xref: /linux/fs/bcachefs/data_update.c (revision fdd5ecbbff751c3b9061d8ebb08e5c96119915b4)
1 // SPDX-License-Identifier: GPL-2.0
2 
3 #include "bcachefs.h"
4 #include "alloc_foreground.h"
5 #include "bkey_buf.h"
6 #include "btree_update.h"
7 #include "buckets.h"
8 #include "data_update.h"
9 #include "ec.h"
10 #include "error.h"
11 #include "extents.h"
12 #include "io_write.h"
13 #include "keylist.h"
14 #include "move.h"
15 #include "nocow_locking.h"
16 #include "rebalance.h"
17 #include "snapshot.h"
18 #include "subvolume.h"
19 #include "trace.h"
20 
21 static void trace_move_extent_finish2(struct bch_fs *c, struct bkey_s_c k)
22 {
23 	if (trace_move_extent_finish_enabled()) {
24 		struct printbuf buf = PRINTBUF;
25 
26 		bch2_bkey_val_to_text(&buf, c, k);
27 		trace_move_extent_finish(c, buf.buf);
28 		printbuf_exit(&buf);
29 	}
30 }
31 
32 static void trace_move_extent_fail2(struct data_update *m,
33 			 struct bkey_s_c new,
34 			 struct bkey_s_c wrote,
35 			 struct bkey_i *insert,
36 			 const char *msg)
37 {
38 	struct bch_fs *c = m->op.c;
39 	struct bkey_s_c old = bkey_i_to_s_c(m->k.k);
40 	const union bch_extent_entry *entry;
41 	struct bch_extent_ptr *ptr;
42 	struct extent_ptr_decoded p;
43 	struct printbuf buf = PRINTBUF;
44 	unsigned i, rewrites_found = 0;
45 
46 	if (!trace_move_extent_fail_enabled())
47 		return;
48 
49 	prt_str(&buf, msg);
50 
51 	if (insert) {
52 		i = 0;
53 		bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry) {
54 			if (((1U << i) & m->data_opts.rewrite_ptrs) &&
55 			    (ptr = bch2_extent_has_ptr(old, p, bkey_i_to_s(insert))) &&
56 			    !ptr->cached)
57 				rewrites_found |= 1U << i;
58 			i++;
59 		}
60 	}
61 
62 	prt_printf(&buf, "\nrewrite ptrs:   %u%u%u%u",
63 		   (m->data_opts.rewrite_ptrs & (1 << 0)) != 0,
64 		   (m->data_opts.rewrite_ptrs & (1 << 1)) != 0,
65 		   (m->data_opts.rewrite_ptrs & (1 << 2)) != 0,
66 		   (m->data_opts.rewrite_ptrs & (1 << 3)) != 0);
67 
68 	prt_printf(&buf, "\nrewrites found: %u%u%u%u",
69 		   (rewrites_found & (1 << 0)) != 0,
70 		   (rewrites_found & (1 << 1)) != 0,
71 		   (rewrites_found & (1 << 2)) != 0,
72 		   (rewrites_found & (1 << 3)) != 0);
73 
74 	prt_str(&buf, "\nold:    ");
75 	bch2_bkey_val_to_text(&buf, c, old);
76 
77 	prt_str(&buf, "\nnew:    ");
78 	bch2_bkey_val_to_text(&buf, c, new);
79 
80 	prt_str(&buf, "\nwrote:  ");
81 	bch2_bkey_val_to_text(&buf, c, wrote);
82 
83 	if (insert) {
84 		prt_str(&buf, "\ninsert: ");
85 		bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
86 	}
87 
88 	trace_move_extent_fail(c, buf.buf);
89 	printbuf_exit(&buf);
90 }
91 
92 static int __bch2_data_update_index_update(struct btree_trans *trans,
93 					   struct bch_write_op *op)
94 {
95 	struct bch_fs *c = op->c;
96 	struct btree_iter iter;
97 	struct data_update *m =
98 		container_of(op, struct data_update, op);
99 	struct keylist *keys = &op->insert_keys;
100 	struct bkey_buf _new, _insert;
101 	int ret = 0;
102 
103 	bch2_bkey_buf_init(&_new);
104 	bch2_bkey_buf_init(&_insert);
105 	bch2_bkey_buf_realloc(&_insert, c, U8_MAX);
106 
107 	bch2_trans_iter_init(trans, &iter, m->btree_id,
108 			     bkey_start_pos(&bch2_keylist_front(keys)->k),
109 			     BTREE_ITER_slots|BTREE_ITER_intent);
110 
111 	while (1) {
112 		struct bkey_s_c k;
113 		struct bkey_s_c old = bkey_i_to_s_c(m->k.k);
114 		struct bkey_i *insert = NULL;
115 		struct bkey_i_extent *new;
116 		const union bch_extent_entry *entry_c;
117 		union bch_extent_entry *entry;
118 		struct extent_ptr_decoded p;
119 		struct bch_extent_ptr *ptr;
120 		const struct bch_extent_ptr *ptr_c;
121 		struct bpos next_pos;
122 		bool should_check_enospc;
123 		s64 i_sectors_delta = 0, disk_sectors_delta = 0;
124 		unsigned rewrites_found = 0, durability, i;
125 
126 		bch2_trans_begin(trans);
127 
128 		k = bch2_btree_iter_peek_slot(&iter);
129 		ret = bkey_err(k);
130 		if (ret)
131 			goto err;
132 
133 		new = bkey_i_to_extent(bch2_keylist_front(keys));
134 
135 		if (!bch2_extents_match(k, old)) {
136 			trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i),
137 						NULL, "no match:");
138 			goto nowork;
139 		}
140 
141 		bkey_reassemble(_insert.k, k);
142 		insert = _insert.k;
143 
144 		bch2_bkey_buf_copy(&_new, c, bch2_keylist_front(keys));
145 		new = bkey_i_to_extent(_new.k);
146 		bch2_cut_front(iter.pos, &new->k_i);
147 
148 		bch2_cut_front(iter.pos,	insert);
149 		bch2_cut_back(new->k.p,		insert);
150 		bch2_cut_back(insert->k.p,	&new->k_i);
151 
152 		/*
153 		 * @old: extent that we read from
154 		 * @insert: key that we're going to update, initialized from
155 		 * extent currently in btree - same as @old unless we raced with
156 		 * other updates
157 		 * @new: extent with new pointers that we'll be adding to @insert
158 		 *
159 		 * Fist, drop rewrite_ptrs from @new:
160 		 */
161 		i = 0;
162 		bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry_c) {
163 			if (((1U << i) & m->data_opts.rewrite_ptrs) &&
164 			    (ptr = bch2_extent_has_ptr(old, p, bkey_i_to_s(insert))) &&
165 			    !ptr->cached) {
166 				bch2_extent_ptr_set_cached(bkey_i_to_s(insert), ptr);
167 				rewrites_found |= 1U << i;
168 			}
169 			i++;
170 		}
171 
172 		if (m->data_opts.rewrite_ptrs &&
173 		    !rewrites_found &&
174 		    bch2_bkey_durability(c, k) >= m->op.opts.data_replicas) {
175 			trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i), insert, "no rewrites found:");
176 			goto nowork;
177 		}
178 
179 		/*
180 		 * A replica that we just wrote might conflict with a replica
181 		 * that we want to keep, due to racing with another move:
182 		 */
183 restart_drop_conflicting_replicas:
184 		extent_for_each_ptr(extent_i_to_s(new), ptr)
185 			if ((ptr_c = bch2_bkey_has_device_c(bkey_i_to_s_c(insert), ptr->dev)) &&
186 			    !ptr_c->cached) {
187 				bch2_bkey_drop_ptr_noerror(bkey_i_to_s(&new->k_i), ptr);
188 				goto restart_drop_conflicting_replicas;
189 			}
190 
191 		if (!bkey_val_u64s(&new->k)) {
192 			trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i), insert, "new replicas conflicted:");
193 			goto nowork;
194 		}
195 
196 		/* Now, drop pointers that conflict with what we just wrote: */
197 		extent_for_each_ptr_decode(extent_i_to_s(new), p, entry)
198 			if ((ptr = bch2_bkey_has_device(bkey_i_to_s(insert), p.ptr.dev)))
199 				bch2_bkey_drop_ptr_noerror(bkey_i_to_s(insert), ptr);
200 
201 		durability = bch2_bkey_durability(c, bkey_i_to_s_c(insert)) +
202 			bch2_bkey_durability(c, bkey_i_to_s_c(&new->k_i));
203 
204 		/* Now, drop excess replicas: */
205 		rcu_read_lock();
206 restart_drop_extra_replicas:
207 		bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs(bkey_i_to_s(insert)), p, entry) {
208 			unsigned ptr_durability = bch2_extent_ptr_durability(c, &p);
209 
210 			if (!p.ptr.cached &&
211 			    durability - ptr_durability >= m->op.opts.data_replicas) {
212 				durability -= ptr_durability;
213 
214 				bch2_extent_ptr_set_cached(bkey_i_to_s(insert), &entry->ptr);
215 				goto restart_drop_extra_replicas;
216 			}
217 		}
218 		rcu_read_unlock();
219 
220 		/* Finally, add the pointers we just wrote: */
221 		extent_for_each_ptr_decode(extent_i_to_s(new), p, entry)
222 			bch2_extent_ptr_decoded_append(insert, &p);
223 
224 		bch2_bkey_narrow_crcs(insert, (struct bch_extent_crc_unpacked) { 0 });
225 		bch2_extent_normalize(c, bkey_i_to_s(insert));
226 
227 		ret = bch2_sum_sector_overwrites(trans, &iter, insert,
228 						 &should_check_enospc,
229 						 &i_sectors_delta,
230 						 &disk_sectors_delta);
231 		if (ret)
232 			goto err;
233 
234 		if (disk_sectors_delta > (s64) op->res.sectors) {
235 			ret = bch2_disk_reservation_add(c, &op->res,
236 						disk_sectors_delta - op->res.sectors,
237 						!should_check_enospc
238 						? BCH_DISK_RESERVATION_NOFAIL : 0);
239 			if (ret)
240 				goto out;
241 		}
242 
243 		next_pos = insert->k.p;
244 
245 		/*
246 		 * Check for nonce offset inconsistency:
247 		 * This is debug code - we've been seeing this bug rarely, and
248 		 * it's been hard to reproduce, so this should give us some more
249 		 * information when it does occur:
250 		 */
251 		struct printbuf err = PRINTBUF;
252 		int invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(insert), __btree_node_type(0, m->btree_id), 0, &err);
253 		printbuf_exit(&err);
254 
255 		if (invalid) {
256 			struct printbuf buf = PRINTBUF;
257 
258 			prt_str(&buf, "about to insert invalid key in data update path");
259 			prt_str(&buf, "\nold: ");
260 			bch2_bkey_val_to_text(&buf, c, old);
261 			prt_str(&buf, "\nk:   ");
262 			bch2_bkey_val_to_text(&buf, c, k);
263 			prt_str(&buf, "\nnew: ");
264 			bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
265 
266 			bch2_print_string_as_lines(KERN_ERR, buf.buf);
267 			printbuf_exit(&buf);
268 
269 			bch2_fatal_error(c);
270 			goto out;
271 		}
272 
273 		if (trace_data_update_enabled()) {
274 			struct printbuf buf = PRINTBUF;
275 
276 			prt_str(&buf, "\nold: ");
277 			bch2_bkey_val_to_text(&buf, c, old);
278 			prt_str(&buf, "\nk:   ");
279 			bch2_bkey_val_to_text(&buf, c, k);
280 			prt_str(&buf, "\nnew: ");
281 			bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
282 
283 			trace_data_update(c, buf.buf);
284 			printbuf_exit(&buf);
285 		}
286 
287 		ret =   bch2_insert_snapshot_whiteouts(trans, m->btree_id,
288 						k.k->p, bkey_start_pos(&insert->k)) ?:
289 			bch2_insert_snapshot_whiteouts(trans, m->btree_id,
290 						k.k->p, insert->k.p) ?:
291 			bch2_bkey_set_needs_rebalance(c, insert, &op->opts) ?:
292 			bch2_trans_update(trans, &iter, insert,
293 				BTREE_UPDATE_internal_snapshot_node) ?:
294 			bch2_trans_commit(trans, &op->res,
295 				NULL,
296 				BCH_TRANS_COMMIT_no_check_rw|
297 				BCH_TRANS_COMMIT_no_enospc|
298 				m->data_opts.btree_insert_flags);
299 		if (!ret) {
300 			bch2_btree_iter_set_pos(&iter, next_pos);
301 
302 			this_cpu_add(c->counters[BCH_COUNTER_move_extent_finish], new->k.size);
303 			trace_move_extent_finish2(c, bkey_i_to_s_c(&new->k_i));
304 		}
305 err:
306 		if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
307 			ret = 0;
308 		if (ret)
309 			break;
310 next:
311 		while (bkey_ge(iter.pos, bch2_keylist_front(keys)->k.p)) {
312 			bch2_keylist_pop_front(keys);
313 			if (bch2_keylist_empty(keys))
314 				goto out;
315 		}
316 		continue;
317 nowork:
318 		if (m->stats) {
319 			BUG_ON(k.k->p.offset <= iter.pos.offset);
320 			atomic64_inc(&m->stats->keys_raced);
321 			atomic64_add(k.k->p.offset - iter.pos.offset,
322 				     &m->stats->sectors_raced);
323 		}
324 
325 		count_event(c, move_extent_fail);
326 
327 		bch2_btree_iter_advance(&iter);
328 		goto next;
329 	}
330 out:
331 	bch2_trans_iter_exit(trans, &iter);
332 	bch2_bkey_buf_exit(&_insert, c);
333 	bch2_bkey_buf_exit(&_new, c);
334 	BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart));
335 	return ret;
336 }
337 
338 int bch2_data_update_index_update(struct bch_write_op *op)
339 {
340 	return bch2_trans_run(op->c, __bch2_data_update_index_update(trans, op));
341 }
342 
343 void bch2_data_update_read_done(struct data_update *m,
344 				struct bch_extent_crc_unpacked crc)
345 {
346 	/* write bio must own pages: */
347 	BUG_ON(!m->op.wbio.bio.bi_vcnt);
348 
349 	m->op.crc = crc;
350 	m->op.wbio.bio.bi_iter.bi_size = crc.compressed_size << 9;
351 
352 	closure_call(&m->op.cl, bch2_write, NULL, NULL);
353 }
354 
355 void bch2_data_update_exit(struct data_update *update)
356 {
357 	struct bch_fs *c = update->op.c;
358 	struct bkey_ptrs_c ptrs =
359 		bch2_bkey_ptrs_c(bkey_i_to_s_c(update->k.k));
360 
361 	bkey_for_each_ptr(ptrs, ptr) {
362 		struct bch_dev *ca = bch2_dev_have_ref(c, ptr->dev);
363 		if (c->opts.nocow_enabled)
364 			bch2_bucket_nocow_unlock(&c->nocow_locks,
365 						 PTR_BUCKET_POS(ca, ptr), 0);
366 		bch2_dev_put(ca);
367 	}
368 
369 	bch2_bkey_buf_exit(&update->k, c);
370 	bch2_disk_reservation_put(c, &update->op.res);
371 	bch2_bio_free_pages_pool(c, &update->op.wbio.bio);
372 }
373 
374 static void bch2_update_unwritten_extent(struct btree_trans *trans,
375 				  struct data_update *update)
376 {
377 	struct bch_fs *c = update->op.c;
378 	struct bio *bio = &update->op.wbio.bio;
379 	struct bkey_i_extent *e;
380 	struct write_point *wp;
381 	struct closure cl;
382 	struct btree_iter iter;
383 	struct bkey_s_c k;
384 	int ret;
385 
386 	closure_init_stack(&cl);
387 	bch2_keylist_init(&update->op.insert_keys, update->op.inline_keys);
388 
389 	while (bio_sectors(bio)) {
390 		unsigned sectors = bio_sectors(bio);
391 
392 		bch2_trans_begin(trans);
393 
394 		bch2_trans_iter_init(trans, &iter, update->btree_id, update->op.pos,
395 				     BTREE_ITER_slots);
396 		ret = lockrestart_do(trans, ({
397 			k = bch2_btree_iter_peek_slot(&iter);
398 			bkey_err(k);
399 		}));
400 		bch2_trans_iter_exit(trans, &iter);
401 
402 		if (ret || !bch2_extents_match(k, bkey_i_to_s_c(update->k.k)))
403 			break;
404 
405 		e = bkey_extent_init(update->op.insert_keys.top);
406 		e->k.p = update->op.pos;
407 
408 		ret = bch2_alloc_sectors_start_trans(trans,
409 				update->op.target,
410 				false,
411 				update->op.write_point,
412 				&update->op.devs_have,
413 				update->op.nr_replicas,
414 				update->op.nr_replicas,
415 				update->op.watermark,
416 				0, &cl, &wp);
417 		if (bch2_err_matches(ret, BCH_ERR_operation_blocked)) {
418 			bch2_trans_unlock(trans);
419 			closure_sync(&cl);
420 			continue;
421 		}
422 
423 		bch_err_fn_ratelimited(c, ret);
424 
425 		if (ret)
426 			return;
427 
428 		sectors = min(sectors, wp->sectors_free);
429 
430 		bch2_key_resize(&e->k, sectors);
431 
432 		bch2_open_bucket_get(c, wp, &update->op.open_buckets);
433 		bch2_alloc_sectors_append_ptrs(c, wp, &e->k_i, sectors, false);
434 		bch2_alloc_sectors_done(c, wp);
435 
436 		bio_advance(bio, sectors << 9);
437 		update->op.pos.offset += sectors;
438 
439 		extent_for_each_ptr(extent_i_to_s(e), ptr)
440 			ptr->unwritten = true;
441 		bch2_keylist_push(&update->op.insert_keys);
442 
443 		ret = __bch2_data_update_index_update(trans, &update->op);
444 
445 		bch2_open_buckets_put(c, &update->op.open_buckets);
446 
447 		if (ret)
448 			break;
449 	}
450 
451 	if (closure_nr_remaining(&cl) != 1) {
452 		bch2_trans_unlock(trans);
453 		closure_sync(&cl);
454 	}
455 }
456 
457 int bch2_extent_drop_ptrs(struct btree_trans *trans,
458 			  struct btree_iter *iter,
459 			  struct bkey_s_c k,
460 			  struct data_update_opts data_opts)
461 {
462 	struct bch_fs *c = trans->c;
463 	struct bkey_i *n;
464 	int ret;
465 
466 	n = bch2_bkey_make_mut_noupdate(trans, k);
467 	ret = PTR_ERR_OR_ZERO(n);
468 	if (ret)
469 		return ret;
470 
471 	while (data_opts.kill_ptrs) {
472 		unsigned i = 0, drop = __fls(data_opts.kill_ptrs);
473 
474 		bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, i++ == drop);
475 		data_opts.kill_ptrs ^= 1U << drop;
476 	}
477 
478 	/*
479 	 * If the new extent no longer has any pointers, bch2_extent_normalize()
480 	 * will do the appropriate thing with it (turning it into a
481 	 * KEY_TYPE_error key, or just a discard if it was a cached extent)
482 	 */
483 	bch2_extent_normalize(c, bkey_i_to_s(n));
484 
485 	/*
486 	 * Since we're not inserting through an extent iterator
487 	 * (BTREE_ITER_all_snapshots iterators aren't extent iterators),
488 	 * we aren't using the extent overwrite path to delete, we're
489 	 * just using the normal key deletion path:
490 	 */
491 	if (bkey_deleted(&n->k) && !(iter->flags & BTREE_ITER_is_extents))
492 		n->k.size = 0;
493 
494 	return bch2_trans_relock(trans) ?:
495 		bch2_trans_update(trans, iter, n, BTREE_UPDATE_internal_snapshot_node) ?:
496 		bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc);
497 }
498 
499 int bch2_data_update_init(struct btree_trans *trans,
500 			  struct btree_iter *iter,
501 			  struct moving_context *ctxt,
502 			  struct data_update *m,
503 			  struct write_point_specifier wp,
504 			  struct bch_io_opts io_opts,
505 			  struct data_update_opts data_opts,
506 			  enum btree_id btree_id,
507 			  struct bkey_s_c k)
508 {
509 	struct bch_fs *c = trans->c;
510 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
511 	const union bch_extent_entry *entry;
512 	struct extent_ptr_decoded p;
513 	unsigned i, reserve_sectors = k.k->size * data_opts.extra_replicas;
514 	unsigned ptrs_locked = 0;
515 	int ret = 0;
516 
517 	/*
518 	 * fs is corrupt  we have a key for a snapshot node that doesn't exist,
519 	 * and we have to check for this because we go rw before repairing the
520 	 * snapshots table - just skip it, we can move it later.
521 	 */
522 	if (unlikely(k.k->p.snapshot && !bch2_snapshot_equiv(c, k.k->p.snapshot)))
523 		return -BCH_ERR_data_update_done;
524 
525 	bch2_bkey_buf_init(&m->k);
526 	bch2_bkey_buf_reassemble(&m->k, c, k);
527 	m->btree_id	= btree_id;
528 	m->data_opts	= data_opts;
529 	m->ctxt		= ctxt;
530 	m->stats	= ctxt ? ctxt->stats : NULL;
531 
532 	bch2_write_op_init(&m->op, c, io_opts);
533 	m->op.pos	= bkey_start_pos(k.k);
534 	m->op.version	= k.k->version;
535 	m->op.target	= data_opts.target;
536 	m->op.write_point = wp;
537 	m->op.nr_replicas = 0;
538 	m->op.flags	|= BCH_WRITE_PAGES_STABLE|
539 		BCH_WRITE_PAGES_OWNED|
540 		BCH_WRITE_DATA_ENCODED|
541 		BCH_WRITE_MOVE|
542 		m->data_opts.write_flags;
543 	m->op.compression_opt	= background_compression(io_opts);
544 	m->op.watermark		= m->data_opts.btree_insert_flags & BCH_WATERMARK_MASK;
545 
546 	bkey_for_each_ptr(ptrs, ptr) {
547 		if (!bch2_dev_tryget(c, ptr->dev)) {
548 			bkey_for_each_ptr(ptrs, ptr2) {
549 				if (ptr2 == ptr)
550 					break;
551 				bch2_dev_put(bch2_dev_have_ref(c, ptr2->dev));
552 			}
553 			return -BCH_ERR_data_update_done;
554 		}
555 	}
556 
557 	unsigned durability_have = 0, durability_removing = 0;
558 
559 	i = 0;
560 	bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
561 		struct bch_dev *ca = bch2_dev_have_ref(c, p.ptr.dev);
562 		struct bpos bucket = PTR_BUCKET_POS(ca, &p.ptr);
563 		bool locked;
564 
565 		rcu_read_lock();
566 		if (((1U << i) & m->data_opts.rewrite_ptrs)) {
567 			BUG_ON(p.ptr.cached);
568 
569 			if (crc_is_compressed(p.crc))
570 				reserve_sectors += k.k->size;
571 
572 			m->op.nr_replicas += bch2_extent_ptr_desired_durability(c, &p);
573 			durability_removing += bch2_extent_ptr_desired_durability(c, &p);
574 		} else if (!p.ptr.cached &&
575 			   !((1U << i) & m->data_opts.kill_ptrs)) {
576 			bch2_dev_list_add_dev(&m->op.devs_have, p.ptr.dev);
577 			durability_have += bch2_extent_ptr_durability(c, &p);
578 		}
579 		rcu_read_unlock();
580 
581 		/*
582 		 * op->csum_type is normally initialized from the fs/file's
583 		 * current options - but if an extent is encrypted, we require
584 		 * that it stays encrypted:
585 		 */
586 		if (bch2_csum_type_is_encryption(p.crc.csum_type)) {
587 			m->op.nonce	= p.crc.nonce + p.crc.offset;
588 			m->op.csum_type = p.crc.csum_type;
589 		}
590 
591 		if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible)
592 			m->op.incompressible = true;
593 
594 		if (c->opts.nocow_enabled) {
595 			if (ctxt) {
596 				move_ctxt_wait_event(ctxt,
597 						(locked = bch2_bucket_nocow_trylock(&c->nocow_locks,
598 									  bucket, 0)) ||
599 						list_empty(&ctxt->ios));
600 
601 				if (!locked)
602 					bch2_bucket_nocow_lock(&c->nocow_locks, bucket, 0);
603 			} else {
604 				if (!bch2_bucket_nocow_trylock(&c->nocow_locks, bucket, 0)) {
605 					ret = -BCH_ERR_nocow_lock_blocked;
606 					goto err;
607 				}
608 			}
609 			ptrs_locked |= (1U << i);
610 		}
611 
612 		i++;
613 	}
614 
615 	unsigned durability_required = max(0, (int) (io_opts.data_replicas - durability_have));
616 
617 	/*
618 	 * If current extent durability is less than io_opts.data_replicas,
619 	 * we're not trying to rereplicate the extent up to data_replicas here -
620 	 * unless extra_replicas was specified
621 	 *
622 	 * Increasing replication is an explicit operation triggered by
623 	 * rereplicate, currently, so that users don't get an unexpected -ENOSPC
624 	 */
625 	if (!(m->data_opts.write_flags & BCH_WRITE_CACHED) &&
626 	    !durability_required) {
627 		m->data_opts.kill_ptrs |= m->data_opts.rewrite_ptrs;
628 		m->data_opts.rewrite_ptrs = 0;
629 		/* if iter == NULL, it's just a promote */
630 		if (iter)
631 			ret = bch2_extent_drop_ptrs(trans, iter, k, m->data_opts);
632 		goto done;
633 	}
634 
635 	m->op.nr_replicas = min(durability_removing, durability_required) +
636 		m->data_opts.extra_replicas;
637 
638 	/*
639 	 * If device(s) were set to durability=0 after data was written to them
640 	 * we can end up with a duribilty=0 extent, and the normal algorithm
641 	 * that tries not to increase durability doesn't work:
642 	 */
643 	if (!(durability_have + durability_removing))
644 		m->op.nr_replicas = max((unsigned) m->op.nr_replicas, 1);
645 
646 	m->op.nr_replicas_required = m->op.nr_replicas;
647 
648 	if (reserve_sectors) {
649 		ret = bch2_disk_reservation_add(c, &m->op.res, reserve_sectors,
650 				m->data_opts.extra_replicas
651 				? 0
652 				: BCH_DISK_RESERVATION_NOFAIL);
653 		if (ret)
654 			goto err;
655 	}
656 
657 	if (bkey_extent_is_unwritten(k)) {
658 		bch2_update_unwritten_extent(trans, m);
659 		goto done;
660 	}
661 
662 	return 0;
663 err:
664 	i = 0;
665 	bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
666 		struct bch_dev *ca = bch2_dev_have_ref(c, p.ptr.dev);
667 		struct bpos bucket = PTR_BUCKET_POS(ca, &p.ptr);
668 		if ((1U << i) & ptrs_locked)
669 			bch2_bucket_nocow_unlock(&c->nocow_locks, bucket, 0);
670 		bch2_dev_put(ca);
671 		i++;
672 	}
673 
674 	bch2_bkey_buf_exit(&m->k, c);
675 	bch2_bio_free_pages_pool(c, &m->op.wbio.bio);
676 	return ret;
677 done:
678 	bch2_data_update_exit(m);
679 	return ret ?: -BCH_ERR_data_update_done;
680 }
681 
682 void bch2_data_update_opts_normalize(struct bkey_s_c k, struct data_update_opts *opts)
683 {
684 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
685 	unsigned i = 0;
686 
687 	bkey_for_each_ptr(ptrs, ptr) {
688 		if ((opts->rewrite_ptrs & (1U << i)) && ptr->cached) {
689 			opts->kill_ptrs |= 1U << i;
690 			opts->rewrite_ptrs ^= 1U << i;
691 		}
692 
693 		i++;
694 	}
695 }
696