(buf, 2);
1406 }
1407 
1408 noinline __cold
1409 void bch2_dump_trans_updates(struct btree_trans *trans)
1410 {
1411 	struct printbuf buf = PRINTBUF;
1412 
1413 	bch2_trans_updates_to_text(&buf, trans);
1414 	bch2_print_string_as_lines(KERN_ERR, buf.buf);
1415 	printbuf_exit(&buf);
1416 }
1417 
1418 static void bch2_btree_path_to_text(struct printbuf *out, struct btree_trans *trans, btree_path_idx_t path_idx)
1419 {
1420 	struct btree_path *path = trans->paths + path_idx;
1421 
1422 	prt_printf(out, "path: idx %2u ref %u:%u %c %c btree=%s l=%u pos ",
1423 		   path_idx, path->ref, path->intent_ref,
1424 		   path->preserve ? 'P' : ' ',
1425 		   path->should_be_locked ? 'S' : ' ',
1426 		   bch2_btree_id_str(path->btree_id),
1427 		   path->level);
1428 	bch2_bpos_to_text(out, path->pos);
1429 
1430 	prt_printf(out, " locks %u", path->nodes_locked);
1431 #ifdef TRACK_PATH_ALLOCATED
1432 	prt_printf(out, " %pS", (void *) path->ip_allocated);
1433 #endif
1434 	prt_newline(out);
1435 }
1436 
1437 static noinline __cold
1438 void __bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans,
1439 				bool nosort)
1440 {
1441 	struct trans_for_each_path_inorder_iter iter;
1442 
1443 	if (!nosort)
1444 		btree_trans_sort_paths(trans);
1445 
1446 	trans_for_each_path_idx_inorder(trans, iter)
1447 		bch2_btree_path_to_text(out, trans, iter.path_idx);
1448 }
1449 
1450 noinline __cold
1451 void bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans)
1452 {
1453 	__bch2_trans_paths_to_text(out, trans, false);
1454 }
1455 
1456 static noinline __cold
1457 void __bch2_dump_trans_paths_updates(struct btree_trans *trans, bool nosort)
1458 {
1459 	struct printbuf buf = PRINTBUF;
1460 
1461 	__bch2_trans_paths_to_text(&buf, trans, nosort);
1462 	bch2_trans_updates_to_text(&buf, trans);
1463 
1464 	bch2_print_string_as_lines(KERN_ERR, buf.buf);
1465 	printbuf_exit(&buf);
1466 }
1467 
1468 noinline __cold
1469 void bch2_dump_trans_paths_updates(struct btree_trans *trans)
1470 {
1471 	__bch2_dump_trans_paths_updates(trans, false);
1472 }
1473 
1474 noinline __cold
1475 static void bch2_trans_update_max_paths(struct btree_trans *trans)
1476 {
1477 	struct btree_transaction_stats *s = btree_trans_stats(trans);
1478 	struct printbuf buf = PRINTBUF;
1479 	size_t nr = bitmap_weight(trans->paths_allocated, trans->nr_paths);
1480 
1481 	bch2_trans_paths_to_text(&buf, trans);
1482 
1483 	if (!buf.allocation_failure) {
1484 		mutex_lock(&s->lock);
1485 		if (nr > s->nr_max_paths) {
1486 			s->nr_max_paths = nr;
1487 			swap(s->max_paths_text, buf.buf);
1488 		}
1489 		mutex_unlock(&s->lock);
1490 	}
1491 
1492 	printbuf_exit(&buf);
1493 
1494 	trans->nr_paths_max = nr;
1495 }
1496 
1497 noinline __cold
1498 int __bch2_btree_trans_too_many_iters(struct btree_trans *trans)
1499 {
1500 	if (trace_trans_restart_too_many_iters_enabled()) {
1501 		struct printbuf buf = PRINTBUF;
1502 
1503 		bch2_trans_paths_to_text(&buf, trans);
1504 		trace_trans_restart_too_many_iters(trans, _THIS_IP_, buf.buf);
1505 		printbuf_exit(&buf);
1506 	}
1507 
1508 	count_event(trans->c, trans_restart_too_many_iters);
1509 
1510 	return btree_trans_restart(trans, BCH_ERR_transaction_restart_too_many_iters);
1511 }
1512 
1513 static noinline void btree_path_overflow(struct btree_trans *trans)
1514 {
1515 	bch2_dump_trans_paths_updates(trans);
1516 	bch_err(trans->c, "trans path overflow");
1517 }
1518 
1519 static noinline void btree_paths_realloc(struct btree_trans *trans)
1520 {
1521 	unsigned nr = trans->nr_paths * 2;
1522 
1523 	void *p = kzalloc(BITS_TO_LONGS(nr) * sizeof(unsigned long) +
1524 			  sizeof(struct btree_trans_paths) +
1525 			  nr * sizeof(struct btree_path) +
1526 			  nr * sizeof(btree_path_idx_t) + 8 +
1527 			  nr * sizeof(struct btree_insert_entry), GFP_KERNEL|__GFP_NOFAIL);
1528 
1529 	unsigned long *paths_allocated = p;
1530 	memcpy(paths_allocated, trans->paths_allocated, BITS_TO_LONGS(trans->nr_paths) * sizeof(unsigned long));
1531 	p += BITS_TO_LONGS(nr) * sizeof(unsigned long);
1532 
1533 	p += sizeof(struct btree_trans_paths);
1534 	struct btree_path *paths = p;
1535 	*trans_paths_nr(paths) = nr;
1536 	memcpy(paths, trans->paths, trans->nr_paths * sizeof(struct btree_path));
1537 	p += nr * sizeof(struct btree_path);
1538 
1539 	btree_path_idx_t *sorted = p;
1540 	memcpy(sorted, trans->sorted, trans->nr_sorted * sizeof(btree_path_idx_t));
1541 	p += nr * sizeof(btree_path_idx_t) + 8;
1542 
1543 	struct btree_insert_entry *updates = p;
1544 	memcpy(updates, trans->updates, trans->nr_paths * sizeof(struct btree_insert_entry));
1545 
1546 	unsigned long *old = trans->paths_allocated;
1547 
1548 	rcu_assign_pointer(trans->paths_allocated,	paths_allocated);
1549 	rcu_assign_pointer(trans->paths,		paths);
1550 	rcu_assign_pointer(trans->sorted,		sorted);
1551 	rcu_assign_pointer(trans->updates,		updates);
1552 
1553 	trans->nr_paths		= nr;
1554 
1555 	if (old != trans->_paths_allocated)
1556 		kfree_rcu_mightsleep(old);
1557 }
1558 
1559 static inline btree_path_idx_t btree_path_alloc(struct btree_trans *trans,
1560 						btree_path_idx_t pos)
1561 {
1562 	btree_path_idx_t idx = find_first_zero_bit(trans->paths_allocated, trans->nr_paths);
1563 
1564 	if (unlikely(idx == trans->nr_paths)) {
1565 		if (trans->nr_paths == BTREE_ITER_MAX) {
1566 			btree_path_overflow(trans);
1567 			return 0;
1568 		}
1569 
1570 		btree_paths_realloc(trans);
1571 	}
1572 
1573 	/*
1574 	 * Do this before marking the new path as allocated, since it won't be
1575 	 * initialized yet:
1576 	 */
1577 	if (unlikely(idx > trans->nr_paths_max))
1578 		bch2_trans_update_max_paths(trans);
1579 
1580 	__set_bit(idx, trans->paths_allocated);
1581 
1582 	struct btree_path *path = &trans->paths[idx];
1583 	path->ref		= 0;
1584 	path->intent_ref	= 0;
1585 	path->nodes_locked	= 0;
1586 
1587 	btree_path_list_add(trans, pos, idx);
1588 	trans->paths_sorted = false;
1589 	return idx;
1590 }
1591 
1592 btree_path_idx_t bch2_path_get(struct btree_trans *trans,
1593 			     enum btree_id btree_id, struct bpos pos,
1594 			     unsigned locks_want, unsigned level,
1595 			     unsigned flags, unsigned long ip)
1596 {
1597 	struct btree_path *path;
1598 	bool cached = flags & BTREE_ITER_CACHED;
1599 	bool intent = flags & BTREE_ITER_INTENT;
1600 	struct trans_for_each_path_inorder_iter iter;
1601 	btree_path_idx_t path_pos = 0, path_idx;
1602 
1603 	bch2_trans_verify_not_in_restart(trans);
1604 	bch2_trans_verify_locks(trans);
1605 
1606 	btree_trans_sort_paths(trans);
1607 
1608 	trans_for_each_path_inorder(trans, path, iter) {
1609 		if (__btree_path_cmp(path,
1610 				     btree_id,
1611 				     cached,
1612 				     pos,
1613 				     level) > 0)
1614 			break;
1615 
1616 		path_pos = iter.path_idx;
1617 	}
1618 
1619 	if (path_pos &&
1620 	    trans->paths[path_pos].cached	== cached &&
1621 	    trans->paths[path_pos].btree_id	== btree_id &&
1622 	    trans->paths[path_pos].level	== level) {
1623 		__btree_path_get(trans->paths + path_pos, intent);
1624 		path_idx = bch2_btree_path_set_pos(trans, path_pos, pos, intent, ip);
1625 		path = trans->paths + path_idx;
1626 	} else {
1627 		path_idx = btree_path_alloc(trans, path_pos);
1628 		path = trans->paths + path_idx;
1629 
1630 		__btree_path_get(path, intent);
1631 		path->pos			= pos;
1632 		path->btree_id			= btree_id;
1633 		path->cached			= cached;
1634 		path->uptodate			= BTREE_ITER_NEED_TRAVERSE;
1635 		path->should_be_locked		= false;
1636 		path->level			= level;
1637 		path->locks_want		= locks_want;
1638 		path->nodes_locked		= 0;
1639 		for (unsigned i = 0; i < ARRAY_SIZE(path->l); i++)
1640 			path->l[i].b		= ERR_PTR(-BCH_ERR_no_btree_node_init);
1641 #ifdef TRACK_PATH_ALLOCATED
1642 		path->ip_allocated		= ip;
1643 #endif
1644 		trans->paths_sorted		= false;
1645 	}
1646 
1647 	if (!(flags & BTREE_ITER_NOPRESERVE))
1648 		path->preserve = true;
1649 
1650 	if (path->intent_ref)
1651 		locks_want = max(locks_want, level + 1);
1652 
1653 	/*
1654 	 * If the path has locks_want greater than requested, we don't downgrade
1655 	 * it here - on transaction restart because btree node split needs to
1656 	 * upgrade locks, we might be putting/getting the iterator again.
1657 	 * Downgrading iterators only happens via bch2_trans_downgrade(), after
1658 	 * a successful transaction commit.
1659 	 */
1660 
1661 	locks_want = min(locks_want, BTREE_MAX_DEPTH);
1662 	if (locks_want > path->locks_want)
1663 		bch2_btree_path_upgrade_noupgrade_sibs(trans, path, locks_want, NULL);
1664 
1665 	return path_idx;
1666 }
1667 
1668 struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u)
1669 {
1670 
1671 	struct btree_path_level *l = path_l(path);
1672 	struct bkey_packed *_k;
1673 	struct bkey_s_c k;
1674 
1675 	if (unlikely(!l->b))
1676 		return bkey_s_c_null;
1677 
1678 	EBUG_ON(path->uptodate != BTREE_ITER_UPTODATE);
1679 	EBUG_ON(!btree_node_locked(path, path->level));
1680 
1681 	if (!path->cached) {
1682 		_k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1683 		k = _k ? bkey_disassemble(l->b, _k, u) : bkey_s_c_null;
1684 
1685 		EBUG_ON(k.k && bkey_deleted(k.k) && bpos_eq(k.k->p, path->pos));
1686 
1687 		if (!k.k || !bpos_eq(path->pos, k.k->p))
1688 			goto hole;
1689 	} else {
1690 		struct bkey_cached *ck = (void *) path->l[0].b;
1691 
1692 		EBUG_ON(ck &&
1693 			(path->btree_id != ck->key.btree_id ||
1694 			 !bkey_eq(path->pos, ck->key.pos)));
1695 		if (!ck || !ck->valid)
1696 			return bkey_s_c_null;
1697 
1698 		*u = ck->k->k;
1699 		k = bkey_i_to_s_c(ck->k);
1700 	}
1701 
1702 	return k;
1703 hole:
1704 	bkey_init(u);
1705 	u->p = path->pos;
1706 	return (struct bkey_s_c) { u, NULL };
1707 }
1708 
1709 /* Btree iterators: */
1710 
1711 int __must_check
1712 __bch2_btree_iter_traverse(struct btree_iter *iter)
1713 {
1714 	return bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1715 }
1716 
1717 int __must_check
1718 bch2_btree_iter_traverse(struct btree_iter *iter)
1719 {
1720 	struct btree_trans *trans = iter->trans;
1721 	int ret;
1722 
1723 	iter->path = bch2_btree_path_set_pos(trans, iter->path,
1724 					btree_iter_search_key(iter),
1725 					iter->flags & BTREE_ITER_INTENT,
1726 					btree_iter_ip_allocated(iter));
1727 
1728 	ret = bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1729 	if (ret)
1730 		return ret;
1731 
1732 	btree_path_set_should_be_locked(trans->paths + iter->path);
1733 	return 0;
1734 }
1735 
1736 /* Iterate across nodes (leaf and interior nodes) */
1737 
1738 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
1739 {
1740 	struct btree_trans *trans = iter->trans;
1741 	struct btree *b = NULL;
1742 	int ret;
1743 
1744 	EBUG_ON(trans->paths[iter->path].cached);
1745 	bch2_btree_iter_verify(iter);
1746 
1747 	ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1748 	if (ret)
1749 		goto err;
1750 
1751 	struct btree_path *path = btree_iter_path(trans, iter);
1752 	b = btree_path_node(path, path->level);
1753 	if (!b)
1754 		goto out;
1755 
1756 	BUG_ON(bpos_lt(b->key.k.p, iter->pos));
1757 
1758 	bkey_init(&iter->k);
1759 	iter->k.p = iter->pos = b->key.k.p;
1760 
1761 	iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1762 					iter->flags & BTREE_ITER_INTENT,
1763 					btree_iter_ip_allocated(iter));
1764 	btree_path_set_should_be_locked(btree_iter_path(trans, iter));
1765 out:
1766 	bch2_btree_iter_verify_entry_exit(iter);
1767 	bch2_btree_iter_verify(iter);
1768 
1769 	return b;
1770 err:
1771 	b = ERR_PTR(ret);
1772 	goto out;
1773 }
1774 
1775 struct btree *bch2_btree_iter_peek_node_and_restart(struct btree_iter *iter)
1776 {
1777 	struct btree *b;
1778 
1779 	while (b = bch2_btree_iter_peek_node(iter),
1780 	       bch2_err_matches(PTR_ERR_OR_ZERO(b), BCH_ERR_transaction_restart))
1781 		bch2_trans_begin(iter->trans);
1782 
1783 	return b;
1784 }
1785 
1786 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
1787 {
1788 	struct btree_trans *trans = iter->trans;
1789 	struct btree *b = NULL;
1790 	int ret;
1791 
1792 	EBUG_ON(trans->paths[iter->path].cached);
1793 	bch2_trans_verify_not_in_restart(trans);
1794 	bch2_btree_iter_verify(iter);
1795 
1796 	struct btree_path *path = btree_iter_path(trans, iter);
1797 
1798 	/* already at end? */
1799 	if (!btree_path_node(path, path->level))
1800 		return NULL;
1801 
1802 	/* got to end? */
1803 	if (!btree_path_node(path, path->level + 1)) {
1804 		btree_path_set_level_up(trans, path);
1805 		return NULL;
1806 	}
1807 
1808 	if (!bch2_btree_node_relock(trans, path, path->level + 1)) {
1809 		__bch2_btree_path_unlock(trans, path);
1810 		path->l[path->level].b		= ERR_PTR(-BCH_ERR_no_btree_node_relock);
1811 		path->l[path->level + 1].b	= ERR_PTR(-BCH_ERR_no_btree_node_relock);
1812 		btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1813 		trace_and_count(trans->c, trans_restart_relock_next_node, trans, _THIS_IP_, path);
1814 		ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
1815 		goto err;
1816 	}
1817 
1818 	b = btree_path_node(path, path->level + 1);
1819 
1820 	if (bpos_eq(iter->pos, b->key.k.p)) {
1821 		__btree_path_set_level_up(trans, path, path->level++);
1822 	} else {
1823 		/*
1824 		 * Haven't gotten to the end of the parent node: go back down to
1825 		 * the next child node
1826 		 */
1827 		iter->path = bch2_btree_path_set_pos(trans, iter->path,
1828 					bpos_successor(iter->pos),
1829 					iter->flags & BTREE_ITER_INTENT,
1830 					btree_iter_ip_allocated(iter));
1831 
1832 		path = btree_iter_path(trans, iter);
1833 		btree_path_set_level_down(trans, path, iter->min_depth);
1834 
1835 		ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1836 		if (ret)
1837 			goto err;
1838 
1839 		path = btree_iter_path(trans, iter);
1840 		b = path->l[path->level].b;
1841 	}
1842 
1843 	bkey_init(&iter->k);
1844 	iter->k.p = iter->pos = b->key.k.p;
1845 
1846 	iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1847 					iter->flags & BTREE_ITER_INTENT,
1848 					btree_iter_ip_allocated(iter));
1849 	btree_path_set_should_be_locked(btree_iter_path(trans, iter));
1850 	EBUG_ON(btree_iter_path(trans, iter)->uptodate);
1851 out:
1852 	bch2_btree_iter_verify_entry_exit(iter);
1853 	bch2_btree_iter_verify(iter);
1854 
1855 	return b;
1856 err:
1857 	b = ERR_PTR(ret);
1858 	goto out;
1859 }
1860 
1861 /* Iterate across keys (in leaf nodes only) */
1862 
1863 inline bool bch2_btree_iter_advance(struct btree_iter *iter)
1864 {
1865 	struct bpos pos = iter->k.p;
1866 	bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1867 		     ? bpos_eq(pos, SPOS_MAX)
1868 		     : bkey_eq(pos, SPOS_MAX));
1869 
1870 	if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1871 		pos = bkey_successor(iter, pos);
1872 	bch2_btree_iter_set_pos(iter, pos);
1873 	return ret;
1874 }
1875 
1876 inline bool bch2_btree_iter_rewind(struct btree_iter *iter)
1877 {
1878 	struct bpos pos = bkey_start_pos(&iter->k);
1879 	bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1880 		     ? bpos_eq(pos, POS_MIN)
1881 		     : bkey_eq(pos, POS_MIN));
1882 
1883 	if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1884 		pos = bkey_predecessor(iter, pos);
1885 	bch2_btree_iter_set_pos(iter, pos);
1886 	return ret;
1887 }
1888 
1889 static noinline
1890 void bch2_btree_trans_peek_prev_updates(struct btree_trans *trans, struct btree_iter *iter,
1891 					struct bkey_s_c *k)
1892 {
1893 	struct bpos end = path_l(btree_iter_path(trans, iter))->b->data->min_key;
1894 
1895 	trans_for_each_update(trans, i)
1896 		if (!i->key_cache_already_flushed &&
1897 		    i->btree_id == iter->btree_id &&
1898 		    bpos_le(i->k->k.p, iter->pos) &&
1899 		    bpos_ge(i->k->k.p, k->k ? k->k->p : end)) {
1900 			iter->k = i->k->k;
1901 			*k = bkey_i_to_s_c(i->k);
1902 		}
1903 }
1904 
1905 static noinline
1906 void bch2_btree_trans_peek_updates(struct btree_trans *trans, struct btree_iter *iter,
1907 				   struct bkey_s_c *k)
1908 {
1909 	struct btree_path *path = btree_iter_path(trans, iter);
1910 	struct bpos end = path_l(path)->b->key.k.p;
1911 
1912 	trans_for_each_update(trans, i)
1913 		if (!i->key_cache_already_flushed &&
1914 		    i->btree_id == iter->btree_id &&
1915 		    bpos_ge(i->k->k.p, path->pos) &&
1916 		    bpos_le(i->k->k.p, k->k ? k->k->p : end)) {
1917 			iter->k = i->k->k;
1918 			*k = bkey_i_to_s_c(i->k);
1919 		}
1920 }
1921 
1922 static noinline
1923 void bch2_btree_trans_peek_slot_updates(struct btree_trans *trans, struct btree_iter *iter,
1924 					struct bkey_s_c *k)
1925 {
1926 	trans_for_each_update(trans, i)
1927 		if (!i->key_cache_already_flushed &&
1928 		    i->btree_id == iter->btree_id &&
1929 		    bpos_eq(i->k->k.p, iter->pos)) {
1930 			iter->k = i->k->k;
1931 			*k = bkey_i_to_s_c(i->k);
1932 		}
1933 }
1934 
1935 static struct bkey_i *bch2_btree_journal_peek(struct btree_trans *trans,
1936 					      struct btree_iter *iter,
1937 					      struct bpos end_pos)
1938 {
1939 	struct btree_path *path = btree_iter_path(trans, iter);
1940 
1941 	return bch2_journal_keys_peek_upto(trans->c, iter->btree_id,
1942 					   path->level,
1943 					   path->pos,
1944 					   end_pos,
1945 					   &iter->journal_idx);
1946 }
1947 
1948 static noinline
1949 struct bkey_s_c btree_trans_peek_slot_journal(struct btree_trans *trans,
1950 					      struct btree_iter *iter)
1951 {
1952 	struct btree_path *path = btree_iter_path(trans, iter);
1953 	struct bkey_i *k = bch2_btree_journal_peek(trans, iter, path->pos);
1954 
1955 	if (k) {
1956 		iter->k = k->k;
1957 		return bkey_i_to_s_c(k);
1958 	} else {
1959 		return bkey_s_c_null;
1960 	}
1961 }
1962 
1963 static noinline
1964 struct bkey_s_c btree_trans_peek_journal(struct btree_trans *trans,
1965 					 struct btree_iter *iter,
1966 					 struct bkey_s_c k)
1967 {
1968 	struct btree_path *path = btree_iter_path(trans, iter);
1969 	struct bkey_i *next_journal =
1970 		bch2_btree_journal_peek(trans, iter,
1971 				k.k ? k.k->p : path_l(path)->b->key.k.p);
1972 
1973 	if (next_journal) {
1974 		iter->k = next_journal->k;
1975 		k = bkey_i_to_s_c(next_journal);
1976 	}
1977 
1978 	return k;
1979 }
1980 
1981 /*
1982  * Checks btree key cache for key at iter->pos and returns it if present, or
1983  * bkey_s_c_null:
1984  */
1985 static noinline
1986 struct bkey_s_c btree_trans_peek_key_cache(struct btree_iter *iter, struct bpos pos)
1987 {
1988 	struct btree_trans *trans = iter->trans;
1989 	struct bch_fs *c = trans->c;
1990 	struct bkey u;
1991 	struct bkey_s_c k;
1992 	int ret;
1993 
1994 	if ((iter->flags & BTREE_ITER_KEY_CACHE_FILL) &&
1995 	    bpos_eq(iter->pos, pos))
1996 		return bkey_s_c_null;
1997 
1998 	if (!bch2_btree_key_cache_find(c, iter->btree_id, pos))
1999 		return bkey_s_c_null;
2000 
2001 	if (!iter->key_cache_path)
2002 		iter->key_cache_path = bch2_path_get(trans, iter->btree_id, pos,
2003 						     iter->flags & BTREE_ITER_INTENT, 0,
2004 						     iter->flags|BTREE_ITER_CACHED|
2005 						     BTREE_ITER_CACHED_NOFILL,
2006 						     _THIS_IP_);
2007 
2008 	iter->key_cache_path = bch2_btree_path_set_pos(trans, iter->key_cache_path, pos,
2009 					iter->flags & BTREE_ITER_INTENT,
2010 					btree_iter_ip_allocated(iter));
2011 
2012 	ret =   bch2_btree_path_traverse(trans, iter->key_cache_path,
2013 					 iter->flags|BTREE_ITER_CACHED) ?:
2014 		bch2_btree_path_relock(trans, btree_iter_path(trans, iter), _THIS_IP_);
2015 	if (unlikely(ret))
2016 		return bkey_s_c_err(ret);
2017 
2018 	btree_path_set_should_be_locked(trans->paths + iter->key_cache_path);
2019 
2020 	k = bch2_btree_path_peek_slot(trans->paths + iter->key_cache_path, &u);
2021 	if (k.k && !bkey_err(k)) {
2022 		iter->k = u;
2023 		k.k = &iter->k;
2024 	}
2025 	return k;
2026 }
2027 
2028 static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bpos search_key)
2029 {
2030 	struct btree_trans *trans = iter->trans;
2031 	struct bkey_s_c k, k2;
2032 	int ret;
2033 
2034 	EBUG_ON(btree_iter_path(trans, iter)->cached);
2035 	bch2_btree_iter_verify(iter);
2036 
2037 	while (1) {
2038 		struct btree_path_level *l;
2039 
2040 		iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2041 					iter->flags & BTREE_ITER_INTENT,
2042 					btree_iter_ip_allocated(iter));
2043 
2044 		ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2045 		if (unlikely(ret)) {
2046 			/* ensure that iter->k is consistent with iter->pos: */
2047 			bch2_btree_iter_set_pos(iter, iter->pos);
2048 			k = bkey_s_c_err(ret);
2049 			goto out;
2050 		}
2051 
2052 		struct btree_path *path = btree_iter_path(trans, iter);
2053 		l = path_l(path);
2054 
2055 		if (unlikely(!l->b)) {
2056 			/* No btree nodes at requested level: */
2057 			bch2_btree_iter_set_pos(iter, SPOS_MAX);
2058 			k = bkey_s_c_null;
2059 			goto out;
2060 		}
2061 
2062 		btree_path_set_should_be_locked(path);
2063 
2064 		k = btree_path_level_peek_all(trans->c, l, &iter->k);
2065 
2066 		if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2067 		    k.k &&
2068 		    (k2 = btree_trans_peek_key_cache(iter, k.k->p)).k) {
2069 			k = k2;
2070 			ret = bkey_err(k);
2071 			if (ret) {
2072 				bch2_btree_iter_set_pos(iter, iter->pos);
2073 				goto out;
2074 			}
2075 		}
2076 
2077 		if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL))
2078 			k = btree_trans_peek_journal(trans, iter, k);
2079 
2080 		if (unlikely((iter->flags & BTREE_ITER_WITH_UPDATES) &&
2081 			     trans->nr_updates))
2082 			bch2_btree_trans_peek_updates(trans, iter, &k);
2083 
2084 		if (k.k && bkey_deleted(k.k)) {
2085 			/*
2086 			 * If we've got a whiteout, and it's after the search
2087 			 * key, advance the search key to the whiteout instead
2088 			 * of just after the whiteout - it might be a btree
2089 			 * whiteout, with a real key at the same position, since
2090 			 * in the btree deleted keys sort before non deleted.
2091 			 */
2092 			search_key = !bpos_eq(search_key, k.k->p)
2093 				? k.k->p
2094 				: bpos_successor(k.k->p);
2095 			continue;
2096 		}
2097 
2098 		if (likely(k.k)) {
2099 			break;
2100 		} else if (likely(!bpos_eq(l->b->key.k.p, SPOS_MAX))) {
2101 			/* Advance to next leaf node: */
2102 			search_key = bpos_successor(l->b->key.k.p);
2103 		} else {
2104 			/* End of btree: */
2105 			bch2_btree_iter_set_pos(iter, SPOS_MAX);
2106 			k = bkey_s_c_null;
2107 			goto out;
2108 		}
2109 	}
2110 out:
2111 	bch2_btree_iter_verify(iter);
2112 
2113 	return k;
2114 }
2115 
2116 /**
2117  * bch2_btree_iter_peek_upto() - returns first key greater than or equal to
2118  * iterator's current position
2119  * @iter:	iterator to peek from
2120  * @end:	search limit: returns keys less than or equal to @end
2121  *
2122  * Returns:	key if found, or an error extractable with bkey_err().
2123  */
2124 struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *iter, struct bpos end)
2125 {
2126 	struct btree_trans *trans = iter->trans;
2127 	struct bpos search_key = btree_iter_search_key(iter);
2128 	struct bkey_s_c k;
2129 	struct bpos iter_pos;
2130 	int ret;
2131 
2132 	EBUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) && bkey_eq(end, POS_MAX));
2133 
2134 	if (iter->update_path) {
2135 		bch2_path_put_nokeep(trans, iter->update_path,
2136 				     iter->flags & BTREE_ITER_INTENT);
2137 		iter->update_path = 0;
2138 	}
2139 
2140 	bch2_btree_iter_verify_entry_exit(iter);
2141 
2142 	while (1) {
2143 		k = __bch2_btree_iter_peek(iter, search_key);
2144 		if (unlikely(!k.k))
2145 			goto end;
2146 		if (unlikely(bkey_err(k)))
2147 			goto out_no_locked;
2148 
2149 		/*
2150 		 * We need to check against @end before FILTER_SNAPSHOTS because
2151 		 * if we get to a different inode that requested we might be
2152 		 * seeing keys for a different snapshot tree that will all be
2153 		 * filtered out.
2154 		 *
2155 		 * But we can't do the full check here, because bkey_start_pos()
2156 		 * isn't monotonically increasing before FILTER_SNAPSHOTS, and
2157 		 * that's what we check against in extents mode:
2158 		 */
2159 		if (k.k->p.inode > end.inode)
2160 			goto end;
2161 
2162 		if (iter->update_path &&
2163 		    !bkey_eq(trans->paths[iter->update_path].pos, k.k->p)) {
2164 			bch2_path_put_nokeep(trans, iter->update_path,
2165 					     iter->flags & BTREE_ITER_INTENT);
2166 			iter->update_path = 0;
2167 		}
2168 
2169 		if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2170 		    (iter->flags & BTREE_ITER_INTENT) &&
2171 		    !(iter->flags & BTREE_ITER_IS_EXTENTS) &&
2172 		    !iter->update_path) {
2173 			struct bpos pos = k.k->p;
2174 
2175 			if (pos.snapshot < iter->snapshot) {
2176 				search_key = bpos_successor(k.k->p);
2177 				continue;
2178 			}
2179 
2180 			pos.snapshot = iter->snapshot;
2181 
2182 			/*
2183 			 * advance, same as on exit for iter->path, but only up
2184 			 * to snapshot
2185 			 */
2186 			__btree_path_get(trans->paths + iter->path, iter->flags & BTREE_ITER_INTENT);
2187 			iter->update_path = iter->path;
2188 
2189 			iter->update_path = bch2_btree_path_set_pos(trans,
2190 						iter->update_path, pos,
2191 						iter->flags & BTREE_ITER_INTENT,
2192 						_THIS_IP_);
2193 			ret = bch2_btree_path_traverse(trans, iter->update_path, iter->flags);
2194 			if (unlikely(ret)) {
2195 				k = bkey_s_c_err(ret);
2196 				goto out_no_locked;
2197 			}
2198 		}
2199 
2200 		/*
2201 		 * We can never have a key in a leaf node at POS_MAX, so
2202 		 * we don't have to check these successor() calls:
2203 		 */
2204 		if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2205 		    !bch2_snapshot_is_ancestor(trans->c,
2206 					       iter->snapshot,
2207 					       k.k->p.snapshot)) {
2208 			search_key = bpos_successor(k.k->p);
2209 			continue;
2210 		}
2211 
2212 		if (bkey_whiteout(k.k) &&
2213 		    !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2214 			search_key = bkey_successor(iter, k.k->p);
2215 			continue;
2216 		}
2217 
2218 		/*
2219 		 * iter->pos should be mononotically increasing, and always be
2220 		 * equal to the key we just returned - except extents can
2221 		 * straddle iter->pos:
2222 		 */
2223 		if (!(iter->flags & BTREE_ITER_IS_EXTENTS))
2224 			iter_pos = k.k->p;
2225 		else
2226 			iter_pos = bkey_max(iter->pos, bkey_start_pos(k.k));
2227 
2228 		if (unlikely(!(iter->flags & BTREE_ITER_IS_EXTENTS)
2229 			     ? bkey_gt(iter_pos, end)
2230 			     : bkey_ge(iter_pos, end)))
2231 			goto end;
2232 
2233 		break;
2234 	}
2235 
2236 	iter->pos = iter_pos;
2237 
2238 	iter->path = bch2_btree_path_set_pos(trans, iter->path, k.k->p,
2239 				iter->flags & BTREE_ITER_INTENT,
2240 				btree_iter_ip_allocated(iter));
2241 
2242 	btree_path_set_should_be_locked(btree_iter_path(trans, iter));
2243 out_no_locked:
2244 	if (iter->update_path) {
2245 		ret = bch2_btree_path_relock(trans, trans->paths + iter->update_path, _THIS_IP_);
2246 		if (unlikely(ret))
2247 			k = bkey_s_c_err(ret);
2248 		else
2249 			btree_path_set_should_be_locked(trans->paths + iter->update_path);
2250 	}
2251 
2252 	if (!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS))
2253 		iter->pos.snapshot = iter->snapshot;
2254 
2255 	ret = bch2_btree_iter_verify_ret(iter, k);
2256 	if (unlikely(ret)) {
2257 		bch2_btree_iter_set_pos(iter, iter->pos);
2258 		k = bkey_s_c_err(ret);
2259 	}
2260 
2261 	bch2_btree_iter_verify_entry_exit(iter);
2262 
2263 	return k;
2264 end:
2265 	bch2_btree_iter_set_pos(iter, end);
2266 	k = bkey_s_c_null;
2267 	goto out_no_locked;
2268 }
2269 
2270 /**
2271  * bch2_btree_iter_next() - returns first key greater than iterator's current
2272  * position
2273  * @iter:	iterator to peek from
2274  *
2275  * Returns:	key if found, or an error extractable with bkey_err().
2276  */
2277 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
2278 {
2279 	if (!bch2_btree_iter_advance(iter))
2280 		return bkey_s_c_null;
2281 
2282 	return bch2_btree_iter_peek(iter);
2283 }
2284 
2285 /**
2286  * bch2_btree_iter_peek_prev() - returns first key less than or equal to
2287  * iterator's current position
2288  * @iter:	iterator to peek from
2289  *
2290  * Returns:	key if found, or an error extractable with bkey_err().
2291  */
2292 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
2293 {
2294 	struct btree_trans *trans = iter->trans;
2295 	struct bpos search_key = iter->pos;
2296 	struct bkey_s_c k;
2297 	struct bkey saved_k;
2298 	const struct bch_val *saved_v;
2299 	btree_path_idx_t saved_path = 0;
2300 	int ret;
2301 
2302 	EBUG_ON(btree_iter_path(trans, iter)->cached ||
2303 		btree_iter_path(trans, iter)->level);
2304 
2305 	if (iter->flags & BTREE_ITER_WITH_JOURNAL)
2306 		return bkey_s_c_err(-EIO);
2307 
2308 	bch2_btree_iter_verify(iter);
2309 	bch2_btree_iter_verify_entry_exit(iter);
2310 
2311 	if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2312 		search_key.snapshot = U32_MAX;
2313 
2314 	while (1) {
2315 		iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2316 						iter->flags & BTREE_ITER_INTENT,
2317 						btree_iter_ip_allocated(iter));
2318 
2319 		ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2320 		if (unlikely(ret)) {
2321 			/* ensure that iter->k is consistent with iter->pos: */
2322 			bch2_btree_iter_set_pos(iter, iter->pos);
2323 			k = bkey_s_c_err(ret);
2324 			goto out_no_locked;
2325 		}
2326 
2327 		struct btree_path *path = btree_iter_path(trans, iter);
2328 
2329 		k = btree_path_level_peek(trans, path, &path->l[0], &iter->k);
2330 		if (!k.k ||
2331 		    ((iter->flags & BTREE_ITER_IS_EXTENTS)
2332 		     ? bpos_ge(bkey_start_pos(k.k), search_key)
2333 		     : bpos_gt(k.k->p, search_key)))
2334 			k = btree_path_level_prev(trans, path, &path->l[0], &iter->k);
2335 
2336 		if (unlikely((iter->flags & BTREE_ITER_WITH_UPDATES) &&
2337 			     trans->nr_updates))
2338 			bch2_btree_trans_peek_prev_updates(trans, iter, &k);
2339 
2340 		if (likely(k.k)) {
2341 			if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) {
2342 				if (k.k->p.snapshot == iter->snapshot)
2343 					goto got_key;
2344 
2345 				/*
2346 				 * If we have a saved candidate, and we're no
2347 				 * longer at the same _key_ (not pos), return
2348 				 * that candidate
2349 				 */
2350 				if (saved_path && !bkey_eq(k.k->p, saved_k.p)) {
2351 					bch2_path_put_nokeep(trans, iter->path,
2352 						      iter->flags & BTREE_ITER_INTENT);
2353 					iter->path = saved_path;
2354 					saved_path = 0;
2355 					iter->k	= saved_k;
2356 					k.v	= saved_v;
2357 					goto got_key;
2358 				}
2359 
2360 				if (bch2_snapshot_is_ancestor(trans->c,
2361 							      iter->snapshot,
2362 							      k.k->p.snapshot)) {
2363 					if (saved_path)
2364 						bch2_path_put_nokeep(trans, saved_path,
2365 						      iter->flags & BTREE_ITER_INTENT);
2366 					saved_path = btree_path_clone(trans, iter->path,
2367 								iter->flags & BTREE_ITER_INTENT);
2368 					path = btree_iter_path(trans, iter);
2369 					saved_k = *k.k;
2370 					saved_v = k.v;
2371 				}
2372 
2373 				search_key = bpos_predecessor(k.k->p);
2374 				continue;
2375 			}
2376 got_key:
2377 			if (bkey_whiteout(k.k) &&
2378 			    !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2379 				search_key = bkey_predecessor(iter, k.k->p);
2380 				if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2381 					search_key.snapshot = U32_MAX;
2382 				continue;
2383 			}
2384 
2385 			btree_path_set_should_be_locked(path);
2386 			break;
2387 		} else if (likely(!bpos_eq(path->l[0].b->data->min_key, POS_MIN))) {
2388 			/* Advance to previous leaf node: */
2389 			search_key = bpos_predecessor(path->l[0].b->data->min_key);
2390 		} else {
2391 			/* Start of btree: */
2392 			bch2_btree_iter_set_pos(iter, POS_MIN);
2393 			k = bkey_s_c_null;
2394 			goto out_no_locked;
2395 		}
2396 	}
2397 
2398 	EBUG_ON(bkey_gt(bkey_start_pos(k.k), iter->pos));
2399 
2400 	/* Extents can straddle iter->pos: */
2401 	if (bkey_lt(k.k->p, iter->pos))
2402 		iter->pos = k.k->p;
2403 
2404 	if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2405 		iter->pos.snapshot = iter->snapshot;
2406 out_no_locked:
2407 	if (saved_path)
2408 		bch2_path_put_nokeep(trans, saved_path, iter->flags & BTREE_ITER_INTENT);
2409 
2410 	bch2_btree_iter_verify_entry_exit(iter);
2411 	bch2_btree_iter_verify(iter);
2412 
2413 	return k;
2414 }
2415 
2416 /**
2417  * bch2_btree_iter_prev() - returns first key less than iterator's current
2418  * position
2419  * @iter:	iterator to peek from
2420  *
2421  * Returns:	key if found, or an error extractable with bkey_err().
2422  */
2423 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
2424 {
2425 	if (!bch2_btree_iter_rewind(iter))
2426 		return bkey_s_c_null;
2427 
2428 	return bch2_btree_iter_peek_prev(iter);
2429 }
2430 
2431 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
2432 {
2433 	struct btree_trans *trans = iter->trans;
2434 	struct bpos search_key;
2435 	struct bkey_s_c k;
2436 	int ret;
2437 
2438 	bch2_btree_iter_verify(iter);
2439 	bch2_btree_iter_verify_entry_exit(iter);
2440 	EBUG_ON(btree_iter_path(trans, iter)->level && (iter->flags & BTREE_ITER_WITH_KEY_CACHE));
2441 
2442 	/* extents can't span inode numbers: */
2443 	if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
2444 	    unlikely(iter->pos.offset == KEY_OFFSET_MAX)) {
2445 		if (iter->pos.inode == KEY_INODE_MAX)
2446 			return bkey_s_c_null;
2447 
2448 		bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
2449 	}
2450 
2451 	search_key = btree_iter_search_key(iter);
2452 	iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2453 					iter->flags & BTREE_ITER_INTENT,
2454 					btree_iter_ip_allocated(iter));
2455 
2456 	ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2457 	if (unlikely(ret)) {
2458 		k = bkey_s_c_err(ret);
2459 		goto out_no_locked;
2460 	}
2461 
2462 	if ((iter->flags & BTREE_ITER_CACHED) ||
2463 	    !(iter->flags & (BTREE_ITER_IS_EXTENTS|BTREE_ITER_FILTER_SNAPSHOTS))) {
2464 		k = bkey_s_c_null;
2465 
2466 		if (unlikely((iter->flags & BTREE_ITER_WITH_UPDATES) &&
2467 			     trans->nr_updates)) {
2468 			bch2_btree_trans_peek_slot_updates(trans, iter, &k);
2469 			if (k.k)
2470 				goto out;
2471 		}
2472 
2473 		if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL) &&
2474 		    (k = btree_trans_peek_slot_journal(trans, iter)).k)
2475 			goto out;
2476 
2477 		if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2478 		    (k = btree_trans_peek_key_cache(iter, iter->pos)).k) {
2479 			if (!bkey_err(k))
2480 				iter->k = *k.k;
2481 			/* We're not returning a key from iter->path: */
2482 			goto out_no_locked;
2483 		}
2484 
2485 		k = bch2_btree_path_peek_slot(trans->paths + iter->path, &iter->k);
2486 		if (unlikely(!k.k))
2487 			goto out_no_locked;
2488 	} else {
2489 		struct bpos next;
2490 		struct bpos end = iter->pos;
2491 
2492 		if (iter->flags & BTREE_ITER_IS_EXTENTS)
2493 			end.offset = U64_MAX;
2494 
2495 		EBUG_ON(btree_iter_path(trans, iter)->level);
2496 
2497 		if (iter->flags & BTREE_ITER_INTENT) {
2498 			struct btree_iter iter2;
2499 
2500 			bch2_trans_copy_iter(&iter2, iter);
2501 			k = bch2_btree_iter_peek_upto(&iter2, end);
2502 
2503 			if (k.k && !bkey_err(k)) {
2504 				iter->k = iter2.k;
2505 				k.k = &iter->k;
2506 			}
2507 			bch2_trans_iter_exit(trans, &iter2);
2508 		} else {
2509 			struct bpos pos = iter->pos;
2510 
2511 			k = bch2_btree_iter_peek_upto(iter, end);
2512 			if (unlikely(bkey_err(k)))
2513 				bch2_btree_iter_set_pos(iter, pos);
2514 			else
2515 				iter->pos = pos;
2516 		}
2517 
2518 		if (unlikely(bkey_err(k)))
2519 			goto out_no_locked;
2520 
2521 		next = k.k ? bkey_start_pos(k.k) : POS_MAX;
2522 
2523 		if (bkey_lt(iter->pos, next)) {
2524 			bkey_init(&iter->k);
2525 			iter->k.p = iter->pos;
2526 
2527 			if (iter->flags & BTREE_ITER_IS_EXTENTS) {
2528 				bch2_key_resize(&iter->k,
2529 						min_t(u64, KEY_SIZE_MAX,
2530 						      (next.inode == iter->pos.inode
2531 						       ? next.offset
2532 						       : KEY_OFFSET_MAX) -
2533 						      iter->pos.offset));
2534 				EBUG_ON(!iter->k.size);
2535 			}
2536 
2537 			k = (struct bkey_s_c) { &iter->k, NULL };
2538 		}
2539 	}
2540 out:
2541 	btree_path_set_should_be_locked(btree_iter_path(trans, iter));
2542 out_no_locked:
2543 	bch2_btree_iter_verify_entry_exit(iter);
2544 	bch2_btree_iter_verify(iter);
2545 	ret = bch2_btree_iter_verify_ret(iter, k);
2546 	if (unlikely(ret))
2547 		return bkey_s_c_err(ret);
2548 
2549 	return k;
2550 }
2551 
2552 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
2553 {
2554 	if (!bch2_btree_iter_advance(iter))
2555 		return bkey_s_c_null;
2556 
2557 	return bch2_btree_iter_peek_slot(iter);
2558 }
2559 
2560 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter)
2561 {
2562 	if (!bch2_btree_iter_rewind(iter))
2563 		return bkey_s_c_null;
2564 
2565 	return bch2_btree_iter_peek_slot(iter);
2566 }
2567 
2568 struct bkey_s_c bch2_btree_iter_peek_and_restart_outlined(struct btree_iter *iter)
2569 {
2570 	struct bkey_s_c k;
2571 
2572 	while (btree_trans_too_many_iters(iter->trans) ||
2573 	       (k = bch2_btree_iter_peek_type(iter, iter->flags),
2574 		bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart)))
2575 		bch2_trans_begin(iter->trans);
2576 
2577 	return k;
2578 }
2579 
2580 /* new transactional stuff: */
2581 
2582 #ifdef CONFIG_BCACHEFS_DEBUG
2583 static void btree_trans_verify_sorted_refs(struct btree_trans *trans)
2584 {
2585 	struct btree_path *path;
2586 	unsigned i;
2587 
2588 	BUG_ON(trans->nr_sorted != bitmap_weight(trans->paths_allocated, trans->nr_paths) - 1);
2589 
2590 	trans_for_each_path(trans, path, i) {
2591 		BUG_ON(path->sorted_idx >= trans->nr_sorted);
2592 		BUG_ON(trans->sorted[path->sorted_idx] != i);
2593 	}
2594 
2595 	for (i = 0; i < trans->nr_sorted; i++) {
2596 		unsigned idx = trans->sorted[i];
2597 
2598 		BUG_ON(!test_bit(idx, trans->paths_allocated));
2599 		BUG_ON(trans->paths[idx].sorted_idx != i);
2600 	}
2601 }
2602 
2603 static void btree_trans_verify_sorted(struct btree_trans *trans)
2604 {
2605 	struct btree_path *path, *prev = NULL;
2606 	struct trans_for_each_path_inorder_iter iter;
2607 
2608 	if (!bch2_debug_check_iterators)
2609 		return;
2610 
2611 	trans_for_each_path_inorder(trans, path, iter) {
2612 		if (prev && btree_path_cmp(prev, path) > 0) {
2613 			__bch2_dump_trans_paths_updates(trans, true);
2614 			panic("trans paths out of order!\n");
2615 		}
2616 		prev = path;
2617 	}
2618 }
2619 #else
2620 static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans) {}
2621 static inline void btree_trans_verify_sorted(struct btree_trans *trans) {}
2622 #endif
2623 
2624 void __bch2_btree_trans_sort_paths(struct btree_trans *trans)
2625 {
2626 	int i, l = 0, r = trans->nr_sorted, inc = 1;
2627 	bool swapped;
2628 
2629 	btree_trans_verify_sorted_refs(trans);
2630 
2631 	if (trans->paths_sorted)
2632 		goto out;
2633 
2634 	/*
2635 	 * Cocktail shaker sort: this is efficient because iterators will be
2636 	 * mostly sorted.
2637 	 */
2638 	do {
2639 		swapped = false;
2640 
2641 		for (i = inc > 0 ? l : r - 2;
2642 		     i + 1 < r && i >= l;
2643 		     i += inc) {
2644 			if (btree_path_cmp(trans->paths + trans->sorted[i],
2645 					   trans->paths + trans->sorted[i + 1]) > 0) {
2646 				swap(trans->sorted[i], trans->sorted[i + 1]);
2647 				trans->paths[trans->sorted[i]].sorted_idx = i;
2648 				trans->paths[trans->sorted[i + 1]].sorted_idx = i + 1;
2649 				swapped = true;
2650 			}
2651 		}
2652 
2653 		if (inc > 0)
2654 			--r;
2655 		else
2656 			l++;
2657 		inc = -inc;
2658 	} while (swapped);
2659 
2660 	trans->paths_sorted = true;
2661 out:
2662 	btree_trans_verify_sorted(trans);
2663 }
2664 
2665 static inline void btree_path_list_remove(struct btree_trans *trans,
2666 					  struct btree_path *path)
2667 {
2668 	EBUG_ON(path->sorted_idx >= trans->nr_sorted);
2669 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2670 	trans->nr_sorted--;
2671 	memmove_u64s_down_small(trans->sorted + path->sorted_idx,
2672 				trans->sorted + path->sorted_idx + 1,
2673 				DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx,
2674 					     sizeof(u64) / sizeof(btree_path_idx_t)));
2675 #else
2676 	array_remove_item(trans->sorted, trans->nr_sorted, path->sorted_idx);
2677 #endif
2678 	for (unsigned i = path->sorted_idx; i < trans->nr_sorted; i++)
2679 		trans->paths[trans->sorted[i]].sorted_idx = i;
2680 }
2681 
2682 static inline void btree_path_list_add(struct btree_trans *trans,
2683 				       btree_path_idx_t pos,
2684 				       btree_path_idx_t path_idx)
2685 {
2686 	struct btree_path *path = trans->paths + path_idx;
2687 
2688 	path->sorted_idx = pos ? trans->paths[pos].sorted_idx + 1 : trans->nr_sorted;
2689 
2690 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2691 	memmove_u64s_up_small(trans->sorted + path->sorted_idx + 1,
2692 			      trans->sorted + path->sorted_idx,
2693 			      DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx,
2694 					   sizeof(u64) / sizeof(btree_path_idx_t)));
2695 	trans->nr_sorted++;
2696 	trans->sorted[path->sorted_idx] = path_idx;
2697 #else
2698 	array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path_idx);
2699 #endif
2700 
2701 	for (unsigned i = path->sorted_idx; i < trans->nr_sorted; i++)
2702 		trans->paths[trans->sorted[i]].sorted_idx = i;
2703 
2704 	btree_trans_verify_sorted_refs(trans);
2705 }
2706 
2707 void bch2_trans_iter_exit(struct btree_trans *trans, struct btree_iter *iter)
2708 {
2709 	if (iter->update_path)
2710 		bch2_path_put_nokeep(trans, iter->update_path,
2711 			      iter->flags & BTREE_ITER_INTENT);
2712 	if (iter->path)
2713 		bch2_path_put(trans, iter->path,
2714 			      iter->flags & BTREE_ITER_INTENT);
2715 	if (iter->key_cache_path)
2716 		bch2_path_put(trans, iter->key_cache_path,
2717 			      iter->flags & BTREE_ITER_INTENT);
2718 	iter->path		= 0;
2719 	iter->update_path	= 0;
2720 	iter->key_cache_path	= 0;
2721 	iter->trans		= NULL;
2722 }
2723 
2724 void bch2_trans_iter_init_outlined(struct btree_trans *trans,
2725 			  struct btree_iter *iter,
2726 			  enum btree_id btree_id, struct bpos pos,
2727 			  unsigned flags)
2728 {
2729 	bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
2730 			       bch2_btree_iter_flags(trans, btree_id, flags),
2731 			       _RET_IP_);
2732 }
2733 
2734 void bch2_trans_node_iter_init(struct btree_trans *trans,
2735 			       struct btree_iter *iter,
2736 			       enum btree_id btree_id,
2737 			       struct bpos pos,
2738 			       unsigned locks_want,
2739 			       unsigned depth,
2740 			       unsigned flags)
2741 {
2742 	flags |= BTREE_ITER_NOT_EXTENTS;
2743 	flags |= __BTREE_ITER_ALL_SNAPSHOTS;
2744 	flags |= BTREE_ITER_ALL_SNAPSHOTS;
2745 
2746 	bch2_trans_iter_init_common(trans, iter, btree_id, pos, locks_want, depth,
2747 			       __bch2_btree_iter_flags(trans, btree_id, flags),
2748 			       _RET_IP_);
2749 
2750 	iter->min_depth	= depth;
2751 
2752 	struct btree_path *path = btree_iter_path(trans, iter);
2753 	BUG_ON(path->locks_want	 < min(locks_want, BTREE_MAX_DEPTH));
2754 	BUG_ON(path->level	!= depth);
2755 	BUG_ON(iter->min_depth	!= depth);
2756 }
2757 
2758 void bch2_trans_copy_iter(struct btree_iter *dst, struct btree_iter *src)
2759 {
2760 	struct btree_trans *trans = src->trans;
2761 
2762 	*dst = *src;
2763 	if (src->path)
2764 		__btree_path_get(trans->paths + src->path, src->flags & BTREE_ITER_INTENT);
2765 	if (src->update_path)
2766 		__btree_path_get(trans->paths + src->update_path, src->flags & BTREE_ITER_INTENT);
2767 	dst->key_cache_path = 0;
2768 }
2769 
2770 void *__bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
2771 {
2772 	struct bch_fs *c = trans->c;
2773 	unsigned new_top = trans->mem_top + size;
2774 	unsigned old_bytes = trans->mem_bytes;
2775 	unsigned new_bytes = roundup_pow_of_two(new_top);
2776 	int ret;
2777 	void *new_mem;
2778 	void *p;
2779 
2780 	WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX);
2781 
2782 	struct btree_transaction_stats *s = btree_trans_stats(trans);
2783 	s->max_mem = max(s->max_mem, new_bytes);
2784 
2785 	new_mem = krealloc(trans->mem, new_bytes, GFP_NOWAIT|__GFP_NOWARN);
2786 	if (unlikely(!new_mem)) {
2787 		bch2_trans_unlock(trans);
2788 
2789 		new_mem = krealloc(trans->mem, new_bytes, GFP_KERNEL);
2790 		if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) {
2791 			new_mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL);
2792 			new_bytes = BTREE_TRANS_MEM_MAX;
2793 			kfree(trans->mem);
2794 		}
2795 
2796 		if (!new_mem)
2797 			return ERR_PTR(-BCH_ERR_ENOMEM_trans_kmalloc);
2798 
2799 		trans->mem = new_mem;
2800 		trans->mem_bytes = new_bytes;
2801 
2802 		ret = bch2_trans_relock(trans);
2803 		if (ret)
2804 			return ERR_PTR(ret);
2805 	}
2806 
2807 	trans->mem = new_mem;
2808 	trans->mem_bytes = new_bytes;
2809 
2810 	if (old_bytes) {
2811 		trace_and_count(c, trans_restart_mem_realloced, trans, _RET_IP_, new_bytes);
2812 		return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_mem_realloced));
2813 	}
2814 
2815 	p = trans->mem + trans->mem_top;
2816 	trans->mem_top += size;
2817 	memset(p, 0, size);
2818 	return p;
2819 }
2820 
2821 static inline void check_srcu_held_too_long(struct btree_trans *trans)
2822 {
2823 	WARN(trans->srcu_held && time_after(jiffies, trans->srcu_lock_time + HZ * 10),
2824 	     "btree trans held srcu lock (delaying memory reclaim) for %lu seconds",
2825 	     (jiffies - trans->srcu_lock_time) / HZ);
2826 }
2827 
2828 void bch2_trans_srcu_unlock(struct btree_trans *trans)
2829 {
2830 	if (trans->srcu_held) {
2831 		struct bch_fs *c = trans->c;
2832 		struct btree_path *path;
2833 		unsigned i;
2834 
2835 		trans_for_each_path(trans, path, i)
2836 			if (path->cached && !btree_node_locked(path, 0))
2837 				path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_srcu_reset);
2838 
2839 		check_srcu_held_too_long(trans);
2840 		srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
2841 		trans->srcu_held = false;
2842 	}
2843 }
2844 
2845 static void bch2_trans_srcu_lock(struct btree_trans *trans)
2846 {
2847 	if (!trans->srcu_held) {
2848 		trans->srcu_idx = srcu_read_lock(&trans->c->btree_trans_barrier);
2849 		trans->srcu_lock_time	= jiffies;
2850 		trans->srcu_held = true;
2851 	}
2852 }
2853 
2854 /**
2855  * bch2_trans_begin() - reset a transaction after a interrupted attempt
2856  * @trans: transaction to reset
2857  *
2858  * Returns:	current restart counter, to be used with trans_was_restarted()
2859  *
2860  * While iterating over nodes or updating nodes a attempt to lock a btree node
2861  * may return BCH_ERR_transaction_restart when the trylock fails. When this
2862  * occurs bch2_trans_begin() should be called and the transaction retried.
2863  */
2864 u32 bch2_trans_begin(struct btree_trans *trans)
2865 {
2866 	struct btree_path *path;
2867 	unsigned i;
2868 	u64 now;
2869 
2870 	bch2_trans_reset_updates(trans);
2871 
2872 	trans->restart_count++;
2873 	trans->mem_top			= 0;
2874 	trans->journal_entries		= NULL;
2875 
2876 	trans_for_each_path(trans, path, i) {
2877 		path->should_be_locked = false;
2878 
2879 		/*
2880 		 * If the transaction wasn't restarted, we're presuming to be
2881 		 * doing something new: dont keep iterators excpt the ones that
2882 		 * are in use - except for the subvolumes btree:
2883 		 */
2884 		if (!trans->restarted && path->btree_id != BTREE_ID_subvolumes)
2885 			path->preserve = false;
2886 
2887 		/*
2888 		 * XXX: we probably shouldn't be doing this if the transaction
2889 		 * was restarted, but currently we still overflow transaction
2890 		 * iterators if we do that
2891 		 */
2892 		if (!path->ref && !path->preserve)
2893 			__bch2_path_free(trans, i);
2894 		else
2895 			path->preserve = false;
2896 	}
2897 
2898 	now = local_clock();
2899 
2900 	if (!IS_ENABLED(CONFIG_BCACHEFS_NO_LATENCY_ACCT) &&
2901 	    time_after64(now, trans->last_begin_time + 10))
2902 		__bch2_time_stats_update(&btree_trans_stats(trans)->duration,
2903 					 trans->last_begin_time, now);
2904 
2905 	if (!trans->restarted &&
2906 	    (need_resched() ||
2907 	     time_after64(now, trans->last_begin_time + BTREE_TRANS_MAX_LOCK_HOLD_TIME_NS))) {
2908 		drop_locks_do(trans, (cond_resched(), 0));
2909 		now = local_clock();
2910 	}
2911 	trans->last_begin_time = now;
2912 
2913 	if (unlikely(trans->srcu_held &&
2914 		     time_after(jiffies, trans->srcu_lock_time + msecs_to_jiffies(10))))
2915 		bch2_trans_srcu_unlock(trans);
2916 
2917 	trans->last_begin_ip = _RET_IP_;
2918 	if (trans->restarted) {
2919 		bch2_btree_path_traverse_all(trans);
2920 		trans->notrace_relock_fail = false;
2921 	}
2922 
2923 	return trans->restart_count;
2924 }
2925 
2926 const char *bch2_btree_transaction_fns[BCH_TRANSACTIONS_NR] = { "(unknown)" };
2927 
2928 unsigned bch2_trans_get_fn_idx(const char *fn)
2929 {
2930 	for (unsigned i = 0; i < ARRAY_SIZE(bch2_btree_transaction_fns); i++)
2931 		if (!bch2_btree_transaction_fns[i] ||
2932 		    bch2_btree_transaction_fns[i] == fn) {
2933 			bch2_btree_transaction_fns[i] = fn;
2934 			return i;
2935 		}
2936 
2937 	pr_warn_once("BCH_TRANSACTIONS_NR not big enough!");
2938 	return 0;
2939 }
2940 
2941 struct btree_trans *__bch2_trans_get(struct bch_fs *c, unsigned fn_idx)
2942 	__acquires(&c->btree_trans_barrier)
2943 {
2944 	struct btree_trans *trans;
2945 
2946 	if (IS_ENABLED(__KERNEL__)) {
2947 		trans = this_cpu_xchg(c->btree_trans_bufs->trans, NULL);
2948 		if (trans) {
2949 			memset(trans, 0, offsetof(struct btree_trans, list));
2950 			goto got_trans;
2951 		}
2952 	}
2953 
2954 	trans = mempool_alloc(&c->btree_trans_pool, GFP_NOFS);
2955 	memset(trans, 0, sizeof(*trans));
2956 	closure_init_stack(&trans->ref);
2957 
2958 	seqmutex_lock(&c->btree_trans_lock);
2959 	if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) {
2960 		struct btree_trans *pos;
2961 		pid_t pid = current->pid;
2962 
2963 		trans->locking_wait.task = current;
2964 
2965 		list_for_each_entry(pos, &c->btree_trans_list, list) {
2966 			struct task_struct *pos_task = READ_ONCE(pos->locking_wait.task);
2967 			/*
2968 			 * We'd much prefer to be stricter here and completely
2969 			 * disallow multiple btree_trans in the same thread -
2970 			 * but the data move path calls bch2_write when we
2971 			 * already have a btree_trans initialized.
2972 			 */
2973 			BUG_ON(pos_task &&
2974 			       pid == pos_task->pid &&
2975 			       bch2_trans_locked(pos));
2976 
2977 			if (pos_task && pid < pos_task->pid) {
2978 				list_add_tail(&trans->list, &pos->list);
2979 				goto list_add_done;
2980 			}
2981 		}
2982 	}
2983 	list_add_tail(&trans->list, &c->btree_trans_list);
2984 list_add_done:
2985 	seqmutex_unlock(&c->btree_trans_lock);
2986 got_trans:
2987 	trans->c		= c;
2988 	trans->last_begin_time	= local_clock();
2989 	trans->fn_idx		= fn_idx;
2990 	trans->locking_wait.task = current;
2991 	trans->journal_replay_not_finished =
2992 		unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)) &&
2993 		atomic_inc_not_zero(&c->journal_keys.ref);
2994 	trans->nr_paths		= ARRAY_SIZE(trans->_paths);
2995 	trans->paths_allocated	= trans->_paths_allocated;
2996 	trans->sorted		= trans->_sorted;
2997 	trans->paths		= trans->_paths;
2998 	trans->updates		= trans->_updates;
2999 
3000 	*trans_paths_nr(trans->paths) = BTREE_ITER_INITIAL;
3001 
3002 	trans->paths_allocated[0] = 1;
3003 
3004 	if (fn_idx < BCH_TRANSACTIONS_NR) {
3005 		trans->fn = bch2_btree_transaction_fns[fn_idx];
3006 
3007 		struct btree_transaction_stats *s = &c->btree_transaction_stats[fn_idx];
3008 
3009 		if (s->max_mem) {
3010 			unsigned expected_mem_bytes = roundup_pow_of_two(s->max_mem);
3011 
3012 			trans->mem = kmalloc(expected_mem_bytes, GFP_KERNEL);
3013 			if (likely(trans->mem))
3014 				trans->mem_bytes = expected_mem_bytes;
3015 		}
3016 
3017 		trans->nr_paths_max = s->nr_max_paths;
3018 		trans->journal_entries_size = s->journal_entries_size;
3019 	}
3020 
3021 	trans->srcu_idx		= srcu_read_lock(&c->btree_trans_barrier);
3022 	trans->srcu_lock_time	= jiffies;
3023 	trans->srcu_held	= true;
3024 	return trans;
3025 }
3026 
3027 static void check_btree_paths_leaked(struct btree_trans *trans)
3028 {
3029 #ifdef CONFIG_BCACHEFS_DEBUG
3030 	struct bch_fs *c = trans->c;
3031 	struct btree_path *path;
3032 	unsigned i;
3033 
3034 	trans_for_each_path(trans, path, i)
3035 		if (path->ref)
3036 			goto leaked;
3037 	return;
3038 leaked:
3039 	bch_err(c, "btree paths leaked from %s!", trans->fn);
3040 	trans_for_each_path(trans, path, i)
3041 		if (path->ref)
3042 			printk(KERN_ERR "  btree %s %pS\n",
3043 			       bch2_btree_id_str(path->btree_id),
3044 			       (void *) path->ip_allocated);
3045 	/* Be noisy about this: */
3046 	bch2_fatal_error(c);
3047 #endif
3048 }
3049 
3050 void bch2_trans_put(struct btree_trans *trans)
3051 	__releases(&c->btree_trans_barrier)
3052 {
3053 	struct bch_fs *c = trans->c;
3054 
3055 	bch2_trans_unlock(trans);
3056 
3057 	trans_for_each_update(trans, i)
3058 		__btree_path_put(trans->paths + i->path, true);
3059 	trans->nr_updates	= 0;
3060 	trans->locking_wait.task = NULL;
3061 
3062 	check_btree_paths_leaked(trans);
3063 
3064 	if (trans->srcu_held) {
3065 		check_srcu_held_too_long(trans);
3066 		srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
3067 	}
3068 
3069 	if (trans->fs_usage_deltas) {
3070 		if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) ==
3071 		    REPLICAS_DELTA_LIST_MAX)
3072 			mempool_free(trans->fs_usage_deltas,
3073 				     &c->replicas_delta_pool);
3074 		else
3075 			kfree(trans->fs_usage_deltas);
3076 	}
3077 
3078 	if (unlikely(trans->journal_replay_not_finished))
3079 		bch2_journal_keys_put(c);
3080 
3081 	unsigned long *paths_allocated = trans->paths_allocated;
3082 	trans->paths_allocated	= NULL;
3083 	trans->paths		= NULL;
3084 
3085 	if (paths_allocated != trans->_paths_allocated)
3086 		kfree_rcu_mightsleep(paths_allocated);
3087 
3088 	if (trans->mem_bytes == BTREE_TRANS_MEM_MAX)
3089 		mempool_free(trans->mem, &c->btree_trans_mem_pool);
3090 	else
3091 		kfree(trans->mem);
3092 
3093 	/* Userspace doesn't have a real percpu implementation: */
3094 	if (IS_ENABLED(__KERNEL__))
3095 		trans = this_cpu_xchg(c->btree_trans_bufs->trans, trans);
3096 
3097 	if (trans) {
3098 		closure_sync(&trans->ref);
3099 
3100 		seqmutex_lock(&c->btree_trans_lock);
3101 		list_del(&trans->list);
3102 		seqmutex_unlock(&c->btree_trans_lock);
3103 
3104 		mempool_free(trans, &c->btree_trans_pool);
3105 	}
3106 }
3107 
3108 static void __maybe_unused
3109 bch2_btree_bkey_cached_common_to_text(struct printbuf *out,
3110 				      struct btree_bkey_cached_common *b)
3111 {
3112 	struct six_lock_count c = six_lock_counts(&b->lock);
3113 	struct task_struct *owner;
3114 	pid_t pid;
3115 
3116 	rcu_read_lock();
3117 	owner = READ_ONCE(b->lock.owner);
3118 	pid = owner ? owner->pid : 0;
3119 	rcu_read_unlock();
3120 
3121 	prt_tab(out);
3122 	prt_printf(out, "%px %c l=%u %s:", b, b->cached ? 'c' : 'b',
3123 		   b->level, bch2_btree_id_str(b->btree_id));
3124 	bch2_bpos_to_text(out, btree_node_pos(b));
3125 
3126 	prt_tab(out);
3127 	prt_printf(out, " locks %u:%u:%u held by pid %u",
3128 		   c.n[0], c.n[1], c.n[2], pid);
3129 }
3130 
3131 void bch2_btree_trans_to_text(struct printbuf *out, struct btree_trans *trans)
3132 {
3133 	struct btree_bkey_cached_common *b;
3134 	static char lock_types[] = { 'r', 'i', 'w' };
3135 	struct task_struct *task = READ_ONCE(trans->locking_wait.task);
3136 	unsigned l, idx;
3137 
3138 	/* before rcu_read_lock(): */
3139 	bch2_printbuf_make_room(out, 4096);
3140 
3141 	if (!out->nr_tabstops) {
3142 		printbuf_tabstop_push(out, 16);
3143 		printbuf_tabstop_push(out, 32);
3144 	}
3145 
3146 	prt_printf(out, "%i %s\n", task ? task->pid : 0, trans->fn);
3147 
3148 	/* trans->paths is rcu protected vs. freeing */
3149 	rcu_read_lock();
3150 	out->atomic++;
3151 
3152 	struct btree_path *paths = rcu_dereference(trans->paths);
3153 	if (!paths)
3154 		goto out;
3155 
3156 	unsigned long *paths_allocated = trans_paths_allocated(paths);
3157 
3158 	trans_for_each_path_idx_from(paths_allocated, *trans_paths_nr(paths), idx, 1) {
3159 		struct btree_path *path = paths + idx;
3160 		if (!path->nodes_locked)
3161 			continue;
3162 
3163 		prt_printf(out, "  path %u %c l=%u %s:",
3164 		       idx,
3165 		       path->cached ? 'c' : 'b',
3166 		       path->level,
3167 		       bch2_btree_id_str(path->btree_id));
3168 		bch2_bpos_to_text(out, path->pos);
3169 		prt_newline(out);
3170 
3171 		for (l = 0; l < BTREE_MAX_DEPTH; l++) {
3172 			if (btree_node_locked(path, l) &&
3173 			    !IS_ERR_OR_NULL(b = (void *) READ_ONCE(path->l[l].b))) {
3174 				prt_printf(out, "    %c l=%u ",
3175 					   lock_types[btree_node_locked_type(path, l)], l);
3176 				bch2_btree_bkey_cached_common_to_text(out, b);
3177 				prt_newline(out);
3178 			}
3179 		}
3180 	}
3181 
3182 	b = READ_ONCE(trans->locking);
3183 	if (b) {
3184 		prt_printf(out, "  blocked for %lluus on",
3185 			   div_u64(local_clock() - trans->locking_wait.start_time,
3186 				   1000));
3187 		prt_newline(out);
3188 		prt_printf(out, "    %c", lock_types[trans->locking_wait.lock_want]);
3189 		bch2_btree_bkey_cached_common_to_text(out, b);
3190 		prt_newline(out);
3191 	}
3192 out:
3193 	--out->atomic;
3194 	rcu_read_unlock();
3195 }
3196 
3197 void bch2_fs_btree_iter_exit(struct bch_fs *c)
3198 {
3199 	struct btree_transaction_stats *s;
3200 	struct btree_trans *trans;
3201 	int cpu;
3202 
3203 	if (c->btree_trans_bufs)
3204 		for_each_possible_cpu(cpu) {
3205 			struct btree_trans *trans =
3206 				per_cpu_ptr(c->btree_trans_bufs, cpu)->trans;
3207 
3208 			if (trans) {
3209 				closure_sync(&trans->ref);
3210 
3211 				seqmutex_lock(&c->btree_trans_lock);
3212 				list_del(&trans->list);
3213 				seqmutex_unlock(&c->btree_trans_lock);
3214 			}
3215 			kfree(trans);
3216 		}
3217 	free_percpu(c->btree_trans_bufs);
3218 
3219 	trans = list_first_entry_or_null(&c->btree_trans_list, struct btree_trans, list);
3220 	if (trans)
3221 		panic("%s leaked btree_trans\n", trans->fn);
3222 
3223 	for (s = c->btree_transaction_stats;
3224 	     s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3225 	     s++) {
3226 		kfree(s->max_paths_text);
3227 		bch2_time_stats_exit(&s->lock_hold_times);
3228 	}
3229 
3230 	if (c->btree_trans_barrier_initialized)
3231 		cleanup_srcu_struct(&c->btree_trans_barrier);
3232 	mempool_exit(&c->btree_trans_mem_pool);
3233 	mempool_exit(&c->btree_trans_pool);
3234 }
3235 
3236 void bch2_fs_btree_iter_init_early(struct bch_fs *c)
3237 {
3238 	struct btree_transaction_stats *s;
3239 
3240 	for (s = c->btree_transaction_stats;
3241 	     s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3242 	     s++) {
3243 		bch2_time_stats_init(&s->duration);
3244 		bch2_time_stats_init(&s->lock_hold_times);
3245 		mutex_init(&s->lock);
3246 	}
3247 
3248 	INIT_LIST_HEAD(&c->btree_trans_list);
3249 	seqmutex_init(&c->btree_trans_lock);
3250 }
3251 
3252 int bch2_fs_btree_iter_init(struct bch_fs *c)
3253 {
3254 	int ret;
3255 
3256 	c->btree_trans_bufs = alloc_percpu(struct btree_trans_buf);
3257 	if (!c->btree_trans_bufs)
3258 		return -ENOMEM;
3259 
3260 	ret   = mempool_init_kmalloc_pool(&c->btree_trans_pool, 1,
3261 					  sizeof(struct btree_trans)) ?:
3262 		mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1,
3263 					  BTREE_TRANS_MEM_MAX) ?:
3264 		init_srcu_struct(&c->btree_trans_barrier);
3265 	if (!ret)
3266 		c->btree_trans_barrier_initialized = true;
3267 	return ret;
3268 }
3269