xref: /linux/fs/bcachefs/io_write.c (revision eb01fe7abbe2d0b38824d2a93fdb4cc3eaf2ccc1)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
4  * Copyright 2012 Google, Inc.
5  */
6 
7 #include "bcachefs.h"
8 #include "alloc_foreground.h"
9 #include "bkey_buf.h"
10 #include "bset.h"
11 #include "btree_update.h"
12 #include "buckets.h"
13 #include "checksum.h"
14 #include "clock.h"
15 #include "compress.h"
16 #include "debug.h"
17 #include "ec.h"
18 #include "error.h"
19 #include "extent_update.h"
20 #include "inode.h"
21 #include "io_write.h"
22 #include "journal.h"
23 #include "keylist.h"
24 #include "move.h"
25 #include "nocow_locking.h"
26 #include "rebalance.h"
27 #include "subvolume.h"
28 #include "super.h"
29 #include "super-io.h"
30 #include "trace.h"
31 
32 #include <linux/blkdev.h>
33 #include <linux/prefetch.h>
34 #include <linux/random.h>
35 #include <linux/sched/mm.h>
36 
37 #ifndef CONFIG_BCACHEFS_NO_LATENCY_ACCT
38 
39 static inline void bch2_congested_acct(struct bch_dev *ca, u64 io_latency,
40 				       u64 now, int rw)
41 {
42 	u64 latency_capable =
43 		ca->io_latency[rw].quantiles.entries[QUANTILE_IDX(1)].m;
44 	/* ideally we'd be taking into account the device's variance here: */
45 	u64 latency_threshold = latency_capable << (rw == READ ? 2 : 3);
46 	s64 latency_over = io_latency - latency_threshold;
47 
48 	if (latency_threshold && latency_over > 0) {
49 		/*
50 		 * bump up congested by approximately latency_over * 4 /
51 		 * latency_threshold - we don't need much accuracy here so don't
52 		 * bother with the divide:
53 		 */
54 		if (atomic_read(&ca->congested) < CONGESTED_MAX)
55 			atomic_add(latency_over >>
56 				   max_t(int, ilog2(latency_threshold) - 2, 0),
57 				   &ca->congested);
58 
59 		ca->congested_last = now;
60 	} else if (atomic_read(&ca->congested) > 0) {
61 		atomic_dec(&ca->congested);
62 	}
63 }
64 
65 void bch2_latency_acct(struct bch_dev *ca, u64 submit_time, int rw)
66 {
67 	atomic64_t *latency = &ca->cur_latency[rw];
68 	u64 now = local_clock();
69 	u64 io_latency = time_after64(now, submit_time)
70 		? now - submit_time
71 		: 0;
72 	u64 old, new, v = atomic64_read(latency);
73 
74 	do {
75 		old = v;
76 
77 		/*
78 		 * If the io latency was reasonably close to the current
79 		 * latency, skip doing the update and atomic operation - most of
80 		 * the time:
81 		 */
82 		if (abs((int) (old - io_latency)) < (old >> 1) &&
83 		    now & ~(~0U << 5))
84 			break;
85 
86 		new = ewma_add(old, io_latency, 5);
87 	} while ((v = atomic64_cmpxchg(latency, old, new)) != old);
88 
89 	bch2_congested_acct(ca, io_latency, now, rw);
90 
91 	__bch2_time_stats_update(&ca->io_latency[rw].stats, submit_time, now);
92 }
93 
94 #endif
95 
96 /* Allocate, free from mempool: */
97 
98 void bch2_bio_free_pages_pool(struct bch_fs *c, struct bio *bio)
99 {
100 	struct bvec_iter_all iter;
101 	struct bio_vec *bv;
102 
103 	bio_for_each_segment_all(bv, bio, iter)
104 		if (bv->bv_page != ZERO_PAGE(0))
105 			mempool_free(bv->bv_page, &c->bio_bounce_pages);
106 	bio->bi_vcnt = 0;
107 }
108 
109 static struct page *__bio_alloc_page_pool(struct bch_fs *c, bool *using_mempool)
110 {
111 	struct page *page;
112 
113 	if (likely(!*using_mempool)) {
114 		page = alloc_page(GFP_NOFS);
115 		if (unlikely(!page)) {
116 			mutex_lock(&c->bio_bounce_pages_lock);
117 			*using_mempool = true;
118 			goto pool_alloc;
119 
120 		}
121 	} else {
122 pool_alloc:
123 		page = mempool_alloc(&c->bio_bounce_pages, GFP_NOFS);
124 	}
125 
126 	return page;
127 }
128 
129 void bch2_bio_alloc_pages_pool(struct bch_fs *c, struct bio *bio,
130 			       size_t size)
131 {
132 	bool using_mempool = false;
133 
134 	while (size) {
135 		struct page *page = __bio_alloc_page_pool(c, &using_mempool);
136 		unsigned len = min_t(size_t, PAGE_SIZE, size);
137 
138 		BUG_ON(!bio_add_page(bio, page, len, 0));
139 		size -= len;
140 	}
141 
142 	if (using_mempool)
143 		mutex_unlock(&c->bio_bounce_pages_lock);
144 }
145 
146 /* Extent update path: */
147 
148 int bch2_sum_sector_overwrites(struct btree_trans *trans,
149 			       struct btree_iter *extent_iter,
150 			       struct bkey_i *new,
151 			       bool *usage_increasing,
152 			       s64 *i_sectors_delta,
153 			       s64 *disk_sectors_delta)
154 {
155 	struct bch_fs *c = trans->c;
156 	struct btree_iter iter;
157 	struct bkey_s_c old;
158 	unsigned new_replicas = bch2_bkey_replicas(c, bkey_i_to_s_c(new));
159 	bool new_compressed = bch2_bkey_sectors_compressed(bkey_i_to_s_c(new));
160 	int ret = 0;
161 
162 	*usage_increasing	= false;
163 	*i_sectors_delta	= 0;
164 	*disk_sectors_delta	= 0;
165 
166 	bch2_trans_copy_iter(&iter, extent_iter);
167 
168 	for_each_btree_key_upto_continue_norestart(iter,
169 				new->k.p, BTREE_ITER_SLOTS, old, ret) {
170 		s64 sectors = min(new->k.p.offset, old.k->p.offset) -
171 			max(bkey_start_offset(&new->k),
172 			    bkey_start_offset(old.k));
173 
174 		*i_sectors_delta += sectors *
175 			(bkey_extent_is_allocation(&new->k) -
176 			 bkey_extent_is_allocation(old.k));
177 
178 		*disk_sectors_delta += sectors * bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(new));
179 		*disk_sectors_delta -= new->k.p.snapshot == old.k->p.snapshot
180 			? sectors * bch2_bkey_nr_ptrs_fully_allocated(old)
181 			: 0;
182 
183 		if (!*usage_increasing &&
184 		    (new->k.p.snapshot != old.k->p.snapshot ||
185 		     new_replicas > bch2_bkey_replicas(c, old) ||
186 		     (!new_compressed && bch2_bkey_sectors_compressed(old))))
187 			*usage_increasing = true;
188 
189 		if (bkey_ge(old.k->p, new->k.p))
190 			break;
191 	}
192 
193 	bch2_trans_iter_exit(trans, &iter);
194 	return ret;
195 }
196 
197 static inline int bch2_extent_update_i_size_sectors(struct btree_trans *trans,
198 						    struct btree_iter *extent_iter,
199 						    u64 new_i_size,
200 						    s64 i_sectors_delta)
201 {
202 	struct btree_iter iter;
203 	struct bkey_i *k;
204 	struct bkey_i_inode_v3 *inode;
205 	/*
206 	 * Crazy performance optimization:
207 	 * Every extent update needs to also update the inode: the inode trigger
208 	 * will set bi->journal_seq to the journal sequence number of this
209 	 * transaction - for fsync.
210 	 *
211 	 * But if that's the only reason we're updating the inode (we're not
212 	 * updating bi_size or bi_sectors), then we don't need the inode update
213 	 * to be journalled - if we crash, the bi_journal_seq update will be
214 	 * lost, but that's fine.
215 	 */
216 	unsigned inode_update_flags = BTREE_UPDATE_NOJOURNAL;
217 	int ret;
218 
219 	k = bch2_bkey_get_mut_noupdate(trans, &iter, BTREE_ID_inodes,
220 			      SPOS(0,
221 				   extent_iter->pos.inode,
222 				   extent_iter->snapshot),
223 			      BTREE_ITER_CACHED);
224 	ret = PTR_ERR_OR_ZERO(k);
225 	if (unlikely(ret))
226 		return ret;
227 
228 	if (unlikely(k->k.type != KEY_TYPE_inode_v3)) {
229 		k = bch2_inode_to_v3(trans, k);
230 		ret = PTR_ERR_OR_ZERO(k);
231 		if (unlikely(ret))
232 			goto err;
233 	}
234 
235 	inode = bkey_i_to_inode_v3(k);
236 
237 	if (!(le64_to_cpu(inode->v.bi_flags) & BCH_INODE_i_size_dirty) &&
238 	    new_i_size > le64_to_cpu(inode->v.bi_size)) {
239 		inode->v.bi_size = cpu_to_le64(new_i_size);
240 		inode_update_flags = 0;
241 	}
242 
243 	if (i_sectors_delta) {
244 		le64_add_cpu(&inode->v.bi_sectors, i_sectors_delta);
245 		inode_update_flags = 0;
246 	}
247 
248 	if (inode->k.p.snapshot != iter.snapshot) {
249 		inode->k.p.snapshot = iter.snapshot;
250 		inode_update_flags = 0;
251 	}
252 
253 	ret = bch2_trans_update(trans, &iter, &inode->k_i,
254 				BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
255 				inode_update_flags);
256 err:
257 	bch2_trans_iter_exit(trans, &iter);
258 	return ret;
259 }
260 
261 int bch2_extent_update(struct btree_trans *trans,
262 		       subvol_inum inum,
263 		       struct btree_iter *iter,
264 		       struct bkey_i *k,
265 		       struct disk_reservation *disk_res,
266 		       u64 new_i_size,
267 		       s64 *i_sectors_delta_total,
268 		       bool check_enospc)
269 {
270 	struct bpos next_pos;
271 	bool usage_increasing;
272 	s64 i_sectors_delta = 0, disk_sectors_delta = 0;
273 	int ret;
274 
275 	/*
276 	 * This traverses us the iterator without changing iter->path->pos to
277 	 * search_key() (which is pos + 1 for extents): we want there to be a
278 	 * path already traversed at iter->pos because
279 	 * bch2_trans_extent_update() will use it to attempt extent merging
280 	 */
281 	ret = __bch2_btree_iter_traverse(iter);
282 	if (ret)
283 		return ret;
284 
285 	ret = bch2_extent_trim_atomic(trans, iter, k);
286 	if (ret)
287 		return ret;
288 
289 	next_pos = k->k.p;
290 
291 	ret = bch2_sum_sector_overwrites(trans, iter, k,
292 			&usage_increasing,
293 			&i_sectors_delta,
294 			&disk_sectors_delta);
295 	if (ret)
296 		return ret;
297 
298 	if (disk_res &&
299 	    disk_sectors_delta > (s64) disk_res->sectors) {
300 		ret = bch2_disk_reservation_add(trans->c, disk_res,
301 					disk_sectors_delta - disk_res->sectors,
302 					!check_enospc || !usage_increasing
303 					? BCH_DISK_RESERVATION_NOFAIL : 0);
304 		if (ret)
305 			return ret;
306 	}
307 
308 	/*
309 	 * Note:
310 	 * We always have to do an inode update - even when i_size/i_sectors
311 	 * aren't changing - for fsync to work properly; fsync relies on
312 	 * inode->bi_journal_seq which is updated by the trigger code:
313 	 */
314 	ret =   bch2_extent_update_i_size_sectors(trans, iter,
315 						  min(k->k.p.offset << 9, new_i_size),
316 						  i_sectors_delta) ?:
317 		bch2_trans_update(trans, iter, k, 0) ?:
318 		bch2_trans_commit(trans, disk_res, NULL,
319 				BCH_TRANS_COMMIT_no_check_rw|
320 				BCH_TRANS_COMMIT_no_enospc);
321 	if (unlikely(ret))
322 		return ret;
323 
324 	if (i_sectors_delta_total)
325 		*i_sectors_delta_total += i_sectors_delta;
326 	bch2_btree_iter_set_pos(iter, next_pos);
327 	return 0;
328 }
329 
330 static int bch2_write_index_default(struct bch_write_op *op)
331 {
332 	struct bch_fs *c = op->c;
333 	struct bkey_buf sk;
334 	struct keylist *keys = &op->insert_keys;
335 	struct bkey_i *k = bch2_keylist_front(keys);
336 	struct btree_trans *trans = bch2_trans_get(c);
337 	struct btree_iter iter;
338 	subvol_inum inum = {
339 		.subvol = op->subvol,
340 		.inum	= k->k.p.inode,
341 	};
342 	int ret;
343 
344 	BUG_ON(!inum.subvol);
345 
346 	bch2_bkey_buf_init(&sk);
347 
348 	do {
349 		bch2_trans_begin(trans);
350 
351 		k = bch2_keylist_front(keys);
352 		bch2_bkey_buf_copy(&sk, c, k);
353 
354 		ret = bch2_subvolume_get_snapshot(trans, inum.subvol,
355 						  &sk.k->k.p.snapshot);
356 		if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
357 			continue;
358 		if (ret)
359 			break;
360 
361 		bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
362 				     bkey_start_pos(&sk.k->k),
363 				     BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
364 
365 		ret =   bch2_bkey_set_needs_rebalance(c, sk.k, &op->opts) ?:
366 			bch2_extent_update(trans, inum, &iter, sk.k,
367 					&op->res,
368 					op->new_i_size, &op->i_sectors_delta,
369 					op->flags & BCH_WRITE_CHECK_ENOSPC);
370 		bch2_trans_iter_exit(trans, &iter);
371 
372 		if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
373 			continue;
374 		if (ret)
375 			break;
376 
377 		if (bkey_ge(iter.pos, k->k.p))
378 			bch2_keylist_pop_front(&op->insert_keys);
379 		else
380 			bch2_cut_front(iter.pos, k);
381 	} while (!bch2_keylist_empty(keys));
382 
383 	bch2_trans_put(trans);
384 	bch2_bkey_buf_exit(&sk, c);
385 
386 	return ret;
387 }
388 
389 /* Writes */
390 
391 void bch2_submit_wbio_replicas(struct bch_write_bio *wbio, struct bch_fs *c,
392 			       enum bch_data_type type,
393 			       const struct bkey_i *k,
394 			       bool nocow)
395 {
396 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(bkey_i_to_s_c(k));
397 	struct bch_write_bio *n;
398 
399 	BUG_ON(c->opts.nochanges);
400 
401 	bkey_for_each_ptr(ptrs, ptr) {
402 		BUG_ON(!bch2_dev_exists2(c, ptr->dev));
403 
404 		struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
405 
406 		if (to_entry(ptr + 1) < ptrs.end) {
407 			n = to_wbio(bio_alloc_clone(NULL, &wbio->bio,
408 						GFP_NOFS, &ca->replica_set));
409 
410 			n->bio.bi_end_io	= wbio->bio.bi_end_io;
411 			n->bio.bi_private	= wbio->bio.bi_private;
412 			n->parent		= wbio;
413 			n->split		= true;
414 			n->bounce		= false;
415 			n->put_bio		= true;
416 			n->bio.bi_opf		= wbio->bio.bi_opf;
417 			bio_inc_remaining(&wbio->bio);
418 		} else {
419 			n = wbio;
420 			n->split		= false;
421 		}
422 
423 		n->c			= c;
424 		n->dev			= ptr->dev;
425 		n->have_ioref		= nocow || bch2_dev_get_ioref(ca,
426 					type == BCH_DATA_btree ? READ : WRITE);
427 		n->nocow		= nocow;
428 		n->submit_time		= local_clock();
429 		n->inode_offset		= bkey_start_offset(&k->k);
430 		n->bio.bi_iter.bi_sector = ptr->offset;
431 
432 		if (likely(n->have_ioref)) {
433 			this_cpu_add(ca->io_done->sectors[WRITE][type],
434 				     bio_sectors(&n->bio));
435 
436 			bio_set_dev(&n->bio, ca->disk_sb.bdev);
437 
438 			if (type != BCH_DATA_btree && unlikely(c->opts.no_data_io)) {
439 				bio_endio(&n->bio);
440 				continue;
441 			}
442 
443 			submit_bio(&n->bio);
444 		} else {
445 			n->bio.bi_status	= BLK_STS_REMOVED;
446 			bio_endio(&n->bio);
447 		}
448 	}
449 }
450 
451 static void __bch2_write(struct bch_write_op *);
452 
453 static void bch2_write_done(struct closure *cl)
454 {
455 	struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
456 	struct bch_fs *c = op->c;
457 
458 	EBUG_ON(op->open_buckets.nr);
459 
460 	bch2_time_stats_update(&c->times[BCH_TIME_data_write], op->start_time);
461 	bch2_disk_reservation_put(c, &op->res);
462 
463 	if (!(op->flags & BCH_WRITE_MOVE))
464 		bch2_write_ref_put(c, BCH_WRITE_REF_write);
465 	bch2_keylist_free(&op->insert_keys, op->inline_keys);
466 
467 	EBUG_ON(cl->parent);
468 	closure_debug_destroy(cl);
469 	if (op->end_io)
470 		op->end_io(op);
471 }
472 
473 static noinline int bch2_write_drop_io_error_ptrs(struct bch_write_op *op)
474 {
475 	struct keylist *keys = &op->insert_keys;
476 	struct bch_extent_ptr *ptr;
477 	struct bkey_i *src, *dst = keys->keys, *n;
478 
479 	for (src = keys->keys; src != keys->top; src = n) {
480 		n = bkey_next(src);
481 
482 		if (bkey_extent_is_direct_data(&src->k)) {
483 			bch2_bkey_drop_ptrs(bkey_i_to_s(src), ptr,
484 					    test_bit(ptr->dev, op->failed.d));
485 
486 			if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(src)))
487 				return -EIO;
488 		}
489 
490 		if (dst != src)
491 			memmove_u64s_down(dst, src, src->k.u64s);
492 		dst = bkey_next(dst);
493 	}
494 
495 	keys->top = dst;
496 	return 0;
497 }
498 
499 /**
500  * __bch2_write_index - after a write, update index to point to new data
501  * @op:		bch_write_op to process
502  */
503 static void __bch2_write_index(struct bch_write_op *op)
504 {
505 	struct bch_fs *c = op->c;
506 	struct keylist *keys = &op->insert_keys;
507 	unsigned dev;
508 	int ret = 0;
509 
510 	if (unlikely(op->flags & BCH_WRITE_IO_ERROR)) {
511 		ret = bch2_write_drop_io_error_ptrs(op);
512 		if (ret)
513 			goto err;
514 	}
515 
516 	if (!bch2_keylist_empty(keys)) {
517 		u64 sectors_start = keylist_sectors(keys);
518 
519 		ret = !(op->flags & BCH_WRITE_MOVE)
520 			? bch2_write_index_default(op)
521 			: bch2_data_update_index_update(op);
522 
523 		BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart));
524 		BUG_ON(keylist_sectors(keys) && !ret);
525 
526 		op->written += sectors_start - keylist_sectors(keys);
527 
528 		if (ret && !bch2_err_matches(ret, EROFS)) {
529 			struct bkey_i *insert = bch2_keylist_front(&op->insert_keys);
530 
531 			bch_err_inum_offset_ratelimited(c,
532 				insert->k.p.inode, insert->k.p.offset << 9,
533 				"%s write error while doing btree update: %s",
534 				op->flags & BCH_WRITE_MOVE ? "move" : "user",
535 				bch2_err_str(ret));
536 		}
537 
538 		if (ret)
539 			goto err;
540 	}
541 out:
542 	/* If some a bucket wasn't written, we can't erasure code it: */
543 	for_each_set_bit(dev, op->failed.d, BCH_SB_MEMBERS_MAX)
544 		bch2_open_bucket_write_error(c, &op->open_buckets, dev);
545 
546 	bch2_open_buckets_put(c, &op->open_buckets);
547 	return;
548 err:
549 	keys->top = keys->keys;
550 	op->error = ret;
551 	op->flags |= BCH_WRITE_DONE;
552 	goto out;
553 }
554 
555 static inline void __wp_update_state(struct write_point *wp, enum write_point_state state)
556 {
557 	if (state != wp->state) {
558 		u64 now = ktime_get_ns();
559 
560 		if (wp->last_state_change &&
561 		    time_after64(now, wp->last_state_change))
562 			wp->time[wp->state] += now - wp->last_state_change;
563 		wp->state = state;
564 		wp->last_state_change = now;
565 	}
566 }
567 
568 static inline void wp_update_state(struct write_point *wp, bool running)
569 {
570 	enum write_point_state state;
571 
572 	state = running			 ? WRITE_POINT_running :
573 		!list_empty(&wp->writes) ? WRITE_POINT_waiting_io
574 					 : WRITE_POINT_stopped;
575 
576 	__wp_update_state(wp, state);
577 }
578 
579 static CLOSURE_CALLBACK(bch2_write_index)
580 {
581 	closure_type(op, struct bch_write_op, cl);
582 	struct write_point *wp = op->wp;
583 	struct workqueue_struct *wq = index_update_wq(op);
584 	unsigned long flags;
585 
586 	if ((op->flags & BCH_WRITE_DONE) &&
587 	    (op->flags & BCH_WRITE_MOVE))
588 		bch2_bio_free_pages_pool(op->c, &op->wbio.bio);
589 
590 	spin_lock_irqsave(&wp->writes_lock, flags);
591 	if (wp->state == WRITE_POINT_waiting_io)
592 		__wp_update_state(wp, WRITE_POINT_waiting_work);
593 	list_add_tail(&op->wp_list, &wp->writes);
594 	spin_unlock_irqrestore (&wp->writes_lock, flags);
595 
596 	queue_work(wq, &wp->index_update_work);
597 }
598 
599 static inline void bch2_write_queue(struct bch_write_op *op, struct write_point *wp)
600 {
601 	op->wp = wp;
602 
603 	if (wp->state == WRITE_POINT_stopped) {
604 		spin_lock_irq(&wp->writes_lock);
605 		__wp_update_state(wp, WRITE_POINT_waiting_io);
606 		spin_unlock_irq(&wp->writes_lock);
607 	}
608 }
609 
610 void bch2_write_point_do_index_updates(struct work_struct *work)
611 {
612 	struct write_point *wp =
613 		container_of(work, struct write_point, index_update_work);
614 	struct bch_write_op *op;
615 
616 	while (1) {
617 		spin_lock_irq(&wp->writes_lock);
618 		op = list_first_entry_or_null(&wp->writes, struct bch_write_op, wp_list);
619 		if (op)
620 			list_del(&op->wp_list);
621 		wp_update_state(wp, op != NULL);
622 		spin_unlock_irq(&wp->writes_lock);
623 
624 		if (!op)
625 			break;
626 
627 		op->flags |= BCH_WRITE_IN_WORKER;
628 
629 		__bch2_write_index(op);
630 
631 		if (!(op->flags & BCH_WRITE_DONE))
632 			__bch2_write(op);
633 		else
634 			bch2_write_done(&op->cl);
635 	}
636 }
637 
638 static void bch2_write_endio(struct bio *bio)
639 {
640 	struct closure *cl		= bio->bi_private;
641 	struct bch_write_op *op		= container_of(cl, struct bch_write_op, cl);
642 	struct bch_write_bio *wbio	= to_wbio(bio);
643 	struct bch_write_bio *parent	= wbio->split ? wbio->parent : NULL;
644 	struct bch_fs *c		= wbio->c;
645 	struct bch_dev *ca		= bch_dev_bkey_exists(c, wbio->dev);
646 
647 	if (bch2_dev_inum_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_write,
648 				    op->pos.inode,
649 				    wbio->inode_offset << 9,
650 				    "data write error: %s",
651 				    bch2_blk_status_to_str(bio->bi_status))) {
652 		set_bit(wbio->dev, op->failed.d);
653 		op->flags |= BCH_WRITE_IO_ERROR;
654 	}
655 
656 	if (wbio->nocow)
657 		set_bit(wbio->dev, op->devs_need_flush->d);
658 
659 	if (wbio->have_ioref) {
660 		bch2_latency_acct(ca, wbio->submit_time, WRITE);
661 		percpu_ref_put(&ca->io_ref);
662 	}
663 
664 	if (wbio->bounce)
665 		bch2_bio_free_pages_pool(c, bio);
666 
667 	if (wbio->put_bio)
668 		bio_put(bio);
669 
670 	if (parent)
671 		bio_endio(&parent->bio);
672 	else
673 		closure_put(cl);
674 }
675 
676 static void init_append_extent(struct bch_write_op *op,
677 			       struct write_point *wp,
678 			       struct bversion version,
679 			       struct bch_extent_crc_unpacked crc)
680 {
681 	struct bkey_i_extent *e;
682 
683 	op->pos.offset += crc.uncompressed_size;
684 
685 	e = bkey_extent_init(op->insert_keys.top);
686 	e->k.p		= op->pos;
687 	e->k.size	= crc.uncompressed_size;
688 	e->k.version	= version;
689 
690 	if (crc.csum_type ||
691 	    crc.compression_type ||
692 	    crc.nonce)
693 		bch2_extent_crc_append(&e->k_i, crc);
694 
695 	bch2_alloc_sectors_append_ptrs_inlined(op->c, wp, &e->k_i, crc.compressed_size,
696 				       op->flags & BCH_WRITE_CACHED);
697 
698 	bch2_keylist_push(&op->insert_keys);
699 }
700 
701 static struct bio *bch2_write_bio_alloc(struct bch_fs *c,
702 					struct write_point *wp,
703 					struct bio *src,
704 					bool *page_alloc_failed,
705 					void *buf)
706 {
707 	struct bch_write_bio *wbio;
708 	struct bio *bio;
709 	unsigned output_available =
710 		min(wp->sectors_free << 9, src->bi_iter.bi_size);
711 	unsigned pages = DIV_ROUND_UP(output_available +
712 				      (buf
713 				       ? ((unsigned long) buf & (PAGE_SIZE - 1))
714 				       : 0), PAGE_SIZE);
715 
716 	pages = min(pages, BIO_MAX_VECS);
717 
718 	bio = bio_alloc_bioset(NULL, pages, 0,
719 			       GFP_NOFS, &c->bio_write);
720 	wbio			= wbio_init(bio);
721 	wbio->put_bio		= true;
722 	/* copy WRITE_SYNC flag */
723 	wbio->bio.bi_opf	= src->bi_opf;
724 
725 	if (buf) {
726 		bch2_bio_map(bio, buf, output_available);
727 		return bio;
728 	}
729 
730 	wbio->bounce		= true;
731 
732 	/*
733 	 * We can't use mempool for more than c->sb.encoded_extent_max
734 	 * worth of pages, but we'd like to allocate more if we can:
735 	 */
736 	bch2_bio_alloc_pages_pool(c, bio,
737 				  min_t(unsigned, output_available,
738 					c->opts.encoded_extent_max));
739 
740 	if (bio->bi_iter.bi_size < output_available)
741 		*page_alloc_failed =
742 			bch2_bio_alloc_pages(bio,
743 					     output_available -
744 					     bio->bi_iter.bi_size,
745 					     GFP_NOFS) != 0;
746 
747 	return bio;
748 }
749 
750 static int bch2_write_rechecksum(struct bch_fs *c,
751 				 struct bch_write_op *op,
752 				 unsigned new_csum_type)
753 {
754 	struct bio *bio = &op->wbio.bio;
755 	struct bch_extent_crc_unpacked new_crc;
756 	int ret;
757 
758 	/* bch2_rechecksum_bio() can't encrypt or decrypt data: */
759 
760 	if (bch2_csum_type_is_encryption(op->crc.csum_type) !=
761 	    bch2_csum_type_is_encryption(new_csum_type))
762 		new_csum_type = op->crc.csum_type;
763 
764 	ret = bch2_rechecksum_bio(c, bio, op->version, op->crc,
765 				  NULL, &new_crc,
766 				  op->crc.offset, op->crc.live_size,
767 				  new_csum_type);
768 	if (ret)
769 		return ret;
770 
771 	bio_advance(bio, op->crc.offset << 9);
772 	bio->bi_iter.bi_size = op->crc.live_size << 9;
773 	op->crc = new_crc;
774 	return 0;
775 }
776 
777 static int bch2_write_decrypt(struct bch_write_op *op)
778 {
779 	struct bch_fs *c = op->c;
780 	struct nonce nonce = extent_nonce(op->version, op->crc);
781 	struct bch_csum csum;
782 	int ret;
783 
784 	if (!bch2_csum_type_is_encryption(op->crc.csum_type))
785 		return 0;
786 
787 	/*
788 	 * If we need to decrypt data in the write path, we'll no longer be able
789 	 * to verify the existing checksum (poly1305 mac, in this case) after
790 	 * it's decrypted - this is the last point we'll be able to reverify the
791 	 * checksum:
792 	 */
793 	csum = bch2_checksum_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
794 	if (bch2_crc_cmp(op->crc.csum, csum) && !c->opts.no_data_io)
795 		return -EIO;
796 
797 	ret = bch2_encrypt_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
798 	op->crc.csum_type = 0;
799 	op->crc.csum = (struct bch_csum) { 0, 0 };
800 	return ret;
801 }
802 
803 static enum prep_encoded_ret {
804 	PREP_ENCODED_OK,
805 	PREP_ENCODED_ERR,
806 	PREP_ENCODED_CHECKSUM_ERR,
807 	PREP_ENCODED_DO_WRITE,
808 } bch2_write_prep_encoded_data(struct bch_write_op *op, struct write_point *wp)
809 {
810 	struct bch_fs *c = op->c;
811 	struct bio *bio = &op->wbio.bio;
812 
813 	if (!(op->flags & BCH_WRITE_DATA_ENCODED))
814 		return PREP_ENCODED_OK;
815 
816 	BUG_ON(bio_sectors(bio) != op->crc.compressed_size);
817 
818 	/* Can we just write the entire extent as is? */
819 	if (op->crc.uncompressed_size == op->crc.live_size &&
820 	    op->crc.uncompressed_size <= c->opts.encoded_extent_max >> 9 &&
821 	    op->crc.compressed_size <= wp->sectors_free &&
822 	    (op->crc.compression_type == bch2_compression_opt_to_type(op->compression_opt) ||
823 	     op->incompressible)) {
824 		if (!crc_is_compressed(op->crc) &&
825 		    op->csum_type != op->crc.csum_type &&
826 		    bch2_write_rechecksum(c, op, op->csum_type) &&
827 		    !c->opts.no_data_io)
828 			return PREP_ENCODED_CHECKSUM_ERR;
829 
830 		return PREP_ENCODED_DO_WRITE;
831 	}
832 
833 	/*
834 	 * If the data is compressed and we couldn't write the entire extent as
835 	 * is, we have to decompress it:
836 	 */
837 	if (crc_is_compressed(op->crc)) {
838 		struct bch_csum csum;
839 
840 		if (bch2_write_decrypt(op))
841 			return PREP_ENCODED_CHECKSUM_ERR;
842 
843 		/* Last point we can still verify checksum: */
844 		csum = bch2_checksum_bio(c, op->crc.csum_type,
845 					 extent_nonce(op->version, op->crc),
846 					 bio);
847 		if (bch2_crc_cmp(op->crc.csum, csum) && !c->opts.no_data_io)
848 			return PREP_ENCODED_CHECKSUM_ERR;
849 
850 		if (bch2_bio_uncompress_inplace(c, bio, &op->crc))
851 			return PREP_ENCODED_ERR;
852 	}
853 
854 	/*
855 	 * No longer have compressed data after this point - data might be
856 	 * encrypted:
857 	 */
858 
859 	/*
860 	 * If the data is checksummed and we're only writing a subset,
861 	 * rechecksum and adjust bio to point to currently live data:
862 	 */
863 	if ((op->crc.live_size != op->crc.uncompressed_size ||
864 	     op->crc.csum_type != op->csum_type) &&
865 	    bch2_write_rechecksum(c, op, op->csum_type) &&
866 	    !c->opts.no_data_io)
867 		return PREP_ENCODED_CHECKSUM_ERR;
868 
869 	/*
870 	 * If we want to compress the data, it has to be decrypted:
871 	 */
872 	if ((op->compression_opt ||
873 	     bch2_csum_type_is_encryption(op->crc.csum_type) !=
874 	     bch2_csum_type_is_encryption(op->csum_type)) &&
875 	    bch2_write_decrypt(op))
876 		return PREP_ENCODED_CHECKSUM_ERR;
877 
878 	return PREP_ENCODED_OK;
879 }
880 
881 static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp,
882 			     struct bio **_dst)
883 {
884 	struct bch_fs *c = op->c;
885 	struct bio *src = &op->wbio.bio, *dst = src;
886 	struct bvec_iter saved_iter;
887 	void *ec_buf;
888 	unsigned total_output = 0, total_input = 0;
889 	bool bounce = false;
890 	bool page_alloc_failed = false;
891 	int ret, more = 0;
892 
893 	BUG_ON(!bio_sectors(src));
894 
895 	ec_buf = bch2_writepoint_ec_buf(c, wp);
896 
897 	switch (bch2_write_prep_encoded_data(op, wp)) {
898 	case PREP_ENCODED_OK:
899 		break;
900 	case PREP_ENCODED_ERR:
901 		ret = -EIO;
902 		goto err;
903 	case PREP_ENCODED_CHECKSUM_ERR:
904 		goto csum_err;
905 	case PREP_ENCODED_DO_WRITE:
906 		/* XXX look for bug here */
907 		if (ec_buf) {
908 			dst = bch2_write_bio_alloc(c, wp, src,
909 						   &page_alloc_failed,
910 						   ec_buf);
911 			bio_copy_data(dst, src);
912 			bounce = true;
913 		}
914 		init_append_extent(op, wp, op->version, op->crc);
915 		goto do_write;
916 	}
917 
918 	if (ec_buf ||
919 	    op->compression_opt ||
920 	    (op->csum_type &&
921 	     !(op->flags & BCH_WRITE_PAGES_STABLE)) ||
922 	    (bch2_csum_type_is_encryption(op->csum_type) &&
923 	     !(op->flags & BCH_WRITE_PAGES_OWNED))) {
924 		dst = bch2_write_bio_alloc(c, wp, src,
925 					   &page_alloc_failed,
926 					   ec_buf);
927 		bounce = true;
928 	}
929 
930 	saved_iter = dst->bi_iter;
931 
932 	do {
933 		struct bch_extent_crc_unpacked crc = { 0 };
934 		struct bversion version = op->version;
935 		size_t dst_len = 0, src_len = 0;
936 
937 		if (page_alloc_failed &&
938 		    dst->bi_iter.bi_size  < (wp->sectors_free << 9) &&
939 		    dst->bi_iter.bi_size < c->opts.encoded_extent_max)
940 			break;
941 
942 		BUG_ON(op->compression_opt &&
943 		       (op->flags & BCH_WRITE_DATA_ENCODED) &&
944 		       bch2_csum_type_is_encryption(op->crc.csum_type));
945 		BUG_ON(op->compression_opt && !bounce);
946 
947 		crc.compression_type = op->incompressible
948 			? BCH_COMPRESSION_TYPE_incompressible
949 			: op->compression_opt
950 			? bch2_bio_compress(c, dst, &dst_len, src, &src_len,
951 					    op->compression_opt)
952 			: 0;
953 		if (!crc_is_compressed(crc)) {
954 			dst_len = min(dst->bi_iter.bi_size, src->bi_iter.bi_size);
955 			dst_len = min_t(unsigned, dst_len, wp->sectors_free << 9);
956 
957 			if (op->csum_type)
958 				dst_len = min_t(unsigned, dst_len,
959 						c->opts.encoded_extent_max);
960 
961 			if (bounce) {
962 				swap(dst->bi_iter.bi_size, dst_len);
963 				bio_copy_data(dst, src);
964 				swap(dst->bi_iter.bi_size, dst_len);
965 			}
966 
967 			src_len = dst_len;
968 		}
969 
970 		BUG_ON(!src_len || !dst_len);
971 
972 		if (bch2_csum_type_is_encryption(op->csum_type)) {
973 			if (bversion_zero(version)) {
974 				version.lo = atomic64_inc_return(&c->key_version);
975 			} else {
976 				crc.nonce = op->nonce;
977 				op->nonce += src_len >> 9;
978 			}
979 		}
980 
981 		if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
982 		    !crc_is_compressed(crc) &&
983 		    bch2_csum_type_is_encryption(op->crc.csum_type) ==
984 		    bch2_csum_type_is_encryption(op->csum_type)) {
985 			u8 compression_type = crc.compression_type;
986 			u16 nonce = crc.nonce;
987 			/*
988 			 * Note: when we're using rechecksum(), we need to be
989 			 * checksumming @src because it has all the data our
990 			 * existing checksum covers - if we bounced (because we
991 			 * were trying to compress), @dst will only have the
992 			 * part of the data the new checksum will cover.
993 			 *
994 			 * But normally we want to be checksumming post bounce,
995 			 * because part of the reason for bouncing is so the
996 			 * data can't be modified (by userspace) while it's in
997 			 * flight.
998 			 */
999 			if (bch2_rechecksum_bio(c, src, version, op->crc,
1000 					&crc, &op->crc,
1001 					src_len >> 9,
1002 					bio_sectors(src) - (src_len >> 9),
1003 					op->csum_type))
1004 				goto csum_err;
1005 			/*
1006 			 * rchecksum_bio sets compression_type on crc from op->crc,
1007 			 * this isn't always correct as sometimes we're changing
1008 			 * an extent from uncompressed to incompressible.
1009 			 */
1010 			crc.compression_type = compression_type;
1011 			crc.nonce = nonce;
1012 		} else {
1013 			if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
1014 			    bch2_rechecksum_bio(c, src, version, op->crc,
1015 					NULL, &op->crc,
1016 					src_len >> 9,
1017 					bio_sectors(src) - (src_len >> 9),
1018 					op->crc.csum_type))
1019 				goto csum_err;
1020 
1021 			crc.compressed_size	= dst_len >> 9;
1022 			crc.uncompressed_size	= src_len >> 9;
1023 			crc.live_size		= src_len >> 9;
1024 
1025 			swap(dst->bi_iter.bi_size, dst_len);
1026 			ret = bch2_encrypt_bio(c, op->csum_type,
1027 					       extent_nonce(version, crc), dst);
1028 			if (ret)
1029 				goto err;
1030 
1031 			crc.csum = bch2_checksum_bio(c, op->csum_type,
1032 					 extent_nonce(version, crc), dst);
1033 			crc.csum_type = op->csum_type;
1034 			swap(dst->bi_iter.bi_size, dst_len);
1035 		}
1036 
1037 		init_append_extent(op, wp, version, crc);
1038 
1039 		if (dst != src)
1040 			bio_advance(dst, dst_len);
1041 		bio_advance(src, src_len);
1042 		total_output	+= dst_len;
1043 		total_input	+= src_len;
1044 	} while (dst->bi_iter.bi_size &&
1045 		 src->bi_iter.bi_size &&
1046 		 wp->sectors_free &&
1047 		 !bch2_keylist_realloc(&op->insert_keys,
1048 				      op->inline_keys,
1049 				      ARRAY_SIZE(op->inline_keys),
1050 				      BKEY_EXTENT_U64s_MAX));
1051 
1052 	more = src->bi_iter.bi_size != 0;
1053 
1054 	dst->bi_iter = saved_iter;
1055 
1056 	if (dst == src && more) {
1057 		BUG_ON(total_output != total_input);
1058 
1059 		dst = bio_split(src, total_input >> 9,
1060 				GFP_NOFS, &c->bio_write);
1061 		wbio_init(dst)->put_bio	= true;
1062 		/* copy WRITE_SYNC flag */
1063 		dst->bi_opf		= src->bi_opf;
1064 	}
1065 
1066 	dst->bi_iter.bi_size = total_output;
1067 do_write:
1068 	*_dst = dst;
1069 	return more;
1070 csum_err:
1071 	bch_err(c, "%s writ error: error verifying existing checksum while rewriting existing data (memory corruption?)",
1072 		op->flags & BCH_WRITE_MOVE ? "move" : "user");
1073 	ret = -EIO;
1074 err:
1075 	if (to_wbio(dst)->bounce)
1076 		bch2_bio_free_pages_pool(c, dst);
1077 	if (to_wbio(dst)->put_bio)
1078 		bio_put(dst);
1079 
1080 	return ret;
1081 }
1082 
1083 static bool bch2_extent_is_writeable(struct bch_write_op *op,
1084 				     struct bkey_s_c k)
1085 {
1086 	struct bch_fs *c = op->c;
1087 	struct bkey_s_c_extent e;
1088 	struct extent_ptr_decoded p;
1089 	const union bch_extent_entry *entry;
1090 	unsigned replicas = 0;
1091 
1092 	if (k.k->type != KEY_TYPE_extent)
1093 		return false;
1094 
1095 	e = bkey_s_c_to_extent(k);
1096 	extent_for_each_ptr_decode(e, p, entry) {
1097 		if (crc_is_encoded(p.crc) || p.has_ec)
1098 			return false;
1099 
1100 		replicas += bch2_extent_ptr_durability(c, &p);
1101 	}
1102 
1103 	return replicas >= op->opts.data_replicas;
1104 }
1105 
1106 static inline void bch2_nocow_write_unlock(struct bch_write_op *op)
1107 {
1108 	struct bch_fs *c = op->c;
1109 
1110 	for_each_keylist_key(&op->insert_keys, k) {
1111 		struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(bkey_i_to_s_c(k));
1112 
1113 		bkey_for_each_ptr(ptrs, ptr)
1114 			bch2_bucket_nocow_unlock(&c->nocow_locks,
1115 						 PTR_BUCKET_POS(c, ptr),
1116 						 BUCKET_NOCOW_LOCK_UPDATE);
1117 	}
1118 }
1119 
1120 static int bch2_nocow_write_convert_one_unwritten(struct btree_trans *trans,
1121 						  struct btree_iter *iter,
1122 						  struct bkey_i *orig,
1123 						  struct bkey_s_c k,
1124 						  u64 new_i_size)
1125 {
1126 	if (!bch2_extents_match(bkey_i_to_s_c(orig), k)) {
1127 		/* trace this */
1128 		return 0;
1129 	}
1130 
1131 	struct bkey_i *new = bch2_bkey_make_mut_noupdate(trans, k);
1132 	int ret = PTR_ERR_OR_ZERO(new);
1133 	if (ret)
1134 		return ret;
1135 
1136 	bch2_cut_front(bkey_start_pos(&orig->k), new);
1137 	bch2_cut_back(orig->k.p, new);
1138 
1139 	struct bkey_ptrs ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
1140 	bkey_for_each_ptr(ptrs, ptr)
1141 		ptr->unwritten = 0;
1142 
1143 	/*
1144 	 * Note that we're not calling bch2_subvol_get_snapshot() in this path -
1145 	 * that was done when we kicked off the write, and here it's important
1146 	 * that we update the extent that we wrote to - even if a snapshot has
1147 	 * since been created. The write is still outstanding, so we're ok
1148 	 * w.r.t. snapshot atomicity:
1149 	 */
1150 	return  bch2_extent_update_i_size_sectors(trans, iter,
1151 					min(new->k.p.offset << 9, new_i_size), 0) ?:
1152 		bch2_trans_update(trans, iter, new,
1153 				  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
1154 }
1155 
1156 static void bch2_nocow_write_convert_unwritten(struct bch_write_op *op)
1157 {
1158 	struct bch_fs *c = op->c;
1159 	struct btree_trans *trans = bch2_trans_get(c);
1160 
1161 	for_each_keylist_key(&op->insert_keys, orig) {
1162 		int ret = for_each_btree_key_upto_commit(trans, iter, BTREE_ID_extents,
1163 				     bkey_start_pos(&orig->k), orig->k.p,
1164 				     BTREE_ITER_INTENT, k,
1165 				     NULL, NULL, BCH_TRANS_COMMIT_no_enospc, ({
1166 			bch2_nocow_write_convert_one_unwritten(trans, &iter, orig, k, op->new_i_size);
1167 		}));
1168 
1169 		if (ret && !bch2_err_matches(ret, EROFS)) {
1170 			struct bkey_i *insert = bch2_keylist_front(&op->insert_keys);
1171 
1172 			bch_err_inum_offset_ratelimited(c,
1173 				insert->k.p.inode, insert->k.p.offset << 9,
1174 				"%s write error while doing btree update: %s",
1175 				op->flags & BCH_WRITE_MOVE ? "move" : "user",
1176 				bch2_err_str(ret));
1177 		}
1178 
1179 		if (ret) {
1180 			op->error = ret;
1181 			break;
1182 		}
1183 	}
1184 
1185 	bch2_trans_put(trans);
1186 }
1187 
1188 static void __bch2_nocow_write_done(struct bch_write_op *op)
1189 {
1190 	bch2_nocow_write_unlock(op);
1191 
1192 	if (unlikely(op->flags & BCH_WRITE_IO_ERROR)) {
1193 		op->error = -EIO;
1194 	} else if (unlikely(op->flags & BCH_WRITE_CONVERT_UNWRITTEN))
1195 		bch2_nocow_write_convert_unwritten(op);
1196 }
1197 
1198 static CLOSURE_CALLBACK(bch2_nocow_write_done)
1199 {
1200 	closure_type(op, struct bch_write_op, cl);
1201 
1202 	__bch2_nocow_write_done(op);
1203 	bch2_write_done(cl);
1204 }
1205 
1206 struct bucket_to_lock {
1207 	struct bpos		b;
1208 	unsigned		gen;
1209 	struct nocow_lock_bucket *l;
1210 };
1211 
1212 static void bch2_nocow_write(struct bch_write_op *op)
1213 {
1214 	struct bch_fs *c = op->c;
1215 	struct btree_trans *trans;
1216 	struct btree_iter iter;
1217 	struct bkey_s_c k;
1218 	DARRAY_PREALLOCATED(struct bucket_to_lock, 3) buckets;
1219 	u32 snapshot;
1220 	struct bucket_to_lock *stale_at;
1221 	int ret;
1222 
1223 	if (op->flags & BCH_WRITE_MOVE)
1224 		return;
1225 
1226 	darray_init(&buckets);
1227 	trans = bch2_trans_get(c);
1228 retry:
1229 	bch2_trans_begin(trans);
1230 
1231 	ret = bch2_subvolume_get_snapshot(trans, op->subvol, &snapshot);
1232 	if (unlikely(ret))
1233 		goto err;
1234 
1235 	bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
1236 			     SPOS(op->pos.inode, op->pos.offset, snapshot),
1237 			     BTREE_ITER_SLOTS);
1238 	while (1) {
1239 		struct bio *bio = &op->wbio.bio;
1240 
1241 		buckets.nr = 0;
1242 
1243 		k = bch2_btree_iter_peek_slot(&iter);
1244 		ret = bkey_err(k);
1245 		if (ret)
1246 			break;
1247 
1248 		/* fall back to normal cow write path? */
1249 		if (unlikely(k.k->p.snapshot != snapshot ||
1250 			     !bch2_extent_is_writeable(op, k)))
1251 			break;
1252 
1253 		if (bch2_keylist_realloc(&op->insert_keys,
1254 					 op->inline_keys,
1255 					 ARRAY_SIZE(op->inline_keys),
1256 					 k.k->u64s))
1257 			break;
1258 
1259 		/* Get iorefs before dropping btree locks: */
1260 		struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1261 		bkey_for_each_ptr(ptrs, ptr) {
1262 			struct bpos b = PTR_BUCKET_POS(c, ptr);
1263 			struct nocow_lock_bucket *l =
1264 				bucket_nocow_lock(&c->nocow_locks, bucket_to_u64(b));
1265 			prefetch(l);
1266 
1267 			if (unlikely(!bch2_dev_get_ioref(bch_dev_bkey_exists(c, ptr->dev), WRITE)))
1268 				goto err_get_ioref;
1269 
1270 			/* XXX allocating memory with btree locks held - rare */
1271 			darray_push_gfp(&buckets, ((struct bucket_to_lock) {
1272 						   .b = b, .gen = ptr->gen, .l = l,
1273 						   }), GFP_KERNEL|__GFP_NOFAIL);
1274 
1275 			if (ptr->unwritten)
1276 				op->flags |= BCH_WRITE_CONVERT_UNWRITTEN;
1277 		}
1278 
1279 		/* Unlock before taking nocow locks, doing IO: */
1280 		bkey_reassemble(op->insert_keys.top, k);
1281 		bch2_trans_unlock(trans);
1282 
1283 		bch2_cut_front(op->pos, op->insert_keys.top);
1284 		if (op->flags & BCH_WRITE_CONVERT_UNWRITTEN)
1285 			bch2_cut_back(POS(op->pos.inode, op->pos.offset + bio_sectors(bio)), op->insert_keys.top);
1286 
1287 		darray_for_each(buckets, i) {
1288 			struct bch_dev *ca = bch_dev_bkey_exists(c, i->b.inode);
1289 
1290 			__bch2_bucket_nocow_lock(&c->nocow_locks, i->l,
1291 						 bucket_to_u64(i->b),
1292 						 BUCKET_NOCOW_LOCK_UPDATE);
1293 
1294 			rcu_read_lock();
1295 			bool stale = gen_after(*bucket_gen(ca, i->b.offset), i->gen);
1296 			rcu_read_unlock();
1297 
1298 			if (unlikely(stale)) {
1299 				stale_at = i;
1300 				goto err_bucket_stale;
1301 			}
1302 		}
1303 
1304 		bio = &op->wbio.bio;
1305 		if (k.k->p.offset < op->pos.offset + bio_sectors(bio)) {
1306 			bio = bio_split(bio, k.k->p.offset - op->pos.offset,
1307 					GFP_KERNEL, &c->bio_write);
1308 			wbio_init(bio)->put_bio = true;
1309 			bio->bi_opf = op->wbio.bio.bi_opf;
1310 		} else {
1311 			op->flags |= BCH_WRITE_DONE;
1312 		}
1313 
1314 		op->pos.offset += bio_sectors(bio);
1315 		op->written += bio_sectors(bio);
1316 
1317 		bio->bi_end_io	= bch2_write_endio;
1318 		bio->bi_private	= &op->cl;
1319 		bio->bi_opf |= REQ_OP_WRITE;
1320 		closure_get(&op->cl);
1321 		bch2_submit_wbio_replicas(to_wbio(bio), c, BCH_DATA_user,
1322 					  op->insert_keys.top, true);
1323 
1324 		bch2_keylist_push(&op->insert_keys);
1325 		if (op->flags & BCH_WRITE_DONE)
1326 			break;
1327 		bch2_btree_iter_advance(&iter);
1328 	}
1329 out:
1330 	bch2_trans_iter_exit(trans, &iter);
1331 err:
1332 	if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1333 		goto retry;
1334 
1335 	if (ret) {
1336 		bch_err_inum_offset_ratelimited(c,
1337 			op->pos.inode, op->pos.offset << 9,
1338 			"%s: btree lookup error %s", __func__, bch2_err_str(ret));
1339 		op->error = ret;
1340 		op->flags |= BCH_WRITE_DONE;
1341 	}
1342 
1343 	bch2_trans_put(trans);
1344 	darray_exit(&buckets);
1345 
1346 	/* fallback to cow write path? */
1347 	if (!(op->flags & BCH_WRITE_DONE)) {
1348 		closure_sync(&op->cl);
1349 		__bch2_nocow_write_done(op);
1350 		op->insert_keys.top = op->insert_keys.keys;
1351 	} else if (op->flags & BCH_WRITE_SYNC) {
1352 		closure_sync(&op->cl);
1353 		bch2_nocow_write_done(&op->cl.work);
1354 	} else {
1355 		/*
1356 		 * XXX
1357 		 * needs to run out of process context because ei_quota_lock is
1358 		 * a mutex
1359 		 */
1360 		continue_at(&op->cl, bch2_nocow_write_done, index_update_wq(op));
1361 	}
1362 	return;
1363 err_get_ioref:
1364 	darray_for_each(buckets, i)
1365 		percpu_ref_put(&bch_dev_bkey_exists(c, i->b.inode)->io_ref);
1366 
1367 	/* Fall back to COW path: */
1368 	goto out;
1369 err_bucket_stale:
1370 	darray_for_each(buckets, i) {
1371 		bch2_bucket_nocow_unlock(&c->nocow_locks, i->b, BUCKET_NOCOW_LOCK_UPDATE);
1372 		if (i == stale_at)
1373 			break;
1374 	}
1375 
1376 	/* We can retry this: */
1377 	ret = -BCH_ERR_transaction_restart;
1378 	goto err_get_ioref;
1379 }
1380 
1381 static void __bch2_write(struct bch_write_op *op)
1382 {
1383 	struct bch_fs *c = op->c;
1384 	struct write_point *wp = NULL;
1385 	struct bio *bio = NULL;
1386 	unsigned nofs_flags;
1387 	int ret;
1388 
1389 	nofs_flags = memalloc_nofs_save();
1390 
1391 	if (unlikely(op->opts.nocow && c->opts.nocow_enabled)) {
1392 		bch2_nocow_write(op);
1393 		if (op->flags & BCH_WRITE_DONE)
1394 			goto out_nofs_restore;
1395 	}
1396 again:
1397 	memset(&op->failed, 0, sizeof(op->failed));
1398 
1399 	do {
1400 		struct bkey_i *key_to_write;
1401 		unsigned key_to_write_offset = op->insert_keys.top_p -
1402 			op->insert_keys.keys_p;
1403 
1404 		/* +1 for possible cache device: */
1405 		if (op->open_buckets.nr + op->nr_replicas + 1 >
1406 		    ARRAY_SIZE(op->open_buckets.v))
1407 			break;
1408 
1409 		if (bch2_keylist_realloc(&op->insert_keys,
1410 					op->inline_keys,
1411 					ARRAY_SIZE(op->inline_keys),
1412 					BKEY_EXTENT_U64s_MAX))
1413 			break;
1414 
1415 		/*
1416 		 * The copygc thread is now global, which means it's no longer
1417 		 * freeing up space on specific disks, which means that
1418 		 * allocations for specific disks may hang arbitrarily long:
1419 		 */
1420 		ret = bch2_trans_do(c, NULL, NULL, 0,
1421 			bch2_alloc_sectors_start_trans(trans,
1422 				op->target,
1423 				op->opts.erasure_code && !(op->flags & BCH_WRITE_CACHED),
1424 				op->write_point,
1425 				&op->devs_have,
1426 				op->nr_replicas,
1427 				op->nr_replicas_required,
1428 				op->watermark,
1429 				op->flags,
1430 				(op->flags & (BCH_WRITE_ALLOC_NOWAIT|
1431 					      BCH_WRITE_ONLY_SPECIFIED_DEVS))
1432 				? NULL : &op->cl, &wp));
1433 		if (unlikely(ret)) {
1434 			if (bch2_err_matches(ret, BCH_ERR_operation_blocked))
1435 				break;
1436 
1437 			goto err;
1438 		}
1439 
1440 		EBUG_ON(!wp);
1441 
1442 		bch2_open_bucket_get(c, wp, &op->open_buckets);
1443 		ret = bch2_write_extent(op, wp, &bio);
1444 
1445 		bch2_alloc_sectors_done_inlined(c, wp);
1446 err:
1447 		if (ret <= 0) {
1448 			op->flags |= BCH_WRITE_DONE;
1449 
1450 			if (ret < 0) {
1451 				if (!(op->flags & BCH_WRITE_ALLOC_NOWAIT))
1452 					bch_err_inum_offset_ratelimited(c,
1453 						op->pos.inode,
1454 						op->pos.offset << 9,
1455 						"%s(): %s error: %s", __func__,
1456 						op->flags & BCH_WRITE_MOVE ? "move" : "user",
1457 						bch2_err_str(ret));
1458 				op->error = ret;
1459 				break;
1460 			}
1461 		}
1462 
1463 		bio->bi_end_io	= bch2_write_endio;
1464 		bio->bi_private	= &op->cl;
1465 		bio->bi_opf |= REQ_OP_WRITE;
1466 
1467 		closure_get(bio->bi_private);
1468 
1469 		key_to_write = (void *) (op->insert_keys.keys_p +
1470 					 key_to_write_offset);
1471 
1472 		bch2_submit_wbio_replicas(to_wbio(bio), c, BCH_DATA_user,
1473 					  key_to_write, false);
1474 	} while (ret);
1475 
1476 	/*
1477 	 * Sync or no?
1478 	 *
1479 	 * If we're running asynchronously, wne may still want to block
1480 	 * synchronously here if we weren't able to submit all of the IO at
1481 	 * once, as that signals backpressure to the caller.
1482 	 */
1483 	if ((op->flags & BCH_WRITE_SYNC) ||
1484 	    (!(op->flags & BCH_WRITE_DONE) &&
1485 	     !(op->flags & BCH_WRITE_IN_WORKER))) {
1486 		closure_sync(&op->cl);
1487 		__bch2_write_index(op);
1488 
1489 		if (!(op->flags & BCH_WRITE_DONE))
1490 			goto again;
1491 		bch2_write_done(&op->cl);
1492 	} else {
1493 		bch2_write_queue(op, wp);
1494 		continue_at(&op->cl, bch2_write_index, NULL);
1495 	}
1496 out_nofs_restore:
1497 	memalloc_nofs_restore(nofs_flags);
1498 }
1499 
1500 static void bch2_write_data_inline(struct bch_write_op *op, unsigned data_len)
1501 {
1502 	struct bio *bio = &op->wbio.bio;
1503 	struct bvec_iter iter;
1504 	struct bkey_i_inline_data *id;
1505 	unsigned sectors;
1506 	int ret;
1507 
1508 	op->flags |= BCH_WRITE_WROTE_DATA_INLINE;
1509 	op->flags |= BCH_WRITE_DONE;
1510 
1511 	bch2_check_set_feature(op->c, BCH_FEATURE_inline_data);
1512 
1513 	ret = bch2_keylist_realloc(&op->insert_keys, op->inline_keys,
1514 				   ARRAY_SIZE(op->inline_keys),
1515 				   BKEY_U64s + DIV_ROUND_UP(data_len, 8));
1516 	if (ret) {
1517 		op->error = ret;
1518 		goto err;
1519 	}
1520 
1521 	sectors = bio_sectors(bio);
1522 	op->pos.offset += sectors;
1523 
1524 	id = bkey_inline_data_init(op->insert_keys.top);
1525 	id->k.p		= op->pos;
1526 	id->k.version	= op->version;
1527 	id->k.size	= sectors;
1528 
1529 	iter = bio->bi_iter;
1530 	iter.bi_size = data_len;
1531 	memcpy_from_bio(id->v.data, bio, iter);
1532 
1533 	while (data_len & 7)
1534 		id->v.data[data_len++] = '\0';
1535 	set_bkey_val_bytes(&id->k, data_len);
1536 	bch2_keylist_push(&op->insert_keys);
1537 
1538 	__bch2_write_index(op);
1539 err:
1540 	bch2_write_done(&op->cl);
1541 }
1542 
1543 /**
1544  * bch2_write() - handle a write to a cache device or flash only volume
1545  * @cl:		&bch_write_op->cl
1546  *
1547  * This is the starting point for any data to end up in a cache device; it could
1548  * be from a normal write, or a writeback write, or a write to a flash only
1549  * volume - it's also used by the moving garbage collector to compact data in
1550  * mostly empty buckets.
1551  *
1552  * It first writes the data to the cache, creating a list of keys to be inserted
1553  * (if the data won't fit in a single open bucket, there will be multiple keys);
1554  * after the data is written it calls bch_journal, and after the keys have been
1555  * added to the next journal write they're inserted into the btree.
1556  *
1557  * If op->discard is true, instead of inserting the data it invalidates the
1558  * region of the cache represented by op->bio and op->inode.
1559  */
1560 CLOSURE_CALLBACK(bch2_write)
1561 {
1562 	closure_type(op, struct bch_write_op, cl);
1563 	struct bio *bio = &op->wbio.bio;
1564 	struct bch_fs *c = op->c;
1565 	unsigned data_len;
1566 
1567 	EBUG_ON(op->cl.parent);
1568 	BUG_ON(!op->nr_replicas);
1569 	BUG_ON(!op->write_point.v);
1570 	BUG_ON(bkey_eq(op->pos, POS_MAX));
1571 
1572 	op->nr_replicas_required = min_t(unsigned, op->nr_replicas_required, op->nr_replicas);
1573 	op->start_time = local_clock();
1574 	bch2_keylist_init(&op->insert_keys, op->inline_keys);
1575 	wbio_init(bio)->put_bio = false;
1576 
1577 	if (bio->bi_iter.bi_size & (c->opts.block_size - 1)) {
1578 		bch_err_inum_offset_ratelimited(c,
1579 			op->pos.inode,
1580 			op->pos.offset << 9,
1581 			"%s write error: misaligned write",
1582 			op->flags & BCH_WRITE_MOVE ? "move" : "user");
1583 		op->error = -EIO;
1584 		goto err;
1585 	}
1586 
1587 	if (c->opts.nochanges) {
1588 		op->error = -BCH_ERR_erofs_no_writes;
1589 		goto err;
1590 	}
1591 
1592 	if (!(op->flags & BCH_WRITE_MOVE) &&
1593 	    !bch2_write_ref_tryget(c, BCH_WRITE_REF_write)) {
1594 		op->error = -BCH_ERR_erofs_no_writes;
1595 		goto err;
1596 	}
1597 
1598 	this_cpu_add(c->counters[BCH_COUNTER_io_write], bio_sectors(bio));
1599 	bch2_increment_clock(c, bio_sectors(bio), WRITE);
1600 
1601 	data_len = min_t(u64, bio->bi_iter.bi_size,
1602 			 op->new_i_size - (op->pos.offset << 9));
1603 
1604 	if (c->opts.inline_data &&
1605 	    data_len <= min(block_bytes(c) / 2, 1024U)) {
1606 		bch2_write_data_inline(op, data_len);
1607 		return;
1608 	}
1609 
1610 	__bch2_write(op);
1611 	return;
1612 err:
1613 	bch2_disk_reservation_put(c, &op->res);
1614 
1615 	closure_debug_destroy(&op->cl);
1616 	if (op->end_io)
1617 		op->end_io(op);
1618 }
1619 
1620 static const char * const bch2_write_flags[] = {
1621 #define x(f)	#f,
1622 	BCH_WRITE_FLAGS()
1623 #undef x
1624 	NULL
1625 };
1626 
1627 void bch2_write_op_to_text(struct printbuf *out, struct bch_write_op *op)
1628 {
1629 	prt_str(out, "pos: ");
1630 	bch2_bpos_to_text(out, op->pos);
1631 	prt_newline(out);
1632 	printbuf_indent_add(out, 2);
1633 
1634 	prt_str(out, "started: ");
1635 	bch2_pr_time_units(out, local_clock() - op->start_time);
1636 	prt_newline(out);
1637 
1638 	prt_str(out, "flags: ");
1639 	prt_bitflags(out, bch2_write_flags, op->flags);
1640 	prt_newline(out);
1641 
1642 	prt_printf(out, "ref: %u", closure_nr_remaining(&op->cl));
1643 	prt_newline(out);
1644 
1645 	printbuf_indent_sub(out, 2);
1646 }
1647 
1648 void bch2_fs_io_write_exit(struct bch_fs *c)
1649 {
1650 	mempool_exit(&c->bio_bounce_pages);
1651 	bioset_exit(&c->bio_write);
1652 }
1653 
1654 int bch2_fs_io_write_init(struct bch_fs *c)
1655 {
1656 	if (bioset_init(&c->bio_write, 1, offsetof(struct bch_write_bio, bio),
1657 			BIOSET_NEED_BVECS))
1658 		return -BCH_ERR_ENOMEM_bio_write_init;
1659 
1660 	if (mempool_init_page_pool(&c->bio_bounce_pages,
1661 				   max_t(unsigned,
1662 					 c->opts.btree_node_size,
1663 					 c->opts.encoded_extent_max) /
1664 				   PAGE_SIZE, 0))
1665 		return -BCH_ERR_ENOMEM_bio_bounce_pages_init;
1666 
1667 	return 0;
1668 }
1669