xref: /linux/fs/bcachefs/io_write.c (revision cb787f4ac0c2e439ea8d7e6387b925f74576bdf8)
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;
73 
74 	old = atomic64_read(latency);
75 	do {
76 		/*
77 		 * If the io latency was reasonably close to the current
78 		 * latency, skip doing the update and atomic operation - most of
79 		 * the time:
80 		 */
81 		if (abs((int) (old - io_latency)) < (old >> 1) &&
82 		    now & ~(~0U << 5))
83 			break;
84 
85 		new = ewma_add(old, io_latency, 5);
86 	} while (!atomic64_try_cmpxchg(latency, &old, new));
87 
88 	bch2_congested_acct(ca, io_latency, now, rw);
89 
90 	__bch2_time_stats_update(&ca->io_latency[rw].stats, submit_time, now);
91 }
92 
93 #endif
94 
95 /* Allocate, free from mempool: */
96 
97 void bch2_bio_free_pages_pool(struct bch_fs *c, struct bio *bio)
98 {
99 	struct bvec_iter_all iter;
100 	struct bio_vec *bv;
101 
102 	bio_for_each_segment_all(bv, bio, iter)
103 		if (bv->bv_page != ZERO_PAGE(0))
104 			mempool_free(bv->bv_page, &c->bio_bounce_pages);
105 	bio->bi_vcnt = 0;
106 }
107 
108 static struct page *__bio_alloc_page_pool(struct bch_fs *c, bool *using_mempool)
109 {
110 	struct page *page;
111 
112 	if (likely(!*using_mempool)) {
113 		page = alloc_page(GFP_NOFS);
114 		if (unlikely(!page)) {
115 			mutex_lock(&c->bio_bounce_pages_lock);
116 			*using_mempool = true;
117 			goto pool_alloc;
118 
119 		}
120 	} else {
121 pool_alloc:
122 		page = mempool_alloc(&c->bio_bounce_pages, GFP_NOFS);
123 	}
124 
125 	return page;
126 }
127 
128 void bch2_bio_alloc_pages_pool(struct bch_fs *c, struct bio *bio,
129 			       size_t size)
130 {
131 	bool using_mempool = false;
132 
133 	while (size) {
134 		struct page *page = __bio_alloc_page_pool(c, &using_mempool);
135 		unsigned len = min_t(size_t, PAGE_SIZE, size);
136 
137 		BUG_ON(!bio_add_page(bio, page, len, 0));
138 		size -= len;
139 	}
140 
141 	if (using_mempool)
142 		mutex_unlock(&c->bio_bounce_pages_lock);
143 }
144 
145 /* Extent update path: */
146 
147 int bch2_sum_sector_overwrites(struct btree_trans *trans,
148 			       struct btree_iter *extent_iter,
149 			       struct bkey_i *new,
150 			       bool *usage_increasing,
151 			       s64 *i_sectors_delta,
152 			       s64 *disk_sectors_delta)
153 {
154 	struct bch_fs *c = trans->c;
155 	struct btree_iter iter;
156 	struct bkey_s_c old;
157 	unsigned new_replicas = bch2_bkey_replicas(c, bkey_i_to_s_c(new));
158 	bool new_compressed = bch2_bkey_sectors_compressed(bkey_i_to_s_c(new));
159 	int ret = 0;
160 
161 	*usage_increasing	= false;
162 	*i_sectors_delta	= 0;
163 	*disk_sectors_delta	= 0;
164 
165 	bch2_trans_copy_iter(&iter, extent_iter);
166 
167 	for_each_btree_key_upto_continue_norestart(iter,
168 				new->k.p, BTREE_ITER_slots, old, ret) {
169 		s64 sectors = min(new->k.p.offset, old.k->p.offset) -
170 			max(bkey_start_offset(&new->k),
171 			    bkey_start_offset(old.k));
172 
173 		*i_sectors_delta += sectors *
174 			(bkey_extent_is_allocation(&new->k) -
175 			 bkey_extent_is_allocation(old.k));
176 
177 		*disk_sectors_delta += sectors * bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(new));
178 		*disk_sectors_delta -= new->k.p.snapshot == old.k->p.snapshot
179 			? sectors * bch2_bkey_nr_ptrs_fully_allocated(old)
180 			: 0;
181 
182 		if (!*usage_increasing &&
183 		    (new->k.p.snapshot != old.k->p.snapshot ||
184 		     new_replicas > bch2_bkey_replicas(c, old) ||
185 		     (!new_compressed && bch2_bkey_sectors_compressed(old))))
186 			*usage_increasing = true;
187 
188 		if (bkey_ge(old.k->p, new->k.p))
189 			break;
190 	}
191 
192 	bch2_trans_iter_exit(trans, &iter);
193 	return ret;
194 }
195 
196 static inline int bch2_extent_update_i_size_sectors(struct btree_trans *trans,
197 						    struct btree_iter *extent_iter,
198 						    u64 new_i_size,
199 						    s64 i_sectors_delta)
200 {
201 	/*
202 	 * Crazy performance optimization:
203 	 * Every extent update needs to also update the inode: the inode trigger
204 	 * will set bi->journal_seq to the journal sequence number of this
205 	 * transaction - for fsync.
206 	 *
207 	 * But if that's the only reason we're updating the inode (we're not
208 	 * updating bi_size or bi_sectors), then we don't need the inode update
209 	 * to be journalled - if we crash, the bi_journal_seq update will be
210 	 * lost, but that's fine.
211 	 */
212 	unsigned inode_update_flags = BTREE_UPDATE_nojournal;
213 
214 	struct btree_iter iter;
215 	struct bkey_s_c k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_inodes,
216 			      SPOS(0,
217 				   extent_iter->pos.inode,
218 				   extent_iter->snapshot),
219 			      BTREE_ITER_cached);
220 	int ret = bkey_err(k);
221 	if (unlikely(ret))
222 		return ret;
223 
224 	/*
225 	 * varint_decode_fast(), in the inode .invalid method, reads up to 7
226 	 * bytes past the end of the buffer:
227 	 */
228 	struct bkey_i *k_mut = bch2_trans_kmalloc_nomemzero(trans, bkey_bytes(k.k) + 8);
229 	ret = PTR_ERR_OR_ZERO(k_mut);
230 	if (unlikely(ret))
231 		goto err;
232 
233 	bkey_reassemble(k_mut, k);
234 
235 	if (unlikely(k_mut->k.type != KEY_TYPE_inode_v3)) {
236 		k_mut = bch2_inode_to_v3(trans, k_mut);
237 		ret = PTR_ERR_OR_ZERO(k_mut);
238 		if (unlikely(ret))
239 			goto err;
240 	}
241 
242 	struct bkey_i_inode_v3 *inode = bkey_i_to_inode_v3(k_mut);
243 
244 	if (!(le64_to_cpu(inode->v.bi_flags) & BCH_INODE_i_size_dirty) &&
245 	    new_i_size > le64_to_cpu(inode->v.bi_size)) {
246 		inode->v.bi_size = cpu_to_le64(new_i_size);
247 		inode_update_flags = 0;
248 	}
249 
250 	if (i_sectors_delta) {
251 		le64_add_cpu(&inode->v.bi_sectors, i_sectors_delta);
252 		inode_update_flags = 0;
253 	}
254 
255 	if (inode->k.p.snapshot != iter.snapshot) {
256 		inode->k.p.snapshot = iter.snapshot;
257 		inode_update_flags = 0;
258 	}
259 
260 	ret = bch2_trans_update(trans, &iter, &inode->k_i,
261 				BTREE_UPDATE_internal_snapshot_node|
262 				inode_update_flags);
263 err:
264 	bch2_trans_iter_exit(trans, &iter);
265 	return ret;
266 }
267 
268 int bch2_extent_update(struct btree_trans *trans,
269 		       subvol_inum inum,
270 		       struct btree_iter *iter,
271 		       struct bkey_i *k,
272 		       struct disk_reservation *disk_res,
273 		       u64 new_i_size,
274 		       s64 *i_sectors_delta_total,
275 		       bool check_enospc)
276 {
277 	struct bpos next_pos;
278 	bool usage_increasing;
279 	s64 i_sectors_delta = 0, disk_sectors_delta = 0;
280 	int ret;
281 
282 	/*
283 	 * This traverses us the iterator without changing iter->path->pos to
284 	 * search_key() (which is pos + 1 for extents): we want there to be a
285 	 * path already traversed at iter->pos because
286 	 * bch2_trans_extent_update() will use it to attempt extent merging
287 	 */
288 	ret = __bch2_btree_iter_traverse(iter);
289 	if (ret)
290 		return ret;
291 
292 	ret = bch2_extent_trim_atomic(trans, iter, k);
293 	if (ret)
294 		return ret;
295 
296 	next_pos = k->k.p;
297 
298 	ret = bch2_sum_sector_overwrites(trans, iter, k,
299 			&usage_increasing,
300 			&i_sectors_delta,
301 			&disk_sectors_delta);
302 	if (ret)
303 		return ret;
304 
305 	if (disk_res &&
306 	    disk_sectors_delta > (s64) disk_res->sectors) {
307 		ret = bch2_disk_reservation_add(trans->c, disk_res,
308 					disk_sectors_delta - disk_res->sectors,
309 					!check_enospc || !usage_increasing
310 					? BCH_DISK_RESERVATION_NOFAIL : 0);
311 		if (ret)
312 			return ret;
313 	}
314 
315 	/*
316 	 * Note:
317 	 * We always have to do an inode update - even when i_size/i_sectors
318 	 * aren't changing - for fsync to work properly; fsync relies on
319 	 * inode->bi_journal_seq which is updated by the trigger code:
320 	 */
321 	ret =   bch2_extent_update_i_size_sectors(trans, iter,
322 						  min(k->k.p.offset << 9, new_i_size),
323 						  i_sectors_delta) ?:
324 		bch2_trans_update(trans, iter, k, 0) ?:
325 		bch2_trans_commit(trans, disk_res, NULL,
326 				BCH_TRANS_COMMIT_no_check_rw|
327 				BCH_TRANS_COMMIT_no_enospc);
328 	if (unlikely(ret))
329 		return ret;
330 
331 	if (i_sectors_delta_total)
332 		*i_sectors_delta_total += i_sectors_delta;
333 	bch2_btree_iter_set_pos(iter, next_pos);
334 	return 0;
335 }
336 
337 static int bch2_write_index_default(struct bch_write_op *op)
338 {
339 	struct bch_fs *c = op->c;
340 	struct bkey_buf sk;
341 	struct keylist *keys = &op->insert_keys;
342 	struct bkey_i *k = bch2_keylist_front(keys);
343 	struct btree_trans *trans = bch2_trans_get(c);
344 	struct btree_iter iter;
345 	subvol_inum inum = {
346 		.subvol = op->subvol,
347 		.inum	= k->k.p.inode,
348 	};
349 	int ret;
350 
351 	BUG_ON(!inum.subvol);
352 
353 	bch2_bkey_buf_init(&sk);
354 
355 	do {
356 		bch2_trans_begin(trans);
357 
358 		k = bch2_keylist_front(keys);
359 		bch2_bkey_buf_copy(&sk, c, k);
360 
361 		ret = bch2_subvolume_get_snapshot(trans, inum.subvol,
362 						  &sk.k->k.p.snapshot);
363 		if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
364 			continue;
365 		if (ret)
366 			break;
367 
368 		bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
369 				     bkey_start_pos(&sk.k->k),
370 				     BTREE_ITER_slots|BTREE_ITER_intent);
371 
372 		ret =   bch2_bkey_set_needs_rebalance(c, sk.k, &op->opts) ?:
373 			bch2_extent_update(trans, inum, &iter, sk.k,
374 					&op->res,
375 					op->new_i_size, &op->i_sectors_delta,
376 					op->flags & BCH_WRITE_CHECK_ENOSPC);
377 		bch2_trans_iter_exit(trans, &iter);
378 
379 		if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
380 			continue;
381 		if (ret)
382 			break;
383 
384 		if (bkey_ge(iter.pos, k->k.p))
385 			bch2_keylist_pop_front(&op->insert_keys);
386 		else
387 			bch2_cut_front(iter.pos, k);
388 	} while (!bch2_keylist_empty(keys));
389 
390 	bch2_trans_put(trans);
391 	bch2_bkey_buf_exit(&sk, c);
392 
393 	return ret;
394 }
395 
396 /* Writes */
397 
398 void bch2_submit_wbio_replicas(struct bch_write_bio *wbio, struct bch_fs *c,
399 			       enum bch_data_type type,
400 			       const struct bkey_i *k,
401 			       bool nocow)
402 {
403 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(bkey_i_to_s_c(k));
404 	struct bch_write_bio *n;
405 
406 	BUG_ON(c->opts.nochanges);
407 
408 	bkey_for_each_ptr(ptrs, ptr) {
409 		struct bch_dev *ca = nocow
410 			? bch2_dev_have_ref(c, ptr->dev)
411 			: bch2_dev_get_ioref(c, ptr->dev, type == BCH_DATA_btree ? READ : WRITE);
412 
413 		if (to_entry(ptr + 1) < ptrs.end) {
414 			n = to_wbio(bio_alloc_clone(NULL, &wbio->bio, GFP_NOFS, &c->replica_set));
415 
416 			n->bio.bi_end_io	= wbio->bio.bi_end_io;
417 			n->bio.bi_private	= wbio->bio.bi_private;
418 			n->parent		= wbio;
419 			n->split		= true;
420 			n->bounce		= false;
421 			n->put_bio		= true;
422 			n->bio.bi_opf		= wbio->bio.bi_opf;
423 			bio_inc_remaining(&wbio->bio);
424 		} else {
425 			n = wbio;
426 			n->split		= false;
427 		}
428 
429 		n->c			= c;
430 		n->dev			= ptr->dev;
431 		n->have_ioref		= ca != NULL;
432 		n->nocow		= nocow;
433 		n->submit_time		= local_clock();
434 		n->inode_offset		= bkey_start_offset(&k->k);
435 		if (nocow)
436 			n->nocow_bucket	= PTR_BUCKET_NR(ca, ptr);
437 		n->bio.bi_iter.bi_sector = ptr->offset;
438 
439 		if (likely(n->have_ioref)) {
440 			this_cpu_add(ca->io_done->sectors[WRITE][type],
441 				     bio_sectors(&n->bio));
442 
443 			bio_set_dev(&n->bio, ca->disk_sb.bdev);
444 
445 			if (type != BCH_DATA_btree && unlikely(c->opts.no_data_io)) {
446 				bio_endio(&n->bio);
447 				continue;
448 			}
449 
450 			submit_bio(&n->bio);
451 		} else {
452 			n->bio.bi_status	= BLK_STS_REMOVED;
453 			bio_endio(&n->bio);
454 		}
455 	}
456 }
457 
458 static void __bch2_write(struct bch_write_op *);
459 
460 static void bch2_write_done(struct closure *cl)
461 {
462 	struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
463 	struct bch_fs *c = op->c;
464 
465 	EBUG_ON(op->open_buckets.nr);
466 
467 	bch2_time_stats_update(&c->times[BCH_TIME_data_write], op->start_time);
468 	bch2_disk_reservation_put(c, &op->res);
469 
470 	if (!(op->flags & BCH_WRITE_MOVE))
471 		bch2_write_ref_put(c, BCH_WRITE_REF_write);
472 	bch2_keylist_free(&op->insert_keys, op->inline_keys);
473 
474 	EBUG_ON(cl->parent);
475 	closure_debug_destroy(cl);
476 	if (op->end_io)
477 		op->end_io(op);
478 }
479 
480 static noinline int bch2_write_drop_io_error_ptrs(struct bch_write_op *op)
481 {
482 	struct keylist *keys = &op->insert_keys;
483 	struct bkey_i *src, *dst = keys->keys, *n;
484 
485 	for (src = keys->keys; src != keys->top; src = n) {
486 		n = bkey_next(src);
487 
488 		if (bkey_extent_is_direct_data(&src->k)) {
489 			bch2_bkey_drop_ptrs(bkey_i_to_s(src), ptr,
490 					    test_bit(ptr->dev, op->failed.d));
491 
492 			if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(src)))
493 				return -EIO;
494 		}
495 
496 		if (dst != src)
497 			memmove_u64s_down(dst, src, src->k.u64s);
498 		dst = bkey_next(dst);
499 	}
500 
501 	keys->top = dst;
502 	return 0;
503 }
504 
505 /**
506  * __bch2_write_index - after a write, update index to point to new data
507  * @op:		bch_write_op to process
508  */
509 static void __bch2_write_index(struct bch_write_op *op)
510 {
511 	struct bch_fs *c = op->c;
512 	struct keylist *keys = &op->insert_keys;
513 	unsigned dev;
514 	int ret = 0;
515 
516 	if (unlikely(op->flags & BCH_WRITE_IO_ERROR)) {
517 		ret = bch2_write_drop_io_error_ptrs(op);
518 		if (ret)
519 			goto err;
520 	}
521 
522 	if (!bch2_keylist_empty(keys)) {
523 		u64 sectors_start = keylist_sectors(keys);
524 
525 		ret = !(op->flags & BCH_WRITE_MOVE)
526 			? bch2_write_index_default(op)
527 			: bch2_data_update_index_update(op);
528 
529 		BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart));
530 		BUG_ON(keylist_sectors(keys) && !ret);
531 
532 		op->written += sectors_start - keylist_sectors(keys);
533 
534 		if (ret && !bch2_err_matches(ret, EROFS)) {
535 			struct bkey_i *insert = bch2_keylist_front(&op->insert_keys);
536 
537 			bch_err_inum_offset_ratelimited(c,
538 				insert->k.p.inode, insert->k.p.offset << 9,
539 				"%s write error while doing btree update: %s",
540 				op->flags & BCH_WRITE_MOVE ? "move" : "user",
541 				bch2_err_str(ret));
542 		}
543 
544 		if (ret)
545 			goto err;
546 	}
547 out:
548 	/* If some a bucket wasn't written, we can't erasure code it: */
549 	for_each_set_bit(dev, op->failed.d, BCH_SB_MEMBERS_MAX)
550 		bch2_open_bucket_write_error(c, &op->open_buckets, dev);
551 
552 	bch2_open_buckets_put(c, &op->open_buckets);
553 	return;
554 err:
555 	keys->top = keys->keys;
556 	op->error = ret;
557 	op->flags |= BCH_WRITE_SUBMITTED;
558 	goto out;
559 }
560 
561 static inline void __wp_update_state(struct write_point *wp, enum write_point_state state)
562 {
563 	if (state != wp->state) {
564 		u64 now = ktime_get_ns();
565 
566 		if (wp->last_state_change &&
567 		    time_after64(now, wp->last_state_change))
568 			wp->time[wp->state] += now - wp->last_state_change;
569 		wp->state = state;
570 		wp->last_state_change = now;
571 	}
572 }
573 
574 static inline void wp_update_state(struct write_point *wp, bool running)
575 {
576 	enum write_point_state state;
577 
578 	state = running			 ? WRITE_POINT_running :
579 		!list_empty(&wp->writes) ? WRITE_POINT_waiting_io
580 					 : WRITE_POINT_stopped;
581 
582 	__wp_update_state(wp, state);
583 }
584 
585 static CLOSURE_CALLBACK(bch2_write_index)
586 {
587 	closure_type(op, struct bch_write_op, cl);
588 	struct write_point *wp = op->wp;
589 	struct workqueue_struct *wq = index_update_wq(op);
590 	unsigned long flags;
591 
592 	if ((op->flags & BCH_WRITE_SUBMITTED) &&
593 	    (op->flags & BCH_WRITE_MOVE))
594 		bch2_bio_free_pages_pool(op->c, &op->wbio.bio);
595 
596 	spin_lock_irqsave(&wp->writes_lock, flags);
597 	if (wp->state == WRITE_POINT_waiting_io)
598 		__wp_update_state(wp, WRITE_POINT_waiting_work);
599 	list_add_tail(&op->wp_list, &wp->writes);
600 	spin_unlock_irqrestore (&wp->writes_lock, flags);
601 
602 	queue_work(wq, &wp->index_update_work);
603 }
604 
605 static inline void bch2_write_queue(struct bch_write_op *op, struct write_point *wp)
606 {
607 	op->wp = wp;
608 
609 	if (wp->state == WRITE_POINT_stopped) {
610 		spin_lock_irq(&wp->writes_lock);
611 		__wp_update_state(wp, WRITE_POINT_waiting_io);
612 		spin_unlock_irq(&wp->writes_lock);
613 	}
614 }
615 
616 void bch2_write_point_do_index_updates(struct work_struct *work)
617 {
618 	struct write_point *wp =
619 		container_of(work, struct write_point, index_update_work);
620 	struct bch_write_op *op;
621 
622 	while (1) {
623 		spin_lock_irq(&wp->writes_lock);
624 		op = list_first_entry_or_null(&wp->writes, struct bch_write_op, wp_list);
625 		if (op)
626 			list_del(&op->wp_list);
627 		wp_update_state(wp, op != NULL);
628 		spin_unlock_irq(&wp->writes_lock);
629 
630 		if (!op)
631 			break;
632 
633 		op->flags |= BCH_WRITE_IN_WORKER;
634 
635 		__bch2_write_index(op);
636 
637 		if (!(op->flags & BCH_WRITE_SUBMITTED))
638 			__bch2_write(op);
639 		else
640 			bch2_write_done(&op->cl);
641 	}
642 }
643 
644 static void bch2_write_endio(struct bio *bio)
645 {
646 	struct closure *cl		= bio->bi_private;
647 	struct bch_write_op *op		= container_of(cl, struct bch_write_op, cl);
648 	struct bch_write_bio *wbio	= to_wbio(bio);
649 	struct bch_write_bio *parent	= wbio->split ? wbio->parent : NULL;
650 	struct bch_fs *c		= wbio->c;
651 	struct bch_dev *ca		= wbio->have_ioref
652 		? bch2_dev_have_ref(c, wbio->dev)
653 		: NULL;
654 
655 	if (bch2_dev_inum_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_write,
656 				    op->pos.inode,
657 				    wbio->inode_offset << 9,
658 				    "data write error: %s",
659 				    bch2_blk_status_to_str(bio->bi_status))) {
660 		set_bit(wbio->dev, op->failed.d);
661 		op->flags |= BCH_WRITE_IO_ERROR;
662 	}
663 
664 	if (wbio->nocow) {
665 		bch2_bucket_nocow_unlock(&c->nocow_locks,
666 					 POS(ca->dev_idx, wbio->nocow_bucket),
667 					 BUCKET_NOCOW_LOCK_UPDATE);
668 		set_bit(wbio->dev, op->devs_need_flush->d);
669 	}
670 
671 	if (wbio->have_ioref) {
672 		bch2_latency_acct(ca, wbio->submit_time, WRITE);
673 		percpu_ref_put(&ca->io_ref);
674 	}
675 
676 	if (wbio->bounce)
677 		bch2_bio_free_pages_pool(c, bio);
678 
679 	if (wbio->put_bio)
680 		bio_put(bio);
681 
682 	if (parent)
683 		bio_endio(&parent->bio);
684 	else
685 		closure_put(cl);
686 }
687 
688 static void init_append_extent(struct bch_write_op *op,
689 			       struct write_point *wp,
690 			       struct bversion version,
691 			       struct bch_extent_crc_unpacked crc)
692 {
693 	struct bkey_i_extent *e;
694 
695 	op->pos.offset += crc.uncompressed_size;
696 
697 	e = bkey_extent_init(op->insert_keys.top);
698 	e->k.p		= op->pos;
699 	e->k.size	= crc.uncompressed_size;
700 	e->k.version	= version;
701 
702 	if (crc.csum_type ||
703 	    crc.compression_type ||
704 	    crc.nonce)
705 		bch2_extent_crc_append(&e->k_i, crc);
706 
707 	bch2_alloc_sectors_append_ptrs_inlined(op->c, wp, &e->k_i, crc.compressed_size,
708 				       op->flags & BCH_WRITE_CACHED);
709 
710 	bch2_keylist_push(&op->insert_keys);
711 }
712 
713 static struct bio *bch2_write_bio_alloc(struct bch_fs *c,
714 					struct write_point *wp,
715 					struct bio *src,
716 					bool *page_alloc_failed,
717 					void *buf)
718 {
719 	struct bch_write_bio *wbio;
720 	struct bio *bio;
721 	unsigned output_available =
722 		min(wp->sectors_free << 9, src->bi_iter.bi_size);
723 	unsigned pages = DIV_ROUND_UP(output_available +
724 				      (buf
725 				       ? ((unsigned long) buf & (PAGE_SIZE - 1))
726 				       : 0), PAGE_SIZE);
727 
728 	pages = min(pages, BIO_MAX_VECS);
729 
730 	bio = bio_alloc_bioset(NULL, pages, 0,
731 			       GFP_NOFS, &c->bio_write);
732 	wbio			= wbio_init(bio);
733 	wbio->put_bio		= true;
734 	/* copy WRITE_SYNC flag */
735 	wbio->bio.bi_opf	= src->bi_opf;
736 
737 	if (buf) {
738 		bch2_bio_map(bio, buf, output_available);
739 		return bio;
740 	}
741 
742 	wbio->bounce		= true;
743 
744 	/*
745 	 * We can't use mempool for more than c->sb.encoded_extent_max
746 	 * worth of pages, but we'd like to allocate more if we can:
747 	 */
748 	bch2_bio_alloc_pages_pool(c, bio,
749 				  min_t(unsigned, output_available,
750 					c->opts.encoded_extent_max));
751 
752 	if (bio->bi_iter.bi_size < output_available)
753 		*page_alloc_failed =
754 			bch2_bio_alloc_pages(bio,
755 					     output_available -
756 					     bio->bi_iter.bi_size,
757 					     GFP_NOFS) != 0;
758 
759 	return bio;
760 }
761 
762 static int bch2_write_rechecksum(struct bch_fs *c,
763 				 struct bch_write_op *op,
764 				 unsigned new_csum_type)
765 {
766 	struct bio *bio = &op->wbio.bio;
767 	struct bch_extent_crc_unpacked new_crc;
768 	int ret;
769 
770 	/* bch2_rechecksum_bio() can't encrypt or decrypt data: */
771 
772 	if (bch2_csum_type_is_encryption(op->crc.csum_type) !=
773 	    bch2_csum_type_is_encryption(new_csum_type))
774 		new_csum_type = op->crc.csum_type;
775 
776 	ret = bch2_rechecksum_bio(c, bio, op->version, op->crc,
777 				  NULL, &new_crc,
778 				  op->crc.offset, op->crc.live_size,
779 				  new_csum_type);
780 	if (ret)
781 		return ret;
782 
783 	bio_advance(bio, op->crc.offset << 9);
784 	bio->bi_iter.bi_size = op->crc.live_size << 9;
785 	op->crc = new_crc;
786 	return 0;
787 }
788 
789 static int bch2_write_decrypt(struct bch_write_op *op)
790 {
791 	struct bch_fs *c = op->c;
792 	struct nonce nonce = extent_nonce(op->version, op->crc);
793 	struct bch_csum csum;
794 	int ret;
795 
796 	if (!bch2_csum_type_is_encryption(op->crc.csum_type))
797 		return 0;
798 
799 	/*
800 	 * If we need to decrypt data in the write path, we'll no longer be able
801 	 * to verify the existing checksum (poly1305 mac, in this case) after
802 	 * it's decrypted - this is the last point we'll be able to reverify the
803 	 * checksum:
804 	 */
805 	csum = bch2_checksum_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
806 	if (bch2_crc_cmp(op->crc.csum, csum) && !c->opts.no_data_io)
807 		return -EIO;
808 
809 	ret = bch2_encrypt_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
810 	op->crc.csum_type = 0;
811 	op->crc.csum = (struct bch_csum) { 0, 0 };
812 	return ret;
813 }
814 
815 static enum prep_encoded_ret {
816 	PREP_ENCODED_OK,
817 	PREP_ENCODED_ERR,
818 	PREP_ENCODED_CHECKSUM_ERR,
819 	PREP_ENCODED_DO_WRITE,
820 } bch2_write_prep_encoded_data(struct bch_write_op *op, struct write_point *wp)
821 {
822 	struct bch_fs *c = op->c;
823 	struct bio *bio = &op->wbio.bio;
824 
825 	if (!(op->flags & BCH_WRITE_DATA_ENCODED))
826 		return PREP_ENCODED_OK;
827 
828 	BUG_ON(bio_sectors(bio) != op->crc.compressed_size);
829 
830 	/* Can we just write the entire extent as is? */
831 	if (op->crc.uncompressed_size == op->crc.live_size &&
832 	    op->crc.uncompressed_size <= c->opts.encoded_extent_max >> 9 &&
833 	    op->crc.compressed_size <= wp->sectors_free &&
834 	    (op->crc.compression_type == bch2_compression_opt_to_type(op->compression_opt) ||
835 	     op->incompressible)) {
836 		if (!crc_is_compressed(op->crc) &&
837 		    op->csum_type != op->crc.csum_type &&
838 		    bch2_write_rechecksum(c, op, op->csum_type) &&
839 		    !c->opts.no_data_io)
840 			return PREP_ENCODED_CHECKSUM_ERR;
841 
842 		return PREP_ENCODED_DO_WRITE;
843 	}
844 
845 	/*
846 	 * If the data is compressed and we couldn't write the entire extent as
847 	 * is, we have to decompress it:
848 	 */
849 	if (crc_is_compressed(op->crc)) {
850 		struct bch_csum csum;
851 
852 		if (bch2_write_decrypt(op))
853 			return PREP_ENCODED_CHECKSUM_ERR;
854 
855 		/* Last point we can still verify checksum: */
856 		csum = bch2_checksum_bio(c, op->crc.csum_type,
857 					 extent_nonce(op->version, op->crc),
858 					 bio);
859 		if (bch2_crc_cmp(op->crc.csum, csum) && !c->opts.no_data_io)
860 			return PREP_ENCODED_CHECKSUM_ERR;
861 
862 		if (bch2_bio_uncompress_inplace(c, bio, &op->crc))
863 			return PREP_ENCODED_ERR;
864 	}
865 
866 	/*
867 	 * No longer have compressed data after this point - data might be
868 	 * encrypted:
869 	 */
870 
871 	/*
872 	 * If the data is checksummed and we're only writing a subset,
873 	 * rechecksum and adjust bio to point to currently live data:
874 	 */
875 	if ((op->crc.live_size != op->crc.uncompressed_size ||
876 	     op->crc.csum_type != op->csum_type) &&
877 	    bch2_write_rechecksum(c, op, op->csum_type) &&
878 	    !c->opts.no_data_io)
879 		return PREP_ENCODED_CHECKSUM_ERR;
880 
881 	/*
882 	 * If we want to compress the data, it has to be decrypted:
883 	 */
884 	if ((op->compression_opt ||
885 	     bch2_csum_type_is_encryption(op->crc.csum_type) !=
886 	     bch2_csum_type_is_encryption(op->csum_type)) &&
887 	    bch2_write_decrypt(op))
888 		return PREP_ENCODED_CHECKSUM_ERR;
889 
890 	return PREP_ENCODED_OK;
891 }
892 
893 static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp,
894 			     struct bio **_dst)
895 {
896 	struct bch_fs *c = op->c;
897 	struct bio *src = &op->wbio.bio, *dst = src;
898 	struct bvec_iter saved_iter;
899 	void *ec_buf;
900 	unsigned total_output = 0, total_input = 0;
901 	bool bounce = false;
902 	bool page_alloc_failed = false;
903 	int ret, more = 0;
904 
905 	BUG_ON(!bio_sectors(src));
906 
907 	ec_buf = bch2_writepoint_ec_buf(c, wp);
908 
909 	switch (bch2_write_prep_encoded_data(op, wp)) {
910 	case PREP_ENCODED_OK:
911 		break;
912 	case PREP_ENCODED_ERR:
913 		ret = -EIO;
914 		goto err;
915 	case PREP_ENCODED_CHECKSUM_ERR:
916 		goto csum_err;
917 	case PREP_ENCODED_DO_WRITE:
918 		/* XXX look for bug here */
919 		if (ec_buf) {
920 			dst = bch2_write_bio_alloc(c, wp, src,
921 						   &page_alloc_failed,
922 						   ec_buf);
923 			bio_copy_data(dst, src);
924 			bounce = true;
925 		}
926 		init_append_extent(op, wp, op->version, op->crc);
927 		goto do_write;
928 	}
929 
930 	if (ec_buf ||
931 	    op->compression_opt ||
932 	    (op->csum_type &&
933 	     !(op->flags & BCH_WRITE_PAGES_STABLE)) ||
934 	    (bch2_csum_type_is_encryption(op->csum_type) &&
935 	     !(op->flags & BCH_WRITE_PAGES_OWNED))) {
936 		dst = bch2_write_bio_alloc(c, wp, src,
937 					   &page_alloc_failed,
938 					   ec_buf);
939 		bounce = true;
940 	}
941 
942 	saved_iter = dst->bi_iter;
943 
944 	do {
945 		struct bch_extent_crc_unpacked crc = { 0 };
946 		struct bversion version = op->version;
947 		size_t dst_len = 0, src_len = 0;
948 
949 		if (page_alloc_failed &&
950 		    dst->bi_iter.bi_size  < (wp->sectors_free << 9) &&
951 		    dst->bi_iter.bi_size < c->opts.encoded_extent_max)
952 			break;
953 
954 		BUG_ON(op->compression_opt &&
955 		       (op->flags & BCH_WRITE_DATA_ENCODED) &&
956 		       bch2_csum_type_is_encryption(op->crc.csum_type));
957 		BUG_ON(op->compression_opt && !bounce);
958 
959 		crc.compression_type = op->incompressible
960 			? BCH_COMPRESSION_TYPE_incompressible
961 			: op->compression_opt
962 			? bch2_bio_compress(c, dst, &dst_len, src, &src_len,
963 					    op->compression_opt)
964 			: 0;
965 		if (!crc_is_compressed(crc)) {
966 			dst_len = min(dst->bi_iter.bi_size, src->bi_iter.bi_size);
967 			dst_len = min_t(unsigned, dst_len, wp->sectors_free << 9);
968 
969 			if (op->csum_type)
970 				dst_len = min_t(unsigned, dst_len,
971 						c->opts.encoded_extent_max);
972 
973 			if (bounce) {
974 				swap(dst->bi_iter.bi_size, dst_len);
975 				bio_copy_data(dst, src);
976 				swap(dst->bi_iter.bi_size, dst_len);
977 			}
978 
979 			src_len = dst_len;
980 		}
981 
982 		BUG_ON(!src_len || !dst_len);
983 
984 		if (bch2_csum_type_is_encryption(op->csum_type)) {
985 			if (bversion_zero(version)) {
986 				version.lo = atomic64_inc_return(&c->key_version);
987 			} else {
988 				crc.nonce = op->nonce;
989 				op->nonce += src_len >> 9;
990 			}
991 		}
992 
993 		if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
994 		    !crc_is_compressed(crc) &&
995 		    bch2_csum_type_is_encryption(op->crc.csum_type) ==
996 		    bch2_csum_type_is_encryption(op->csum_type)) {
997 			u8 compression_type = crc.compression_type;
998 			u16 nonce = crc.nonce;
999 			/*
1000 			 * Note: when we're using rechecksum(), we need to be
1001 			 * checksumming @src because it has all the data our
1002 			 * existing checksum covers - if we bounced (because we
1003 			 * were trying to compress), @dst will only have the
1004 			 * part of the data the new checksum will cover.
1005 			 *
1006 			 * But normally we want to be checksumming post bounce,
1007 			 * because part of the reason for bouncing is so the
1008 			 * data can't be modified (by userspace) while it's in
1009 			 * flight.
1010 			 */
1011 			if (bch2_rechecksum_bio(c, src, version, op->crc,
1012 					&crc, &op->crc,
1013 					src_len >> 9,
1014 					bio_sectors(src) - (src_len >> 9),
1015 					op->csum_type))
1016 				goto csum_err;
1017 			/*
1018 			 * rchecksum_bio sets compression_type on crc from op->crc,
1019 			 * this isn't always correct as sometimes we're changing
1020 			 * an extent from uncompressed to incompressible.
1021 			 */
1022 			crc.compression_type = compression_type;
1023 			crc.nonce = nonce;
1024 		} else {
1025 			if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
1026 			    bch2_rechecksum_bio(c, src, version, op->crc,
1027 					NULL, &op->crc,
1028 					src_len >> 9,
1029 					bio_sectors(src) - (src_len >> 9),
1030 					op->crc.csum_type))
1031 				goto csum_err;
1032 
1033 			crc.compressed_size	= dst_len >> 9;
1034 			crc.uncompressed_size	= src_len >> 9;
1035 			crc.live_size		= src_len >> 9;
1036 
1037 			swap(dst->bi_iter.bi_size, dst_len);
1038 			ret = bch2_encrypt_bio(c, op->csum_type,
1039 					       extent_nonce(version, crc), dst);
1040 			if (ret)
1041 				goto err;
1042 
1043 			crc.csum = bch2_checksum_bio(c, op->csum_type,
1044 					 extent_nonce(version, crc), dst);
1045 			crc.csum_type = op->csum_type;
1046 			swap(dst->bi_iter.bi_size, dst_len);
1047 		}
1048 
1049 		init_append_extent(op, wp, version, crc);
1050 
1051 		if (dst != src)
1052 			bio_advance(dst, dst_len);
1053 		bio_advance(src, src_len);
1054 		total_output	+= dst_len;
1055 		total_input	+= src_len;
1056 	} while (dst->bi_iter.bi_size &&
1057 		 src->bi_iter.bi_size &&
1058 		 wp->sectors_free &&
1059 		 !bch2_keylist_realloc(&op->insert_keys,
1060 				      op->inline_keys,
1061 				      ARRAY_SIZE(op->inline_keys),
1062 				      BKEY_EXTENT_U64s_MAX));
1063 
1064 	more = src->bi_iter.bi_size != 0;
1065 
1066 	dst->bi_iter = saved_iter;
1067 
1068 	if (dst == src && more) {
1069 		BUG_ON(total_output != total_input);
1070 
1071 		dst = bio_split(src, total_input >> 9,
1072 				GFP_NOFS, &c->bio_write);
1073 		wbio_init(dst)->put_bio	= true;
1074 		/* copy WRITE_SYNC flag */
1075 		dst->bi_opf		= src->bi_opf;
1076 	}
1077 
1078 	dst->bi_iter.bi_size = total_output;
1079 do_write:
1080 	*_dst = dst;
1081 	return more;
1082 csum_err:
1083 	bch_err_inum_offset_ratelimited(c,
1084 		op->pos.inode,
1085 		op->pos.offset << 9,
1086 		"%s write error: error verifying existing checksum while rewriting existing data (memory corruption?)",
1087 		op->flags & BCH_WRITE_MOVE ? "move" : "user");
1088 	ret = -EIO;
1089 err:
1090 	if (to_wbio(dst)->bounce)
1091 		bch2_bio_free_pages_pool(c, dst);
1092 	if (to_wbio(dst)->put_bio)
1093 		bio_put(dst);
1094 
1095 	return ret;
1096 }
1097 
1098 static bool bch2_extent_is_writeable(struct bch_write_op *op,
1099 				     struct bkey_s_c k)
1100 {
1101 	struct bch_fs *c = op->c;
1102 	struct bkey_s_c_extent e;
1103 	struct extent_ptr_decoded p;
1104 	const union bch_extent_entry *entry;
1105 	unsigned replicas = 0;
1106 
1107 	if (k.k->type != KEY_TYPE_extent)
1108 		return false;
1109 
1110 	e = bkey_s_c_to_extent(k);
1111 
1112 	rcu_read_lock();
1113 	extent_for_each_ptr_decode(e, p, entry) {
1114 		if (crc_is_encoded(p.crc) || p.has_ec) {
1115 			rcu_read_unlock();
1116 			return false;
1117 		}
1118 
1119 		replicas += bch2_extent_ptr_durability(c, &p);
1120 	}
1121 	rcu_read_unlock();
1122 
1123 	return replicas >= op->opts.data_replicas;
1124 }
1125 
1126 static int bch2_nocow_write_convert_one_unwritten(struct btree_trans *trans,
1127 						  struct btree_iter *iter,
1128 						  struct bkey_i *orig,
1129 						  struct bkey_s_c k,
1130 						  u64 new_i_size)
1131 {
1132 	if (!bch2_extents_match(bkey_i_to_s_c(orig), k)) {
1133 		/* trace this */
1134 		return 0;
1135 	}
1136 
1137 	struct bkey_i *new = bch2_bkey_make_mut_noupdate(trans, k);
1138 	int ret = PTR_ERR_OR_ZERO(new);
1139 	if (ret)
1140 		return ret;
1141 
1142 	bch2_cut_front(bkey_start_pos(&orig->k), new);
1143 	bch2_cut_back(orig->k.p, new);
1144 
1145 	struct bkey_ptrs ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
1146 	bkey_for_each_ptr(ptrs, ptr)
1147 		ptr->unwritten = 0;
1148 
1149 	/*
1150 	 * Note that we're not calling bch2_subvol_get_snapshot() in this path -
1151 	 * that was done when we kicked off the write, and here it's important
1152 	 * that we update the extent that we wrote to - even if a snapshot has
1153 	 * since been created. The write is still outstanding, so we're ok
1154 	 * w.r.t. snapshot atomicity:
1155 	 */
1156 	return  bch2_extent_update_i_size_sectors(trans, iter,
1157 					min(new->k.p.offset << 9, new_i_size), 0) ?:
1158 		bch2_trans_update(trans, iter, new,
1159 				  BTREE_UPDATE_internal_snapshot_node);
1160 }
1161 
1162 static void bch2_nocow_write_convert_unwritten(struct bch_write_op *op)
1163 {
1164 	struct bch_fs *c = op->c;
1165 	struct btree_trans *trans = bch2_trans_get(c);
1166 
1167 	for_each_keylist_key(&op->insert_keys, orig) {
1168 		int ret = for_each_btree_key_upto_commit(trans, iter, BTREE_ID_extents,
1169 				     bkey_start_pos(&orig->k), orig->k.p,
1170 				     BTREE_ITER_intent, k,
1171 				     NULL, NULL, BCH_TRANS_COMMIT_no_enospc, ({
1172 			bch2_nocow_write_convert_one_unwritten(trans, &iter, orig, k, op->new_i_size);
1173 		}));
1174 
1175 		if (ret && !bch2_err_matches(ret, EROFS)) {
1176 			struct bkey_i *insert = bch2_keylist_front(&op->insert_keys);
1177 
1178 			bch_err_inum_offset_ratelimited(c,
1179 				insert->k.p.inode, insert->k.p.offset << 9,
1180 				"%s write error while doing btree update: %s",
1181 				op->flags & BCH_WRITE_MOVE ? "move" : "user",
1182 				bch2_err_str(ret));
1183 		}
1184 
1185 		if (ret) {
1186 			op->error = ret;
1187 			break;
1188 		}
1189 	}
1190 
1191 	bch2_trans_put(trans);
1192 }
1193 
1194 static void __bch2_nocow_write_done(struct bch_write_op *op)
1195 {
1196 	if (unlikely(op->flags & BCH_WRITE_IO_ERROR)) {
1197 		op->error = -EIO;
1198 	} else if (unlikely(op->flags & BCH_WRITE_CONVERT_UNWRITTEN))
1199 		bch2_nocow_write_convert_unwritten(op);
1200 }
1201 
1202 static CLOSURE_CALLBACK(bch2_nocow_write_done)
1203 {
1204 	closure_type(op, struct bch_write_op, cl);
1205 
1206 	__bch2_nocow_write_done(op);
1207 	bch2_write_done(cl);
1208 }
1209 
1210 struct bucket_to_lock {
1211 	struct bpos		b;
1212 	unsigned		gen;
1213 	struct nocow_lock_bucket *l;
1214 };
1215 
1216 static void bch2_nocow_write(struct bch_write_op *op)
1217 {
1218 	struct bch_fs *c = op->c;
1219 	struct btree_trans *trans;
1220 	struct btree_iter iter;
1221 	struct bkey_s_c k;
1222 	DARRAY_PREALLOCATED(struct bucket_to_lock, 3) buckets;
1223 	u32 snapshot;
1224 	struct bucket_to_lock *stale_at;
1225 	int stale, ret;
1226 
1227 	if (op->flags & BCH_WRITE_MOVE)
1228 		return;
1229 
1230 	darray_init(&buckets);
1231 	trans = bch2_trans_get(c);
1232 retry:
1233 	bch2_trans_begin(trans);
1234 
1235 	ret = bch2_subvolume_get_snapshot(trans, op->subvol, &snapshot);
1236 	if (unlikely(ret))
1237 		goto err;
1238 
1239 	bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
1240 			     SPOS(op->pos.inode, op->pos.offset, snapshot),
1241 			     BTREE_ITER_slots);
1242 	while (1) {
1243 		struct bio *bio = &op->wbio.bio;
1244 
1245 		buckets.nr = 0;
1246 
1247 		ret = bch2_trans_relock(trans);
1248 		if (ret)
1249 			break;
1250 
1251 		k = bch2_btree_iter_peek_slot(&iter);
1252 		ret = bkey_err(k);
1253 		if (ret)
1254 			break;
1255 
1256 		/* fall back to normal cow write path? */
1257 		if (unlikely(k.k->p.snapshot != snapshot ||
1258 			     !bch2_extent_is_writeable(op, k)))
1259 			break;
1260 
1261 		if (bch2_keylist_realloc(&op->insert_keys,
1262 					 op->inline_keys,
1263 					 ARRAY_SIZE(op->inline_keys),
1264 					 k.k->u64s))
1265 			break;
1266 
1267 		/* Get iorefs before dropping btree locks: */
1268 		struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1269 		bkey_for_each_ptr(ptrs, ptr) {
1270 			struct bch_dev *ca = bch2_dev_get_ioref(c, ptr->dev, WRITE);
1271 			if (unlikely(!ca))
1272 				goto err_get_ioref;
1273 
1274 			struct bpos b = PTR_BUCKET_POS(ca, ptr);
1275 			struct nocow_lock_bucket *l =
1276 				bucket_nocow_lock(&c->nocow_locks, bucket_to_u64(b));
1277 			prefetch(l);
1278 
1279 			/* XXX allocating memory with btree locks held - rare */
1280 			darray_push_gfp(&buckets, ((struct bucket_to_lock) {
1281 						   .b = b, .gen = ptr->gen, .l = l,
1282 						   }), GFP_KERNEL|__GFP_NOFAIL);
1283 
1284 			if (ptr->unwritten)
1285 				op->flags |= BCH_WRITE_CONVERT_UNWRITTEN;
1286 		}
1287 
1288 		/* Unlock before taking nocow locks, doing IO: */
1289 		bkey_reassemble(op->insert_keys.top, k);
1290 		bch2_trans_unlock(trans);
1291 
1292 		bch2_cut_front(op->pos, op->insert_keys.top);
1293 		if (op->flags & BCH_WRITE_CONVERT_UNWRITTEN)
1294 			bch2_cut_back(POS(op->pos.inode, op->pos.offset + bio_sectors(bio)), op->insert_keys.top);
1295 
1296 		darray_for_each(buckets, i) {
1297 			struct bch_dev *ca = bch2_dev_have_ref(c, i->b.inode);
1298 
1299 			__bch2_bucket_nocow_lock(&c->nocow_locks, i->l,
1300 						 bucket_to_u64(i->b),
1301 						 BUCKET_NOCOW_LOCK_UPDATE);
1302 
1303 			rcu_read_lock();
1304 			u8 *gen = bucket_gen(ca, i->b.offset);
1305 			stale = !gen ? -1 : gen_after(*gen, i->gen);
1306 			rcu_read_unlock();
1307 
1308 			if (unlikely(stale)) {
1309 				stale_at = i;
1310 				goto err_bucket_stale;
1311 			}
1312 		}
1313 
1314 		bio = &op->wbio.bio;
1315 		if (k.k->p.offset < op->pos.offset + bio_sectors(bio)) {
1316 			bio = bio_split(bio, k.k->p.offset - op->pos.offset,
1317 					GFP_KERNEL, &c->bio_write);
1318 			wbio_init(bio)->put_bio = true;
1319 			bio->bi_opf = op->wbio.bio.bi_opf;
1320 		} else {
1321 			op->flags |= BCH_WRITE_SUBMITTED;
1322 		}
1323 
1324 		op->pos.offset += bio_sectors(bio);
1325 		op->written += bio_sectors(bio);
1326 
1327 		bio->bi_end_io	= bch2_write_endio;
1328 		bio->bi_private	= &op->cl;
1329 		bio->bi_opf |= REQ_OP_WRITE;
1330 		closure_get(&op->cl);
1331 		bch2_submit_wbio_replicas(to_wbio(bio), c, BCH_DATA_user,
1332 					  op->insert_keys.top, true);
1333 
1334 		bch2_keylist_push(&op->insert_keys);
1335 		if (op->flags & BCH_WRITE_SUBMITTED)
1336 			break;
1337 		bch2_btree_iter_advance(&iter);
1338 	}
1339 out:
1340 	bch2_trans_iter_exit(trans, &iter);
1341 err:
1342 	if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1343 		goto retry;
1344 
1345 	if (ret) {
1346 		bch_err_inum_offset_ratelimited(c,
1347 			op->pos.inode, op->pos.offset << 9,
1348 			"%s: btree lookup error %s", __func__, bch2_err_str(ret));
1349 		op->error = ret;
1350 		op->flags |= BCH_WRITE_SUBMITTED;
1351 	}
1352 
1353 	bch2_trans_put(trans);
1354 	darray_exit(&buckets);
1355 
1356 	/* fallback to cow write path? */
1357 	if (!(op->flags & BCH_WRITE_SUBMITTED)) {
1358 		closure_sync(&op->cl);
1359 		__bch2_nocow_write_done(op);
1360 		op->insert_keys.top = op->insert_keys.keys;
1361 	} else if (op->flags & BCH_WRITE_SYNC) {
1362 		closure_sync(&op->cl);
1363 		bch2_nocow_write_done(&op->cl.work);
1364 	} else {
1365 		/*
1366 		 * XXX
1367 		 * needs to run out of process context because ei_quota_lock is
1368 		 * a mutex
1369 		 */
1370 		continue_at(&op->cl, bch2_nocow_write_done, index_update_wq(op));
1371 	}
1372 	return;
1373 err_get_ioref:
1374 	darray_for_each(buckets, i)
1375 		percpu_ref_put(&bch2_dev_have_ref(c, i->b.inode)->io_ref);
1376 
1377 	/* Fall back to COW path: */
1378 	goto out;
1379 err_bucket_stale:
1380 	darray_for_each(buckets, i) {
1381 		bch2_bucket_nocow_unlock(&c->nocow_locks, i->b, BUCKET_NOCOW_LOCK_UPDATE);
1382 		if (i == stale_at)
1383 			break;
1384 	}
1385 
1386 	struct printbuf buf = PRINTBUF;
1387 	if (bch2_fs_inconsistent_on(stale < 0, c,
1388 				    "pointer to invalid bucket in nocow path on device %llu\n  %s",
1389 				    stale_at->b.inode,
1390 				    (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1391 		ret = -EIO;
1392 	} else {
1393 		/* We can retry this: */
1394 		ret = -BCH_ERR_transaction_restart;
1395 	}
1396 	printbuf_exit(&buf);
1397 
1398 	goto err_get_ioref;
1399 }
1400 
1401 static void __bch2_write(struct bch_write_op *op)
1402 {
1403 	struct bch_fs *c = op->c;
1404 	struct write_point *wp = NULL;
1405 	struct bio *bio = NULL;
1406 	unsigned nofs_flags;
1407 	int ret;
1408 
1409 	nofs_flags = memalloc_nofs_save();
1410 
1411 	if (unlikely(op->opts.nocow && c->opts.nocow_enabled)) {
1412 		bch2_nocow_write(op);
1413 		if (op->flags & BCH_WRITE_SUBMITTED)
1414 			goto out_nofs_restore;
1415 	}
1416 again:
1417 	memset(&op->failed, 0, sizeof(op->failed));
1418 
1419 	do {
1420 		struct bkey_i *key_to_write;
1421 		unsigned key_to_write_offset = op->insert_keys.top_p -
1422 			op->insert_keys.keys_p;
1423 
1424 		/* +1 for possible cache device: */
1425 		if (op->open_buckets.nr + op->nr_replicas + 1 >
1426 		    ARRAY_SIZE(op->open_buckets.v))
1427 			break;
1428 
1429 		if (bch2_keylist_realloc(&op->insert_keys,
1430 					op->inline_keys,
1431 					ARRAY_SIZE(op->inline_keys),
1432 					BKEY_EXTENT_U64s_MAX))
1433 			break;
1434 
1435 		/*
1436 		 * The copygc thread is now global, which means it's no longer
1437 		 * freeing up space on specific disks, which means that
1438 		 * allocations for specific disks may hang arbitrarily long:
1439 		 */
1440 		ret = bch2_trans_do(c, NULL, NULL, 0,
1441 			bch2_alloc_sectors_start_trans(trans,
1442 				op->target,
1443 				op->opts.erasure_code && !(op->flags & BCH_WRITE_CACHED),
1444 				op->write_point,
1445 				&op->devs_have,
1446 				op->nr_replicas,
1447 				op->nr_replicas_required,
1448 				op->watermark,
1449 				op->flags,
1450 				&op->cl, &wp));
1451 		if (unlikely(ret)) {
1452 			if (bch2_err_matches(ret, BCH_ERR_operation_blocked))
1453 				break;
1454 
1455 			goto err;
1456 		}
1457 
1458 		EBUG_ON(!wp);
1459 
1460 		bch2_open_bucket_get(c, wp, &op->open_buckets);
1461 		ret = bch2_write_extent(op, wp, &bio);
1462 
1463 		bch2_alloc_sectors_done_inlined(c, wp);
1464 err:
1465 		if (ret <= 0) {
1466 			op->flags |= BCH_WRITE_SUBMITTED;
1467 
1468 			if (ret < 0) {
1469 				if (!(op->flags & BCH_WRITE_ALLOC_NOWAIT))
1470 					bch_err_inum_offset_ratelimited(c,
1471 						op->pos.inode,
1472 						op->pos.offset << 9,
1473 						"%s(): %s error: %s", __func__,
1474 						op->flags & BCH_WRITE_MOVE ? "move" : "user",
1475 						bch2_err_str(ret));
1476 				op->error = ret;
1477 				break;
1478 			}
1479 		}
1480 
1481 		bio->bi_end_io	= bch2_write_endio;
1482 		bio->bi_private	= &op->cl;
1483 		bio->bi_opf |= REQ_OP_WRITE;
1484 
1485 		closure_get(bio->bi_private);
1486 
1487 		key_to_write = (void *) (op->insert_keys.keys_p +
1488 					 key_to_write_offset);
1489 
1490 		bch2_submit_wbio_replicas(to_wbio(bio), c, BCH_DATA_user,
1491 					  key_to_write, false);
1492 	} while (ret);
1493 
1494 	/*
1495 	 * Sync or no?
1496 	 *
1497 	 * If we're running asynchronously, wne may still want to block
1498 	 * synchronously here if we weren't able to submit all of the IO at
1499 	 * once, as that signals backpressure to the caller.
1500 	 */
1501 	if ((op->flags & BCH_WRITE_SYNC) ||
1502 	    (!(op->flags & BCH_WRITE_SUBMITTED) &&
1503 	     !(op->flags & BCH_WRITE_IN_WORKER))) {
1504 		bch2_wait_on_allocator(c, &op->cl);
1505 
1506 		__bch2_write_index(op);
1507 
1508 		if (!(op->flags & BCH_WRITE_SUBMITTED))
1509 			goto again;
1510 		bch2_write_done(&op->cl);
1511 	} else {
1512 		bch2_write_queue(op, wp);
1513 		continue_at(&op->cl, bch2_write_index, NULL);
1514 	}
1515 out_nofs_restore:
1516 	memalloc_nofs_restore(nofs_flags);
1517 }
1518 
1519 static void bch2_write_data_inline(struct bch_write_op *op, unsigned data_len)
1520 {
1521 	struct bio *bio = &op->wbio.bio;
1522 	struct bvec_iter iter;
1523 	struct bkey_i_inline_data *id;
1524 	unsigned sectors;
1525 	int ret;
1526 
1527 	memset(&op->failed, 0, sizeof(op->failed));
1528 
1529 	op->flags |= BCH_WRITE_WROTE_DATA_INLINE;
1530 	op->flags |= BCH_WRITE_SUBMITTED;
1531 
1532 	bch2_check_set_feature(op->c, BCH_FEATURE_inline_data);
1533 
1534 	ret = bch2_keylist_realloc(&op->insert_keys, op->inline_keys,
1535 				   ARRAY_SIZE(op->inline_keys),
1536 				   BKEY_U64s + DIV_ROUND_UP(data_len, 8));
1537 	if (ret) {
1538 		op->error = ret;
1539 		goto err;
1540 	}
1541 
1542 	sectors = bio_sectors(bio);
1543 	op->pos.offset += sectors;
1544 
1545 	id = bkey_inline_data_init(op->insert_keys.top);
1546 	id->k.p		= op->pos;
1547 	id->k.version	= op->version;
1548 	id->k.size	= sectors;
1549 
1550 	iter = bio->bi_iter;
1551 	iter.bi_size = data_len;
1552 	memcpy_from_bio(id->v.data, bio, iter);
1553 
1554 	while (data_len & 7)
1555 		id->v.data[data_len++] = '\0';
1556 	set_bkey_val_bytes(&id->k, data_len);
1557 	bch2_keylist_push(&op->insert_keys);
1558 
1559 	__bch2_write_index(op);
1560 err:
1561 	bch2_write_done(&op->cl);
1562 }
1563 
1564 /**
1565  * bch2_write() - handle a write to a cache device or flash only volume
1566  * @cl:		&bch_write_op->cl
1567  *
1568  * This is the starting point for any data to end up in a cache device; it could
1569  * be from a normal write, or a writeback write, or a write to a flash only
1570  * volume - it's also used by the moving garbage collector to compact data in
1571  * mostly empty buckets.
1572  *
1573  * It first writes the data to the cache, creating a list of keys to be inserted
1574  * (if the data won't fit in a single open bucket, there will be multiple keys);
1575  * after the data is written it calls bch_journal, and after the keys have been
1576  * added to the next journal write they're inserted into the btree.
1577  *
1578  * If op->discard is true, instead of inserting the data it invalidates the
1579  * region of the cache represented by op->bio and op->inode.
1580  */
1581 CLOSURE_CALLBACK(bch2_write)
1582 {
1583 	closure_type(op, struct bch_write_op, cl);
1584 	struct bio *bio = &op->wbio.bio;
1585 	struct bch_fs *c = op->c;
1586 	unsigned data_len;
1587 
1588 	EBUG_ON(op->cl.parent);
1589 	BUG_ON(!op->nr_replicas);
1590 	BUG_ON(!op->write_point.v);
1591 	BUG_ON(bkey_eq(op->pos, POS_MAX));
1592 
1593 	if (op->flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)
1594 		op->flags |= BCH_WRITE_ALLOC_NOWAIT;
1595 
1596 	op->nr_replicas_required = min_t(unsigned, op->nr_replicas_required, op->nr_replicas);
1597 	op->start_time = local_clock();
1598 	bch2_keylist_init(&op->insert_keys, op->inline_keys);
1599 	wbio_init(bio)->put_bio = false;
1600 
1601 	if (bio->bi_iter.bi_size & (c->opts.block_size - 1)) {
1602 		bch_err_inum_offset_ratelimited(c,
1603 			op->pos.inode,
1604 			op->pos.offset << 9,
1605 			"%s write error: misaligned write",
1606 			op->flags & BCH_WRITE_MOVE ? "move" : "user");
1607 		op->error = -EIO;
1608 		goto err;
1609 	}
1610 
1611 	if (c->opts.nochanges) {
1612 		op->error = -BCH_ERR_erofs_no_writes;
1613 		goto err;
1614 	}
1615 
1616 	if (!(op->flags & BCH_WRITE_MOVE) &&
1617 	    !bch2_write_ref_tryget(c, BCH_WRITE_REF_write)) {
1618 		op->error = -BCH_ERR_erofs_no_writes;
1619 		goto err;
1620 	}
1621 
1622 	this_cpu_add(c->counters[BCH_COUNTER_io_write], bio_sectors(bio));
1623 	bch2_increment_clock(c, bio_sectors(bio), WRITE);
1624 
1625 	data_len = min_t(u64, bio->bi_iter.bi_size,
1626 			 op->new_i_size - (op->pos.offset << 9));
1627 
1628 	if (c->opts.inline_data &&
1629 	    data_len <= min(block_bytes(c) / 2, 1024U)) {
1630 		bch2_write_data_inline(op, data_len);
1631 		return;
1632 	}
1633 
1634 	__bch2_write(op);
1635 	return;
1636 err:
1637 	bch2_disk_reservation_put(c, &op->res);
1638 
1639 	closure_debug_destroy(&op->cl);
1640 	if (op->end_io)
1641 		op->end_io(op);
1642 }
1643 
1644 static const char * const bch2_write_flags[] = {
1645 #define x(f)	#f,
1646 	BCH_WRITE_FLAGS()
1647 #undef x
1648 	NULL
1649 };
1650 
1651 void bch2_write_op_to_text(struct printbuf *out, struct bch_write_op *op)
1652 {
1653 	prt_str(out, "pos: ");
1654 	bch2_bpos_to_text(out, op->pos);
1655 	prt_newline(out);
1656 	printbuf_indent_add(out, 2);
1657 
1658 	prt_str(out, "started: ");
1659 	bch2_pr_time_units(out, local_clock() - op->start_time);
1660 	prt_newline(out);
1661 
1662 	prt_str(out, "flags: ");
1663 	prt_bitflags(out, bch2_write_flags, op->flags);
1664 	prt_newline(out);
1665 
1666 	prt_printf(out, "ref: %u\n", closure_nr_remaining(&op->cl));
1667 
1668 	printbuf_indent_sub(out, 2);
1669 }
1670 
1671 void bch2_fs_io_write_exit(struct bch_fs *c)
1672 {
1673 	mempool_exit(&c->bio_bounce_pages);
1674 	bioset_exit(&c->replica_set);
1675 	bioset_exit(&c->bio_write);
1676 }
1677 
1678 int bch2_fs_io_write_init(struct bch_fs *c)
1679 {
1680 	if (bioset_init(&c->bio_write,   1, offsetof(struct bch_write_bio, bio), BIOSET_NEED_BVECS) ||
1681 	    bioset_init(&c->replica_set, 4, offsetof(struct bch_write_bio, bio), 0))
1682 		return -BCH_ERR_ENOMEM_bio_write_init;
1683 
1684 	if (mempool_init_page_pool(&c->bio_bounce_pages,
1685 				   max_t(unsigned,
1686 					 c->opts.btree_node_size,
1687 					 c->opts.encoded_extent_max) /
1688 				   PAGE_SIZE, 0))
1689 		return -BCH_ERR_ENOMEM_bio_bounce_pages_init;
1690 
1691 	return 0;
1692 }
1693