xref: /linux/fs/bcachefs/btree_trans_commit.c (revision f057b57270c2a17d3f45c177e9434fa5745caa48)
1 // SPDX-License-Identifier: GPL-2.0
2 
3 #include "bcachefs.h"
4 #include "alloc_foreground.h"
5 #include "btree_gc.h"
6 #include "btree_io.h"
7 #include "btree_iter.h"
8 #include "btree_journal_iter.h"
9 #include "btree_key_cache.h"
10 #include "btree_update_interior.h"
11 #include "btree_write_buffer.h"
12 #include "buckets.h"
13 #include "disk_accounting.h"
14 #include "errcode.h"
15 #include "error.h"
16 #include "journal.h"
17 #include "journal_io.h"
18 #include "journal_reclaim.h"
19 #include "replicas.h"
20 #include "snapshot.h"
21 
22 #include <linux/prefetch.h>
23 
24 static const char * const trans_commit_flags_strs[] = {
25 #define x(n, ...) #n,
26 	BCH_TRANS_COMMIT_FLAGS()
27 #undef x
28 	NULL
29 };
30 
31 void bch2_trans_commit_flags_to_text(struct printbuf *out, enum bch_trans_commit_flags flags)
32 {
33 	enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
34 
35 	prt_printf(out, "watermark=%s", bch2_watermarks[watermark]);
36 
37 	flags >>= BCH_WATERMARK_BITS;
38 	if (flags) {
39 		prt_char(out, ' ');
40 		bch2_prt_bitflags(out, trans_commit_flags_strs, flags);
41 	}
42 }
43 
44 static void verify_update_old_key(struct btree_trans *trans, struct btree_insert_entry *i)
45 {
46 #ifdef CONFIG_BCACHEFS_DEBUG
47 	struct bch_fs *c = trans->c;
48 	struct bkey u;
49 	struct bkey_s_c k = bch2_btree_path_peek_slot_exact(trans->paths + i->path, &u);
50 
51 	if (unlikely(trans->journal_replay_not_finished)) {
52 		struct bkey_i *j_k =
53 			bch2_journal_keys_peek_slot(c, i->btree_id, i->level, i->k->k.p);
54 
55 		if (j_k)
56 			k = bkey_i_to_s_c(j_k);
57 	}
58 
59 	u = *k.k;
60 	u.needs_whiteout = i->old_k.needs_whiteout;
61 
62 	BUG_ON(memcmp(&i->old_k, &u, sizeof(struct bkey)));
63 	BUG_ON(i->old_v != k.v);
64 #endif
65 }
66 
67 static inline struct btree_path_level *insert_l(struct btree_trans *trans, struct btree_insert_entry *i)
68 {
69 	return (trans->paths + i->path)->l + i->level;
70 }
71 
72 static inline bool same_leaf_as_prev(struct btree_trans *trans,
73 				     struct btree_insert_entry *i)
74 {
75 	return i != trans->updates &&
76 		insert_l(trans, &i[0])->b == insert_l(trans, &i[-1])->b;
77 }
78 
79 static inline bool same_leaf_as_next(struct btree_trans *trans,
80 				     struct btree_insert_entry *i)
81 {
82 	return i + 1 < trans->updates + trans->nr_updates &&
83 		insert_l(trans, &i[0])->b == insert_l(trans, &i[1])->b;
84 }
85 
86 inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
87 					   struct btree_path *path,
88 					   struct btree *b)
89 {
90 	struct bch_fs *c = trans->c;
91 
92 	if (unlikely(btree_node_just_written(b)) &&
93 	    bch2_btree_post_write_cleanup(c, b))
94 		bch2_trans_node_reinit_iter(trans, b);
95 
96 	/*
97 	 * If the last bset has been written, or if it's gotten too big - start
98 	 * a new bset to insert into:
99 	 */
100 	if (want_new_bset(c, b))
101 		bch2_btree_init_next(trans, b);
102 }
103 
104 static noinline int trans_lock_write_fail(struct btree_trans *trans, struct btree_insert_entry *i)
105 {
106 	while (--i >= trans->updates) {
107 		if (same_leaf_as_prev(trans, i))
108 			continue;
109 
110 		bch2_btree_node_unlock_write(trans, trans->paths + i->path, insert_l(trans, i)->b);
111 	}
112 
113 	trace_and_count(trans->c, trans_restart_would_deadlock_write, trans);
114 	return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
115 }
116 
117 static inline int bch2_trans_lock_write(struct btree_trans *trans)
118 {
119 	EBUG_ON(trans->write_locked);
120 
121 	trans_for_each_update(trans, i) {
122 		if (same_leaf_as_prev(trans, i))
123 			continue;
124 
125 		if (bch2_btree_node_lock_write(trans, trans->paths + i->path, &insert_l(trans, i)->b->c))
126 			return trans_lock_write_fail(trans, i);
127 
128 		if (!i->cached)
129 			bch2_btree_node_prep_for_write(trans, trans->paths + i->path, insert_l(trans, i)->b);
130 	}
131 
132 	trans->write_locked = true;
133 	return 0;
134 }
135 
136 static inline void bch2_trans_unlock_write(struct btree_trans *trans)
137 {
138 	if (likely(trans->write_locked)) {
139 		trans_for_each_update(trans, i)
140 			if (btree_node_locked_type(trans->paths + i->path, i->level) ==
141 			    BTREE_NODE_WRITE_LOCKED)
142 				bch2_btree_node_unlock_write_inlined(trans,
143 						trans->paths + i->path, insert_l(trans, i)->b);
144 		trans->write_locked = false;
145 	}
146 }
147 
148 /* Inserting into a given leaf node (last stage of insert): */
149 
150 /* Handle overwrites and do insert, for non extents: */
151 bool bch2_btree_bset_insert_key(struct btree_trans *trans,
152 				struct btree_path *path,
153 				struct btree *b,
154 				struct btree_node_iter *node_iter,
155 				struct bkey_i *insert)
156 {
157 	struct bkey_packed *k;
158 	unsigned clobber_u64s = 0, new_u64s = 0;
159 
160 	EBUG_ON(btree_node_just_written(b));
161 	EBUG_ON(bset_written(b, btree_bset_last(b)));
162 	EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
163 	EBUG_ON(bpos_lt(insert->k.p, b->data->min_key));
164 	EBUG_ON(bpos_gt(insert->k.p, b->data->max_key));
165 	EBUG_ON(insert->k.u64s > bch2_btree_keys_u64s_remaining(b));
166 	EBUG_ON(!b->c.level && !bpos_eq(insert->k.p, path->pos));
167 
168 	k = bch2_btree_node_iter_peek_all(node_iter, b);
169 	if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
170 		k = NULL;
171 
172 	/* @k is the key being overwritten/deleted, if any: */
173 	EBUG_ON(k && bkey_deleted(k));
174 
175 	/* Deleting, but not found? nothing to do: */
176 	if (bkey_deleted(&insert->k) && !k)
177 		return false;
178 
179 	if (bkey_deleted(&insert->k)) {
180 		/* Deleting: */
181 		btree_account_key_drop(b, k);
182 		k->type = KEY_TYPE_deleted;
183 
184 		if (k->needs_whiteout)
185 			push_whiteout(b, insert->k.p);
186 		k->needs_whiteout = false;
187 
188 		if (k >= btree_bset_last(b)->start) {
189 			clobber_u64s = k->u64s;
190 			bch2_bset_delete(b, k, clobber_u64s);
191 			goto fix_iter;
192 		} else {
193 			bch2_btree_path_fix_key_modified(trans, b, k);
194 		}
195 
196 		return true;
197 	}
198 
199 	if (k) {
200 		/* Overwriting: */
201 		btree_account_key_drop(b, k);
202 		k->type = KEY_TYPE_deleted;
203 
204 		insert->k.needs_whiteout = k->needs_whiteout;
205 		k->needs_whiteout = false;
206 
207 		if (k >= btree_bset_last(b)->start) {
208 			clobber_u64s = k->u64s;
209 			goto overwrite;
210 		} else {
211 			bch2_btree_path_fix_key_modified(trans, b, k);
212 		}
213 	}
214 
215 	k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
216 overwrite:
217 	bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
218 	new_u64s = k->u64s;
219 fix_iter:
220 	if (clobber_u64s != new_u64s)
221 		bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
222 					 clobber_u64s, new_u64s);
223 	return true;
224 }
225 
226 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
227 			       unsigned i, u64 seq)
228 {
229 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
230 	struct btree_write *w = container_of(pin, struct btree_write, journal);
231 	struct btree *b = container_of(w, struct btree, writes[i]);
232 	struct btree_trans *trans = bch2_trans_get(c);
233 	unsigned long old, new;
234 	unsigned idx = w - b->writes;
235 
236 	btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
237 
238 	old = READ_ONCE(b->flags);
239 	do {
240 		new = old;
241 
242 		if (!(old & (1 << BTREE_NODE_dirty)) ||
243 		    !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
244 		    w->journal.seq != seq)
245 			break;
246 
247 		new &= ~BTREE_WRITE_TYPE_MASK;
248 		new |= BTREE_WRITE_journal_reclaim;
249 		new |= 1 << BTREE_NODE_need_write;
250 	} while (!try_cmpxchg(&b->flags, &old, new));
251 
252 	btree_node_write_if_need(c, b, SIX_LOCK_read);
253 	six_unlock_read(&b->c.lock);
254 
255 	bch2_trans_put(trans);
256 	return 0;
257 }
258 
259 int bch2_btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
260 {
261 	return __btree_node_flush(j, pin, 0, seq);
262 }
263 
264 int bch2_btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
265 {
266 	return __btree_node_flush(j, pin, 1, seq);
267 }
268 
269 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
270 				       struct btree *b, u64 seq)
271 {
272 	struct btree_write *w = btree_current_write(b);
273 
274 	bch2_journal_pin_add(&c->journal, seq, &w->journal,
275 			     btree_node_write_idx(b) == 0
276 			     ? bch2_btree_node_flush0
277 			     : bch2_btree_node_flush1);
278 }
279 
280 /**
281  * bch2_btree_insert_key_leaf() - insert a key one key into a leaf node
282  * @trans:		btree transaction object
283  * @path:		path pointing to @insert's pos
284  * @insert:		key to insert
285  * @journal_seq:	sequence number of journal reservation
286  */
287 inline void bch2_btree_insert_key_leaf(struct btree_trans *trans,
288 				       struct btree_path *path,
289 				       struct bkey_i *insert,
290 				       u64 journal_seq)
291 {
292 	struct bch_fs *c = trans->c;
293 	struct btree *b = path_l(path)->b;
294 	struct bset_tree *t = bset_tree_last(b);
295 	struct bset *i = bset(b, t);
296 	int old_u64s = bset_u64s(t);
297 	int old_live_u64s = b->nr.live_u64s;
298 	int live_u64s_added, u64s_added;
299 
300 	if (unlikely(!bch2_btree_bset_insert_key(trans, path, b,
301 					&path_l(path)->iter, insert)))
302 		return;
303 
304 	i->journal_seq = cpu_to_le64(max(journal_seq, le64_to_cpu(i->journal_seq)));
305 
306 	bch2_btree_add_journal_pin(c, b, journal_seq);
307 
308 	if (unlikely(!btree_node_dirty(b))) {
309 		EBUG_ON(test_bit(BCH_FS_clean_shutdown, &c->flags));
310 		set_btree_node_dirty_acct(c, b);
311 	}
312 
313 	live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
314 	u64s_added = (int) bset_u64s(t) - old_u64s;
315 
316 	if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
317 		b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
318 	if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
319 		b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
320 
321 	if (u64s_added > live_u64s_added &&
322 	    bch2_maybe_compact_whiteouts(c, b))
323 		bch2_trans_node_reinit_iter(trans, b);
324 }
325 
326 /* Cached btree updates: */
327 
328 /* Normal update interface: */
329 
330 static inline void btree_insert_entry_checks(struct btree_trans *trans,
331 					     struct btree_insert_entry *i)
332 {
333 	struct btree_path *path = trans->paths + i->path;
334 
335 	BUG_ON(!bpos_eq(i->k->k.p, path->pos));
336 	BUG_ON(i->cached	!= path->cached);
337 	BUG_ON(i->level		!= path->level);
338 	BUG_ON(i->btree_id	!= path->btree_id);
339 	EBUG_ON(!i->level &&
340 		btree_type_has_snapshots(i->btree_id) &&
341 		!(i->flags & BTREE_UPDATE_internal_snapshot_node) &&
342 		test_bit(JOURNAL_replay_done, &trans->c->journal.flags) &&
343 		i->k->k.p.snapshot &&
344 		bch2_snapshot_is_internal_node(trans->c, i->k->k.p.snapshot) > 0);
345 }
346 
347 static __always_inline int bch2_trans_journal_res_get(struct btree_trans *trans,
348 						      unsigned flags)
349 {
350 	return bch2_journal_res_get(&trans->c->journal, &trans->journal_res,
351 				    trans->journal_u64s, flags);
352 }
353 
354 #define JSET_ENTRY_LOG_U64s		4
355 
356 static noinline void journal_transaction_name(struct btree_trans *trans)
357 {
358 	struct bch_fs *c = trans->c;
359 	struct journal *j = &c->journal;
360 	struct jset_entry *entry =
361 		bch2_journal_add_entry(j, &trans->journal_res,
362 				       BCH_JSET_ENTRY_log, 0, 0,
363 				       JSET_ENTRY_LOG_U64s);
364 	struct jset_entry_log *l =
365 		container_of(entry, struct jset_entry_log, entry);
366 
367 	strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
368 }
369 
370 static inline int btree_key_can_insert(struct btree_trans *trans,
371 				       struct btree *b, unsigned u64s)
372 {
373 	if (!bch2_btree_node_insert_fits(b, u64s))
374 		return -BCH_ERR_btree_insert_btree_node_full;
375 
376 	return 0;
377 }
378 
379 noinline static int
380 btree_key_can_insert_cached_slowpath(struct btree_trans *trans, unsigned flags,
381 				     struct btree_path *path, unsigned new_u64s)
382 {
383 	struct bkey_cached *ck = (void *) path->l[0].b;
384 	struct bkey_i *new_k;
385 	int ret;
386 
387 	bch2_trans_unlock_write(trans);
388 	bch2_trans_unlock(trans);
389 
390 	new_k = kmalloc(new_u64s * sizeof(u64), GFP_KERNEL);
391 	if (!new_k) {
392 		bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
393 			bch2_btree_id_str(path->btree_id), new_u64s);
394 		return -BCH_ERR_ENOMEM_btree_key_cache_insert;
395 	}
396 
397 	ret =   bch2_trans_relock(trans) ?:
398 		bch2_trans_lock_write(trans);
399 	if (unlikely(ret)) {
400 		kfree(new_k);
401 		return ret;
402 	}
403 
404 	memcpy(new_k, ck->k, ck->u64s * sizeof(u64));
405 
406 	trans_for_each_update(trans, i)
407 		if (i->old_v == &ck->k->v)
408 			i->old_v = &new_k->v;
409 
410 	kfree(ck->k);
411 	ck->u64s	= new_u64s;
412 	ck->k		= new_k;
413 	return 0;
414 }
415 
416 static int btree_key_can_insert_cached(struct btree_trans *trans, unsigned flags,
417 				       struct btree_path *path, unsigned u64s)
418 {
419 	struct bch_fs *c = trans->c;
420 	struct bkey_cached *ck = (void *) path->l[0].b;
421 	unsigned new_u64s;
422 	struct bkey_i *new_k;
423 	unsigned watermark = flags & BCH_WATERMARK_MASK;
424 
425 	EBUG_ON(path->level);
426 
427 	if (watermark < BCH_WATERMARK_reclaim &&
428 	    !test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
429 	    bch2_btree_key_cache_must_wait(c))
430 		return -BCH_ERR_btree_insert_need_journal_reclaim;
431 
432 	/*
433 	 * bch2_varint_decode can read past the end of the buffer by at most 7
434 	 * bytes (it won't be used):
435 	 */
436 	u64s += 1;
437 
438 	if (u64s <= ck->u64s)
439 		return 0;
440 
441 	new_u64s	= roundup_pow_of_two(u64s);
442 	new_k		= krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOWAIT|__GFP_NOWARN);
443 	if (unlikely(!new_k))
444 		return btree_key_can_insert_cached_slowpath(trans, flags, path, new_u64s);
445 
446 	trans_for_each_update(trans, i)
447 		if (i->old_v == &ck->k->v)
448 			i->old_v = &new_k->v;
449 
450 	ck->u64s	= new_u64s;
451 	ck->k		= new_k;
452 	return 0;
453 }
454 
455 /* Triggers: */
456 
457 static int run_one_mem_trigger(struct btree_trans *trans,
458 			       struct btree_insert_entry *i,
459 			       unsigned flags)
460 {
461 	verify_update_old_key(trans, i);
462 
463 	if (unlikely(flags & BTREE_TRIGGER_norun))
464 		return 0;
465 
466 	struct bkey_s_c old = { &i->old_k, i->old_v };
467 	struct bkey_i *new = i->k;
468 	const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
469 	const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
470 
471 	if (old_ops->trigger == new_ops->trigger)
472 		return bch2_key_trigger(trans, i->btree_id, i->level,
473 				old, bkey_i_to_s(new),
474 				BTREE_TRIGGER_insert|BTREE_TRIGGER_overwrite|flags);
475 	else
476 		return bch2_key_trigger_new(trans, i->btree_id, i->level,
477 				bkey_i_to_s(new), flags) ?:
478 		       bch2_key_trigger_old(trans, i->btree_id, i->level,
479 				old, flags);
480 }
481 
482 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
483 				 bool overwrite)
484 {
485 	verify_update_old_key(trans, i);
486 
487 	if ((i->flags & BTREE_TRIGGER_norun) ||
488 	    !btree_node_type_has_trans_triggers(i->bkey_type))
489 		return 0;
490 
491 	/*
492 	 * Transactional triggers create new btree_insert_entries, so we can't
493 	 * pass them a pointer to a btree_insert_entry, that memory is going to
494 	 * move:
495 	 */
496 	struct bkey old_k = i->old_k;
497 	struct bkey_s_c old = { &old_k, i->old_v };
498 	const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
499 	const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
500 	unsigned flags = i->flags|BTREE_TRIGGER_transactional;
501 
502 	if (!i->insert_trigger_run &&
503 	    !i->overwrite_trigger_run &&
504 	    old_ops->trigger == new_ops->trigger) {
505 		i->overwrite_trigger_run = true;
506 		i->insert_trigger_run = true;
507 		return bch2_key_trigger(trans, i->btree_id, i->level, old, bkey_i_to_s(i->k),
508 					BTREE_TRIGGER_insert|
509 					BTREE_TRIGGER_overwrite|flags) ?: 1;
510 	} else if (overwrite && !i->overwrite_trigger_run) {
511 		i->overwrite_trigger_run = true;
512 		return bch2_key_trigger_old(trans, i->btree_id, i->level, old, flags) ?: 1;
513 	} else if (!overwrite && !i->insert_trigger_run) {
514 		i->insert_trigger_run = true;
515 		return bch2_key_trigger_new(trans, i->btree_id, i->level, bkey_i_to_s(i->k), flags) ?: 1;
516 	} else {
517 		return 0;
518 	}
519 }
520 
521 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
522 			      unsigned btree_id_start)
523 {
524 	for (int overwrite = 1; overwrite >= 0; --overwrite) {
525 		bool trans_trigger_run;
526 
527 		/*
528 		 * Running triggers will append more updates to the list of updates as
529 		 * we're walking it:
530 		 */
531 		do {
532 			trans_trigger_run = false;
533 
534 			for (unsigned i = btree_id_start;
535 			     i < trans->nr_updates && trans->updates[i].btree_id <= btree_id;
536 			     i++) {
537 				if (trans->updates[i].btree_id != btree_id)
538 					continue;
539 
540 				int ret = run_one_trans_trigger(trans, trans->updates + i, overwrite);
541 				if (ret < 0)
542 					return ret;
543 				if (ret)
544 					trans_trigger_run = true;
545 			}
546 		} while (trans_trigger_run);
547 	}
548 
549 	return 0;
550 }
551 
552 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
553 {
554 	unsigned btree_id = 0, btree_id_start = 0;
555 	int ret = 0;
556 
557 	/*
558 	 *
559 	 * For a given btree, this algorithm runs insert triggers before
560 	 * overwrite triggers: this is so that when extents are being moved
561 	 * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
562 	 * they are re-added.
563 	 */
564 	for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
565 		if (btree_id == BTREE_ID_alloc)
566 			continue;
567 
568 		while (btree_id_start < trans->nr_updates &&
569 		       trans->updates[btree_id_start].btree_id < btree_id)
570 			btree_id_start++;
571 
572 		ret = run_btree_triggers(trans, btree_id, btree_id_start);
573 		if (ret)
574 			return ret;
575 	}
576 
577 	for (unsigned idx = 0; idx < trans->nr_updates; idx++) {
578 		struct btree_insert_entry *i = trans->updates + idx;
579 
580 		if (i->btree_id > BTREE_ID_alloc)
581 			break;
582 		if (i->btree_id == BTREE_ID_alloc) {
583 			ret = run_btree_triggers(trans, BTREE_ID_alloc, idx);
584 			if (ret)
585 				return ret;
586 			break;
587 		}
588 	}
589 
590 #ifdef CONFIG_BCACHEFS_DEBUG
591 	trans_for_each_update(trans, i)
592 		BUG_ON(!(i->flags & BTREE_TRIGGER_norun) &&
593 		       btree_node_type_has_trans_triggers(i->bkey_type) &&
594 		       (!i->insert_trigger_run || !i->overwrite_trigger_run));
595 #endif
596 	return 0;
597 }
598 
599 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
600 {
601 	trans_for_each_update(trans, i)
602 		if (btree_node_type_has_triggers(i->bkey_type) &&
603 		    gc_visited(trans->c, gc_pos_btree(i->btree_id, i->level, i->k->k.p))) {
604 			int ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_gc);
605 			if (ret)
606 				return ret;
607 		}
608 
609 	return 0;
610 }
611 
612 static struct bversion journal_pos_to_bversion(struct journal_res *res, unsigned offset)
613 {
614 	return (struct bversion) {
615 		.hi = res->seq >> 32,
616 		.lo = (res->seq << 32) | (res->offset + offset),
617 	};
618 }
619 
620 static inline int
621 bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
622 			       struct btree_insert_entry **stopped_at,
623 			       unsigned long trace_ip)
624 {
625 	struct bch_fs *c = trans->c;
626 	struct btree_trans_commit_hook *h;
627 	unsigned u64s = 0;
628 	int ret = 0;
629 
630 	bch2_trans_verify_not_unlocked(trans);
631 	bch2_trans_verify_not_in_restart(trans);
632 
633 	if (race_fault()) {
634 		trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
635 		return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
636 	}
637 
638 	/*
639 	 * Check if the insert will fit in the leaf node with the write lock
640 	 * held, otherwise another thread could write the node changing the
641 	 * amount of space available:
642 	 */
643 
644 	prefetch(&trans->c->journal.flags);
645 
646 	trans_for_each_update(trans, i) {
647 		/* Multiple inserts might go to same leaf: */
648 		if (!same_leaf_as_prev(trans, i))
649 			u64s = 0;
650 
651 		u64s += i->k->k.u64s;
652 		ret = !i->cached
653 			? btree_key_can_insert(trans, insert_l(trans, i)->b, u64s)
654 			: btree_key_can_insert_cached(trans, flags, trans->paths + i->path, u64s);
655 		if (ret) {
656 			*stopped_at = i;
657 			return ret;
658 		}
659 
660 		i->k->k.needs_whiteout = false;
661 	}
662 
663 	/*
664 	 * Don't get journal reservation until after we know insert will
665 	 * succeed:
666 	 */
667 	if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res))) {
668 		ret = bch2_trans_journal_res_get(trans,
669 				(flags & BCH_WATERMARK_MASK)|
670 				JOURNAL_RES_GET_NONBLOCK);
671 		if (ret)
672 			return ret;
673 
674 		if (unlikely(trans->journal_transaction_names))
675 			journal_transaction_name(trans);
676 	}
677 
678 	/*
679 	 * Not allowed to fail after we've gotten our journal reservation - we
680 	 * have to use it:
681 	 */
682 
683 	if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
684 	    !(flags & BCH_TRANS_COMMIT_no_journal_res)) {
685 		if (bch2_journal_seq_verify)
686 			trans_for_each_update(trans, i)
687 				i->k->k.version.lo = trans->journal_res.seq;
688 		else if (bch2_inject_invalid_keys)
689 			trans_for_each_update(trans, i)
690 				i->k->k.version = MAX_VERSION;
691 	}
692 
693 	h = trans->hooks;
694 	while (h) {
695 		ret = h->fn(trans, h);
696 		if (ret)
697 			return ret;
698 		h = h->next;
699 	}
700 
701 	struct jset_entry *entry = trans->journal_entries;
702 
703 	if (likely(!(flags & BCH_TRANS_COMMIT_skip_accounting_apply))) {
704 		percpu_down_read(&c->mark_lock);
705 
706 		for (entry = trans->journal_entries;
707 		     entry != (void *) ((u64 *) trans->journal_entries + trans->journal_entries_u64s);
708 		     entry = vstruct_next(entry))
709 			if (jset_entry_is_key(entry) && entry->start->k.type == KEY_TYPE_accounting) {
710 				struct bkey_i_accounting *a = bkey_i_to_accounting(entry->start);
711 
712 				a->k.version = journal_pos_to_bversion(&trans->journal_res,
713 								(u64 *) entry - (u64 *) trans->journal_entries);
714 				BUG_ON(bversion_zero(a->k.version));
715 				ret = bch2_accounting_mem_mod_locked(trans, accounting_i_to_s_c(a), false, false);
716 				if (ret)
717 					goto revert_fs_usage;
718 			}
719 		percpu_up_read(&c->mark_lock);
720 
721 		/* XXX: we only want to run this if deltas are nonzero */
722 		bch2_trans_account_disk_usage_change(trans);
723 	}
724 
725 	trans_for_each_update(trans, i)
726 		if (btree_node_type_has_atomic_triggers(i->bkey_type)) {
727 			ret = run_one_mem_trigger(trans, i, BTREE_TRIGGER_atomic|i->flags);
728 			if (ret)
729 				goto fatal_err;
730 		}
731 
732 	if (unlikely(c->gc_pos.phase)) {
733 		ret = bch2_trans_commit_run_gc_triggers(trans);
734 		if  (ret)
735 			goto fatal_err;
736 	}
737 
738 	if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res))) {
739 		struct journal *j = &c->journal;
740 		struct jset_entry *entry;
741 
742 		trans_for_each_update(trans, i) {
743 			if (i->key_cache_already_flushed)
744 				continue;
745 
746 			if (i->flags & BTREE_UPDATE_nojournal)
747 				continue;
748 
749 			verify_update_old_key(trans, i);
750 
751 			if (trans->journal_transaction_names) {
752 				entry = bch2_journal_add_entry(j, &trans->journal_res,
753 						       BCH_JSET_ENTRY_overwrite,
754 						       i->btree_id, i->level,
755 						       i->old_k.u64s);
756 				bkey_reassemble((struct bkey_i *) entry->start,
757 						(struct bkey_s_c) { &i->old_k, i->old_v });
758 			}
759 
760 			entry = bch2_journal_add_entry(j, &trans->journal_res,
761 					       BCH_JSET_ENTRY_btree_keys,
762 					       i->btree_id, i->level,
763 					       i->k->k.u64s);
764 			bkey_copy((struct bkey_i *) entry->start, i->k);
765 		}
766 
767 		memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
768 				  trans->journal_entries,
769 				  trans->journal_entries_u64s);
770 
771 		trans->journal_res.offset	+= trans->journal_entries_u64s;
772 		trans->journal_res.u64s		-= trans->journal_entries_u64s;
773 
774 		if (trans->journal_seq)
775 			*trans->journal_seq = trans->journal_res.seq;
776 	}
777 
778 	trans_for_each_update(trans, i) {
779 		struct btree_path *path = trans->paths + i->path;
780 
781 		if (!i->cached)
782 			bch2_btree_insert_key_leaf(trans, path, i->k, trans->journal_res.seq);
783 		else if (!i->key_cache_already_flushed)
784 			bch2_btree_insert_key_cached(trans, flags, i);
785 		else
786 			bch2_btree_key_cache_drop(trans, path);
787 	}
788 
789 	return 0;
790 fatal_err:
791 	bch2_fs_fatal_error(c, "fatal error in transaction commit: %s", bch2_err_str(ret));
792 	percpu_down_read(&c->mark_lock);
793 revert_fs_usage:
794 	for (struct jset_entry *entry2 = trans->journal_entries;
795 	     entry2 != entry;
796 	     entry2 = vstruct_next(entry2))
797 		if (jset_entry_is_key(entry2) && entry2->start->k.type == KEY_TYPE_accounting) {
798 			struct bkey_s_accounting a = bkey_i_to_s_accounting(entry2->start);
799 
800 			bch2_accounting_neg(a);
801 			bch2_accounting_mem_mod_locked(trans, a.c, false, false);
802 			bch2_accounting_neg(a);
803 		}
804 	percpu_up_read(&c->mark_lock);
805 	return ret;
806 }
807 
808 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
809 {
810 	/*
811 	 * Accounting keys aren't deduped in the journal: we have to compare
812 	 * each individual update against what's in the btree to see if it has
813 	 * been applied yet, and accounting updates also don't overwrite,
814 	 * they're deltas that accumulate.
815 	 */
816 	trans_for_each_update(trans, i)
817 		if (i->k->k.type != KEY_TYPE_accounting)
818 			bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
819 }
820 
821 static int bch2_trans_commit_journal_pin_flush(struct journal *j,
822 				struct journal_entry_pin *_pin, u64 seq)
823 {
824 	return 0;
825 }
826 
827 /*
828  * Get journal reservation, take write locks, and attempt to do btree update(s):
829  */
830 static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags,
831 				       struct btree_insert_entry **stopped_at,
832 				       unsigned long trace_ip)
833 {
834 	struct bch_fs *c = trans->c;
835 	int ret = 0, u64s_delta = 0;
836 
837 	for (unsigned idx = 0; idx < trans->nr_updates; idx++) {
838 		struct btree_insert_entry *i = trans->updates + idx;
839 		if (i->cached)
840 			continue;
841 
842 		u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
843 		u64s_delta -= i->old_btree_u64s;
844 
845 		if (!same_leaf_as_next(trans, i)) {
846 			if (u64s_delta <= 0) {
847 				ret = bch2_foreground_maybe_merge(trans, i->path,
848 							i->level, flags);
849 				if (unlikely(ret))
850 					return ret;
851 			}
852 
853 			u64s_delta = 0;
854 		}
855 	}
856 
857 	ret = bch2_trans_lock_write(trans);
858 	if (unlikely(ret))
859 		return ret;
860 
861 	ret = bch2_trans_commit_write_locked(trans, flags, stopped_at, trace_ip);
862 
863 	if (!ret && unlikely(trans->journal_replay_not_finished))
864 		bch2_drop_overwrites_from_journal(trans);
865 
866 	bch2_trans_unlock_write(trans);
867 
868 	if (!ret && trans->journal_pin)
869 		bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
870 				     trans->journal_pin,
871 				     bch2_trans_commit_journal_pin_flush);
872 
873 	/*
874 	 * Drop journal reservation after dropping write locks, since dropping
875 	 * the journal reservation may kick off a journal write:
876 	 */
877 	if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res)))
878 		bch2_journal_res_put(&c->journal, &trans->journal_res);
879 
880 	return ret;
881 }
882 
883 static int journal_reclaim_wait_done(struct bch_fs *c)
884 {
885 	int ret = bch2_journal_error(&c->journal) ?:
886 		bch2_btree_key_cache_wait_done(c);
887 
888 	if (!ret)
889 		journal_reclaim_kick(&c->journal);
890 	return ret;
891 }
892 
893 static noinline
894 int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
895 			    struct btree_insert_entry *i,
896 			    int ret, unsigned long trace_ip)
897 {
898 	struct bch_fs *c = trans->c;
899 	enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
900 
901 	switch (ret) {
902 	case -BCH_ERR_btree_insert_btree_node_full:
903 		ret = bch2_btree_split_leaf(trans, i->path, flags);
904 		if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
905 			trace_and_count(c, trans_restart_btree_node_split, trans,
906 					trace_ip, trans->paths + i->path);
907 		break;
908 	case -BCH_ERR_btree_insert_need_mark_replicas:
909 		ret = drop_locks_do(trans,
910 			bch2_accounting_update_sb(trans));
911 		break;
912 	case -BCH_ERR_journal_res_get_blocked:
913 		/*
914 		 * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
915 		 * flag
916 		 */
917 		if ((flags & BCH_TRANS_COMMIT_journal_reclaim) &&
918 		    watermark < BCH_WATERMARK_reclaim) {
919 			ret = -BCH_ERR_journal_reclaim_would_deadlock;
920 			break;
921 		}
922 
923 		ret = drop_locks_do(trans,
924 			bch2_trans_journal_res_get(trans,
925 					(flags & BCH_WATERMARK_MASK)|
926 					JOURNAL_RES_GET_CHECK));
927 		break;
928 	case -BCH_ERR_btree_insert_need_journal_reclaim:
929 		bch2_trans_unlock(trans);
930 
931 		trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
932 		track_event_change(&c->times[BCH_TIME_blocked_key_cache_flush], true);
933 
934 		wait_event_freezable(c->journal.reclaim_wait,
935 				     (ret = journal_reclaim_wait_done(c)));
936 
937 		track_event_change(&c->times[BCH_TIME_blocked_key_cache_flush], false);
938 
939 		if (ret < 0)
940 			break;
941 
942 		ret = bch2_trans_relock(trans);
943 		break;
944 	default:
945 		BUG_ON(ret >= 0);
946 		break;
947 	}
948 
949 	BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
950 
951 	bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
952 				(flags & BCH_TRANS_COMMIT_no_enospc), c,
953 		"%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
954 
955 	return ret;
956 }
957 
958 static noinline int
959 bch2_trans_commit_get_rw_cold(struct btree_trans *trans, unsigned flags)
960 {
961 	struct bch_fs *c = trans->c;
962 	int ret;
963 
964 	if (likely(!(flags & BCH_TRANS_COMMIT_lazy_rw)) ||
965 	    test_bit(BCH_FS_started, &c->flags))
966 		return -BCH_ERR_erofs_trans_commit;
967 
968 	ret = drop_locks_do(trans, bch2_fs_read_write_early(c));
969 	if (ret)
970 		return ret;
971 
972 	bch2_write_ref_get(c, BCH_WRITE_REF_trans);
973 	return 0;
974 }
975 
976 /*
977  * This is for updates done in the early part of fsck - btree_gc - before we've
978  * gone RW. we only add the new key to the list of keys for journal replay to
979  * do.
980  */
981 static noinline int
982 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
983 {
984 	struct bch_fs *c = trans->c;
985 
986 	trans_for_each_update(trans, i) {
987 		int ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
988 		if (ret)
989 			return ret;
990 	}
991 
992 	for (struct jset_entry *i = trans->journal_entries;
993 	     i != (void *) ((u64 *) trans->journal_entries + trans->journal_entries_u64s);
994 	     i = vstruct_next(i))
995 		if (i->type == BCH_JSET_ENTRY_btree_keys ||
996 		    i->type == BCH_JSET_ENTRY_write_buffer_keys) {
997 			int ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->start);
998 			if (ret)
999 				return ret;
1000 		}
1001 
1002 	return 0;
1003 }
1004 
1005 int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
1006 {
1007 	struct btree_insert_entry *errored_at = NULL;
1008 	struct bch_fs *c = trans->c;
1009 	int ret = 0;
1010 
1011 	bch2_trans_verify_not_unlocked(trans);
1012 	bch2_trans_verify_not_in_restart(trans);
1013 
1014 	if (!trans->nr_updates &&
1015 	    !trans->journal_entries_u64s)
1016 		goto out_reset;
1017 
1018 	ret = bch2_trans_commit_run_triggers(trans);
1019 	if (ret)
1020 		goto out_reset;
1021 
1022 	trans_for_each_update(trans, i) {
1023 		enum bch_validate_flags invalid_flags = 0;
1024 
1025 		if (!(flags & BCH_TRANS_COMMIT_no_journal_res))
1026 			invalid_flags |= BCH_VALIDATE_write|BCH_VALIDATE_commit;
1027 
1028 		ret = bch2_bkey_validate(c, bkey_i_to_s_c(i->k),
1029 					 i->bkey_type, invalid_flags);
1030 		if (unlikely(ret)){
1031 			bch2_trans_inconsistent(trans, "invalid bkey on insert from %s -> %ps\n",
1032 						trans->fn, (void *) i->ip_allocated);
1033 			return ret;
1034 		}
1035 		btree_insert_entry_checks(trans, i);
1036 	}
1037 
1038 	for (struct jset_entry *i = trans->journal_entries;
1039 	     i != (void *) ((u64 *) trans->journal_entries + trans->journal_entries_u64s);
1040 	     i = vstruct_next(i)) {
1041 		enum bch_validate_flags invalid_flags = 0;
1042 
1043 		if (!(flags & BCH_TRANS_COMMIT_no_journal_res))
1044 			invalid_flags |= BCH_VALIDATE_write|BCH_VALIDATE_commit;
1045 
1046 		ret = bch2_journal_entry_validate(c, NULL, i,
1047 						  bcachefs_metadata_version_current,
1048 						  CPU_BIG_ENDIAN, invalid_flags);
1049 		if (unlikely(ret)) {
1050 			bch2_trans_inconsistent(trans, "invalid journal entry on insert from %s\n",
1051 						trans->fn);
1052 			return ret;
1053 		}
1054 	}
1055 
1056 	if (unlikely(!test_bit(BCH_FS_may_go_rw, &c->flags))) {
1057 		ret = do_bch2_trans_commit_to_journal_replay(trans);
1058 		goto out_reset;
1059 	}
1060 
1061 	if (!(flags & BCH_TRANS_COMMIT_no_check_rw) &&
1062 	    unlikely(!bch2_write_ref_tryget(c, BCH_WRITE_REF_trans))) {
1063 		ret = bch2_trans_commit_get_rw_cold(trans, flags);
1064 		if (ret)
1065 			goto out_reset;
1066 	}
1067 
1068 	EBUG_ON(test_bit(BCH_FS_clean_shutdown, &c->flags));
1069 
1070 	trans->journal_u64s		= trans->journal_entries_u64s;
1071 	trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
1072 	if (trans->journal_transaction_names)
1073 		trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1074 
1075 	trans_for_each_update(trans, i) {
1076 		struct btree_path *path = trans->paths + i->path;
1077 
1078 		EBUG_ON(!path->should_be_locked);
1079 
1080 		ret = bch2_btree_path_upgrade(trans, path, i->level + 1);
1081 		if (unlikely(ret))
1082 			goto out;
1083 
1084 		EBUG_ON(!btree_node_intent_locked(path, i->level));
1085 
1086 		if (i->key_cache_already_flushed)
1087 			continue;
1088 
1089 		if (i->flags & BTREE_UPDATE_nojournal)
1090 			continue;
1091 
1092 		/* we're going to journal the key being updated: */
1093 		trans->journal_u64s += jset_u64s(i->k->k.u64s);
1094 
1095 		/* and we're also going to log the overwrite: */
1096 		if (trans->journal_transaction_names)
1097 			trans->journal_u64s += jset_u64s(i->old_k.u64s);
1098 	}
1099 
1100 	if (trans->extra_disk_res) {
1101 		ret = bch2_disk_reservation_add(c, trans->disk_res,
1102 				trans->extra_disk_res,
1103 				(flags & BCH_TRANS_COMMIT_no_enospc)
1104 				? BCH_DISK_RESERVATION_NOFAIL : 0);
1105 		if (ret)
1106 			goto err;
1107 	}
1108 retry:
1109 	errored_at = NULL;
1110 	bch2_trans_verify_not_unlocked(trans);
1111 	bch2_trans_verify_not_in_restart(trans);
1112 	if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res)))
1113 		memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1114 	memset(&trans->fs_usage_delta, 0, sizeof(trans->fs_usage_delta));
1115 
1116 	ret = do_bch2_trans_commit(trans, flags, &errored_at, _RET_IP_);
1117 
1118 	/* make sure we didn't drop or screw up locks: */
1119 	bch2_trans_verify_locks(trans);
1120 
1121 	if (ret)
1122 		goto err;
1123 
1124 	trace_and_count(c, transaction_commit, trans, _RET_IP_);
1125 out:
1126 	if (likely(!(flags & BCH_TRANS_COMMIT_no_check_rw)))
1127 		bch2_write_ref_put(c, BCH_WRITE_REF_trans);
1128 out_reset:
1129 	if (!ret)
1130 		bch2_trans_downgrade(trans);
1131 	bch2_trans_reset_updates(trans);
1132 
1133 	return ret;
1134 err:
1135 	ret = bch2_trans_commit_error(trans, flags, errored_at, ret, _RET_IP_);
1136 	if (ret)
1137 		goto out;
1138 
1139 	/*
1140 	 * We might have done another transaction commit in the error path -
1141 	 * i.e. btree write buffer flush - which will have made use of
1142 	 * trans->journal_res, but with BCH_TRANS_COMMIT_no_journal_res that is
1143 	 * how the journal sequence number to pin is passed in - so we must
1144 	 * restart:
1145 	 */
1146 	if (flags & BCH_TRANS_COMMIT_no_journal_res) {
1147 		ret = -BCH_ERR_transaction_restart_nested;
1148 		goto out;
1149 	}
1150 
1151 	goto retry;
1152 }
1153