xref: /linux/fs/bcachefs/io_write.c (revision be239684b18e1cdcafcf8c7face4a2f562c745ad)
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], 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 				"write error while doing btree update: %s",
534 				bch2_err_str(ret));
535 		}
536 
537 		if (ret)
538 			goto err;
539 	}
540 out:
541 	/* If some a bucket wasn't written, we can't erasure code it: */
542 	for_each_set_bit(dev, op->failed.d, BCH_SB_MEMBERS_MAX)
543 		bch2_open_bucket_write_error(c, &op->open_buckets, dev);
544 
545 	bch2_open_buckets_put(c, &op->open_buckets);
546 	return;
547 err:
548 	keys->top = keys->keys;
549 	op->error = ret;
550 	op->flags |= BCH_WRITE_DONE;
551 	goto out;
552 }
553 
554 static inline void __wp_update_state(struct write_point *wp, enum write_point_state state)
555 {
556 	if (state != wp->state) {
557 		u64 now = ktime_get_ns();
558 
559 		if (wp->last_state_change &&
560 		    time_after64(now, wp->last_state_change))
561 			wp->time[wp->state] += now - wp->last_state_change;
562 		wp->state = state;
563 		wp->last_state_change = now;
564 	}
565 }
566 
567 static inline void wp_update_state(struct write_point *wp, bool running)
568 {
569 	enum write_point_state state;
570 
571 	state = running			 ? WRITE_POINT_running :
572 		!list_empty(&wp->writes) ? WRITE_POINT_waiting_io
573 					 : WRITE_POINT_stopped;
574 
575 	__wp_update_state(wp, state);
576 }
577 
578 static CLOSURE_CALLBACK(bch2_write_index)
579 {
580 	closure_type(op, struct bch_write_op, cl);
581 	struct write_point *wp = op->wp;
582 	struct workqueue_struct *wq = index_update_wq(op);
583 	unsigned long flags;
584 
585 	if ((op->flags & BCH_WRITE_DONE) &&
586 	    (op->flags & BCH_WRITE_MOVE))
587 		bch2_bio_free_pages_pool(op->c, &op->wbio.bio);
588 
589 	spin_lock_irqsave(&wp->writes_lock, flags);
590 	if (wp->state == WRITE_POINT_waiting_io)
591 		__wp_update_state(wp, WRITE_POINT_waiting_work);
592 	list_add_tail(&op->wp_list, &wp->writes);
593 	spin_unlock_irqrestore (&wp->writes_lock, flags);
594 
595 	queue_work(wq, &wp->index_update_work);
596 }
597 
598 static inline void bch2_write_queue(struct bch_write_op *op, struct write_point *wp)
599 {
600 	op->wp = wp;
601 
602 	if (wp->state == WRITE_POINT_stopped) {
603 		spin_lock_irq(&wp->writes_lock);
604 		__wp_update_state(wp, WRITE_POINT_waiting_io);
605 		spin_unlock_irq(&wp->writes_lock);
606 	}
607 }
608 
609 void bch2_write_point_do_index_updates(struct work_struct *work)
610 {
611 	struct write_point *wp =
612 		container_of(work, struct write_point, index_update_work);
613 	struct bch_write_op *op;
614 
615 	while (1) {
616 		spin_lock_irq(&wp->writes_lock);
617 		op = list_first_entry_or_null(&wp->writes, struct bch_write_op, wp_list);
618 		if (op)
619 			list_del(&op->wp_list);
620 		wp_update_state(wp, op != NULL);
621 		spin_unlock_irq(&wp->writes_lock);
622 
623 		if (!op)
624 			break;
625 
626 		op->flags |= BCH_WRITE_IN_WORKER;
627 
628 		__bch2_write_index(op);
629 
630 		if (!(op->flags & BCH_WRITE_DONE))
631 			__bch2_write(op);
632 		else
633 			bch2_write_done(&op->cl);
634 	}
635 }
636 
637 static void bch2_write_endio(struct bio *bio)
638 {
639 	struct closure *cl		= bio->bi_private;
640 	struct bch_write_op *op		= container_of(cl, struct bch_write_op, cl);
641 	struct bch_write_bio *wbio	= to_wbio(bio);
642 	struct bch_write_bio *parent	= wbio->split ? wbio->parent : NULL;
643 	struct bch_fs *c		= wbio->c;
644 	struct bch_dev *ca		= bch_dev_bkey_exists(c, wbio->dev);
645 
646 	if (bch2_dev_inum_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_write,
647 				    op->pos.inode,
648 				    wbio->inode_offset << 9,
649 				    "data write error: %s",
650 				    bch2_blk_status_to_str(bio->bi_status))) {
651 		set_bit(wbio->dev, op->failed.d);
652 		op->flags |= BCH_WRITE_IO_ERROR;
653 	}
654 
655 	if (wbio->nocow)
656 		set_bit(wbio->dev, op->devs_need_flush->d);
657 
658 	if (wbio->have_ioref) {
659 		bch2_latency_acct(ca, wbio->submit_time, WRITE);
660 		percpu_ref_put(&ca->io_ref);
661 	}
662 
663 	if (wbio->bounce)
664 		bch2_bio_free_pages_pool(c, bio);
665 
666 	if (wbio->put_bio)
667 		bio_put(bio);
668 
669 	if (parent)
670 		bio_endio(&parent->bio);
671 	else
672 		closure_put(cl);
673 }
674 
675 static void init_append_extent(struct bch_write_op *op,
676 			       struct write_point *wp,
677 			       struct bversion version,
678 			       struct bch_extent_crc_unpacked crc)
679 {
680 	struct bkey_i_extent *e;
681 
682 	op->pos.offset += crc.uncompressed_size;
683 
684 	e = bkey_extent_init(op->insert_keys.top);
685 	e->k.p		= op->pos;
686 	e->k.size	= crc.uncompressed_size;
687 	e->k.version	= version;
688 
689 	if (crc.csum_type ||
690 	    crc.compression_type ||
691 	    crc.nonce)
692 		bch2_extent_crc_append(&e->k_i, crc);
693 
694 	bch2_alloc_sectors_append_ptrs_inlined(op->c, wp, &e->k_i, crc.compressed_size,
695 				       op->flags & BCH_WRITE_CACHED);
696 
697 	bch2_keylist_push(&op->insert_keys);
698 }
699 
700 static struct bio *bch2_write_bio_alloc(struct bch_fs *c,
701 					struct write_point *wp,
702 					struct bio *src,
703 					bool *page_alloc_failed,
704 					void *buf)
705 {
706 	struct bch_write_bio *wbio;
707 	struct bio *bio;
708 	unsigned output_available =
709 		min(wp->sectors_free << 9, src->bi_iter.bi_size);
710 	unsigned pages = DIV_ROUND_UP(output_available +
711 				      (buf
712 				       ? ((unsigned long) buf & (PAGE_SIZE - 1))
713 				       : 0), PAGE_SIZE);
714 
715 	pages = min(pages, BIO_MAX_VECS);
716 
717 	bio = bio_alloc_bioset(NULL, pages, 0,
718 			       GFP_NOFS, &c->bio_write);
719 	wbio			= wbio_init(bio);
720 	wbio->put_bio		= true;
721 	/* copy WRITE_SYNC flag */
722 	wbio->bio.bi_opf	= src->bi_opf;
723 
724 	if (buf) {
725 		bch2_bio_map(bio, buf, output_available);
726 		return bio;
727 	}
728 
729 	wbio->bounce		= true;
730 
731 	/*
732 	 * We can't use mempool for more than c->sb.encoded_extent_max
733 	 * worth of pages, but we'd like to allocate more if we can:
734 	 */
735 	bch2_bio_alloc_pages_pool(c, bio,
736 				  min_t(unsigned, output_available,
737 					c->opts.encoded_extent_max));
738 
739 	if (bio->bi_iter.bi_size < output_available)
740 		*page_alloc_failed =
741 			bch2_bio_alloc_pages(bio,
742 					     output_available -
743 					     bio->bi_iter.bi_size,
744 					     GFP_NOFS) != 0;
745 
746 	return bio;
747 }
748 
749 static int bch2_write_rechecksum(struct bch_fs *c,
750 				 struct bch_write_op *op,
751 				 unsigned new_csum_type)
752 {
753 	struct bio *bio = &op->wbio.bio;
754 	struct bch_extent_crc_unpacked new_crc;
755 	int ret;
756 
757 	/* bch2_rechecksum_bio() can't encrypt or decrypt data: */
758 
759 	if (bch2_csum_type_is_encryption(op->crc.csum_type) !=
760 	    bch2_csum_type_is_encryption(new_csum_type))
761 		new_csum_type = op->crc.csum_type;
762 
763 	ret = bch2_rechecksum_bio(c, bio, op->version, op->crc,
764 				  NULL, &new_crc,
765 				  op->crc.offset, op->crc.live_size,
766 				  new_csum_type);
767 	if (ret)
768 		return ret;
769 
770 	bio_advance(bio, op->crc.offset << 9);
771 	bio->bi_iter.bi_size = op->crc.live_size << 9;
772 	op->crc = new_crc;
773 	return 0;
774 }
775 
776 static int bch2_write_decrypt(struct bch_write_op *op)
777 {
778 	struct bch_fs *c = op->c;
779 	struct nonce nonce = extent_nonce(op->version, op->crc);
780 	struct bch_csum csum;
781 	int ret;
782 
783 	if (!bch2_csum_type_is_encryption(op->crc.csum_type))
784 		return 0;
785 
786 	/*
787 	 * If we need to decrypt data in the write path, we'll no longer be able
788 	 * to verify the existing checksum (poly1305 mac, in this case) after
789 	 * it's decrypted - this is the last point we'll be able to reverify the
790 	 * checksum:
791 	 */
792 	csum = bch2_checksum_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
793 	if (bch2_crc_cmp(op->crc.csum, csum) && !c->opts.no_data_io)
794 		return -EIO;
795 
796 	ret = bch2_encrypt_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
797 	op->crc.csum_type = 0;
798 	op->crc.csum = (struct bch_csum) { 0, 0 };
799 	return ret;
800 }
801 
802 static enum prep_encoded_ret {
803 	PREP_ENCODED_OK,
804 	PREP_ENCODED_ERR,
805 	PREP_ENCODED_CHECKSUM_ERR,
806 	PREP_ENCODED_DO_WRITE,
807 } bch2_write_prep_encoded_data(struct bch_write_op *op, struct write_point *wp)
808 {
809 	struct bch_fs *c = op->c;
810 	struct bio *bio = &op->wbio.bio;
811 
812 	if (!(op->flags & BCH_WRITE_DATA_ENCODED))
813 		return PREP_ENCODED_OK;
814 
815 	BUG_ON(bio_sectors(bio) != op->crc.compressed_size);
816 
817 	/* Can we just write the entire extent as is? */
818 	if (op->crc.uncompressed_size == op->crc.live_size &&
819 	    op->crc.uncompressed_size <= c->opts.encoded_extent_max >> 9 &&
820 	    op->crc.compressed_size <= wp->sectors_free &&
821 	    (op->crc.compression_type == bch2_compression_opt_to_type(op->compression_opt) ||
822 	     op->incompressible)) {
823 		if (!crc_is_compressed(op->crc) &&
824 		    op->csum_type != op->crc.csum_type &&
825 		    bch2_write_rechecksum(c, op, op->csum_type) &&
826 		    !c->opts.no_data_io)
827 			return PREP_ENCODED_CHECKSUM_ERR;
828 
829 		return PREP_ENCODED_DO_WRITE;
830 	}
831 
832 	/*
833 	 * If the data is compressed and we couldn't write the entire extent as
834 	 * is, we have to decompress it:
835 	 */
836 	if (crc_is_compressed(op->crc)) {
837 		struct bch_csum csum;
838 
839 		if (bch2_write_decrypt(op))
840 			return PREP_ENCODED_CHECKSUM_ERR;
841 
842 		/* Last point we can still verify checksum: */
843 		csum = bch2_checksum_bio(c, op->crc.csum_type,
844 					 extent_nonce(op->version, op->crc),
845 					 bio);
846 		if (bch2_crc_cmp(op->crc.csum, csum) && !c->opts.no_data_io)
847 			return PREP_ENCODED_CHECKSUM_ERR;
848 
849 		if (bch2_bio_uncompress_inplace(c, bio, &op->crc))
850 			return PREP_ENCODED_ERR;
851 	}
852 
853 	/*
854 	 * No longer have compressed data after this point - data might be
855 	 * encrypted:
856 	 */
857 
858 	/*
859 	 * If the data is checksummed and we're only writing a subset,
860 	 * rechecksum and adjust bio to point to currently live data:
861 	 */
862 	if ((op->crc.live_size != op->crc.uncompressed_size ||
863 	     op->crc.csum_type != op->csum_type) &&
864 	    bch2_write_rechecksum(c, op, op->csum_type) &&
865 	    !c->opts.no_data_io)
866 		return PREP_ENCODED_CHECKSUM_ERR;
867 
868 	/*
869 	 * If we want to compress the data, it has to be decrypted:
870 	 */
871 	if ((op->compression_opt ||
872 	     bch2_csum_type_is_encryption(op->crc.csum_type) !=
873 	     bch2_csum_type_is_encryption(op->csum_type)) &&
874 	    bch2_write_decrypt(op))
875 		return PREP_ENCODED_CHECKSUM_ERR;
876 
877 	return PREP_ENCODED_OK;
878 }
879 
880 static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp,
881 			     struct bio **_dst)
882 {
883 	struct bch_fs *c = op->c;
884 	struct bio *src = &op->wbio.bio, *dst = src;
885 	struct bvec_iter saved_iter;
886 	void *ec_buf;
887 	unsigned total_output = 0, total_input = 0;
888 	bool bounce = false;
889 	bool page_alloc_failed = false;
890 	int ret, more = 0;
891 
892 	BUG_ON(!bio_sectors(src));
893 
894 	ec_buf = bch2_writepoint_ec_buf(c, wp);
895 
896 	switch (bch2_write_prep_encoded_data(op, wp)) {
897 	case PREP_ENCODED_OK:
898 		break;
899 	case PREP_ENCODED_ERR:
900 		ret = -EIO;
901 		goto err;
902 	case PREP_ENCODED_CHECKSUM_ERR:
903 		goto csum_err;
904 	case PREP_ENCODED_DO_WRITE:
905 		/* XXX look for bug here */
906 		if (ec_buf) {
907 			dst = bch2_write_bio_alloc(c, wp, src,
908 						   &page_alloc_failed,
909 						   ec_buf);
910 			bio_copy_data(dst, src);
911 			bounce = true;
912 		}
913 		init_append_extent(op, wp, op->version, op->crc);
914 		goto do_write;
915 	}
916 
917 	if (ec_buf ||
918 	    op->compression_opt ||
919 	    (op->csum_type &&
920 	     !(op->flags & BCH_WRITE_PAGES_STABLE)) ||
921 	    (bch2_csum_type_is_encryption(op->csum_type) &&
922 	     !(op->flags & BCH_WRITE_PAGES_OWNED))) {
923 		dst = bch2_write_bio_alloc(c, wp, src,
924 					   &page_alloc_failed,
925 					   ec_buf);
926 		bounce = true;
927 	}
928 
929 	saved_iter = dst->bi_iter;
930 
931 	do {
932 		struct bch_extent_crc_unpacked crc = { 0 };
933 		struct bversion version = op->version;
934 		size_t dst_len = 0, src_len = 0;
935 
936 		if (page_alloc_failed &&
937 		    dst->bi_iter.bi_size  < (wp->sectors_free << 9) &&
938 		    dst->bi_iter.bi_size < c->opts.encoded_extent_max)
939 			break;
940 
941 		BUG_ON(op->compression_opt &&
942 		       (op->flags & BCH_WRITE_DATA_ENCODED) &&
943 		       bch2_csum_type_is_encryption(op->crc.csum_type));
944 		BUG_ON(op->compression_opt && !bounce);
945 
946 		crc.compression_type = op->incompressible
947 			? BCH_COMPRESSION_TYPE_incompressible
948 			: op->compression_opt
949 			? bch2_bio_compress(c, dst, &dst_len, src, &src_len,
950 					    op->compression_opt)
951 			: 0;
952 		if (!crc_is_compressed(crc)) {
953 			dst_len = min(dst->bi_iter.bi_size, src->bi_iter.bi_size);
954 			dst_len = min_t(unsigned, dst_len, wp->sectors_free << 9);
955 
956 			if (op->csum_type)
957 				dst_len = min_t(unsigned, dst_len,
958 						c->opts.encoded_extent_max);
959 
960 			if (bounce) {
961 				swap(dst->bi_iter.bi_size, dst_len);
962 				bio_copy_data(dst, src);
963 				swap(dst->bi_iter.bi_size, dst_len);
964 			}
965 
966 			src_len = dst_len;
967 		}
968 
969 		BUG_ON(!src_len || !dst_len);
970 
971 		if (bch2_csum_type_is_encryption(op->csum_type)) {
972 			if (bversion_zero(version)) {
973 				version.lo = atomic64_inc_return(&c->key_version);
974 			} else {
975 				crc.nonce = op->nonce;
976 				op->nonce += src_len >> 9;
977 			}
978 		}
979 
980 		if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
981 		    !crc_is_compressed(crc) &&
982 		    bch2_csum_type_is_encryption(op->crc.csum_type) ==
983 		    bch2_csum_type_is_encryption(op->csum_type)) {
984 			u8 compression_type = crc.compression_type;
985 			u16 nonce = crc.nonce;
986 			/*
987 			 * Note: when we're using rechecksum(), we need to be
988 			 * checksumming @src because it has all the data our
989 			 * existing checksum covers - if we bounced (because we
990 			 * were trying to compress), @dst will only have the
991 			 * part of the data the new checksum will cover.
992 			 *
993 			 * But normally we want to be checksumming post bounce,
994 			 * because part of the reason for bouncing is so the
995 			 * data can't be modified (by userspace) while it's in
996 			 * flight.
997 			 */
998 			if (bch2_rechecksum_bio(c, src, version, op->crc,
999 					&crc, &op->crc,
1000 					src_len >> 9,
1001 					bio_sectors(src) - (src_len >> 9),
1002 					op->csum_type))
1003 				goto csum_err;
1004 			/*
1005 			 * rchecksum_bio sets compression_type on crc from op->crc,
1006 			 * this isn't always correct as sometimes we're changing
1007 			 * an extent from uncompressed to incompressible.
1008 			 */
1009 			crc.compression_type = compression_type;
1010 			crc.nonce = nonce;
1011 		} else {
1012 			if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
1013 			    bch2_rechecksum_bio(c, src, version, op->crc,
1014 					NULL, &op->crc,
1015 					src_len >> 9,
1016 					bio_sectors(src) - (src_len >> 9),
1017 					op->crc.csum_type))
1018 				goto csum_err;
1019 
1020 			crc.compressed_size	= dst_len >> 9;
1021 			crc.uncompressed_size	= src_len >> 9;
1022 			crc.live_size		= src_len >> 9;
1023 
1024 			swap(dst->bi_iter.bi_size, dst_len);
1025 			ret = bch2_encrypt_bio(c, op->csum_type,
1026 					       extent_nonce(version, crc), dst);
1027 			if (ret)
1028 				goto err;
1029 
1030 			crc.csum = bch2_checksum_bio(c, op->csum_type,
1031 					 extent_nonce(version, crc), dst);
1032 			crc.csum_type = op->csum_type;
1033 			swap(dst->bi_iter.bi_size, dst_len);
1034 		}
1035 
1036 		init_append_extent(op, wp, version, crc);
1037 
1038 		if (dst != src)
1039 			bio_advance(dst, dst_len);
1040 		bio_advance(src, src_len);
1041 		total_output	+= dst_len;
1042 		total_input	+= src_len;
1043 	} while (dst->bi_iter.bi_size &&
1044 		 src->bi_iter.bi_size &&
1045 		 wp->sectors_free &&
1046 		 !bch2_keylist_realloc(&op->insert_keys,
1047 				      op->inline_keys,
1048 				      ARRAY_SIZE(op->inline_keys),
1049 				      BKEY_EXTENT_U64s_MAX));
1050 
1051 	more = src->bi_iter.bi_size != 0;
1052 
1053 	dst->bi_iter = saved_iter;
1054 
1055 	if (dst == src && more) {
1056 		BUG_ON(total_output != total_input);
1057 
1058 		dst = bio_split(src, total_input >> 9,
1059 				GFP_NOFS, &c->bio_write);
1060 		wbio_init(dst)->put_bio	= true;
1061 		/* copy WRITE_SYNC flag */
1062 		dst->bi_opf		= src->bi_opf;
1063 	}
1064 
1065 	dst->bi_iter.bi_size = total_output;
1066 do_write:
1067 	*_dst = dst;
1068 	return more;
1069 csum_err:
1070 	bch_err(c, "error verifying existing checksum while rewriting existing data (memory corruption?)");
1071 	ret = -EIO;
1072 err:
1073 	if (to_wbio(dst)->bounce)
1074 		bch2_bio_free_pages_pool(c, dst);
1075 	if (to_wbio(dst)->put_bio)
1076 		bio_put(dst);
1077 
1078 	return ret;
1079 }
1080 
1081 static bool bch2_extent_is_writeable(struct bch_write_op *op,
1082 				     struct bkey_s_c k)
1083 {
1084 	struct bch_fs *c = op->c;
1085 	struct bkey_s_c_extent e;
1086 	struct extent_ptr_decoded p;
1087 	const union bch_extent_entry *entry;
1088 	unsigned replicas = 0;
1089 
1090 	if (k.k->type != KEY_TYPE_extent)
1091 		return false;
1092 
1093 	e = bkey_s_c_to_extent(k);
1094 	extent_for_each_ptr_decode(e, p, entry) {
1095 		if (crc_is_encoded(p.crc) || p.has_ec)
1096 			return false;
1097 
1098 		replicas += bch2_extent_ptr_durability(c, &p);
1099 	}
1100 
1101 	return replicas >= op->opts.data_replicas;
1102 }
1103 
1104 static inline void bch2_nocow_write_unlock(struct bch_write_op *op)
1105 {
1106 	struct bch_fs *c = op->c;
1107 
1108 	for_each_keylist_key(&op->insert_keys, k) {
1109 		struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(bkey_i_to_s_c(k));
1110 
1111 		bkey_for_each_ptr(ptrs, ptr)
1112 			bch2_bucket_nocow_unlock(&c->nocow_locks,
1113 						 PTR_BUCKET_POS(c, ptr),
1114 						 BUCKET_NOCOW_LOCK_UPDATE);
1115 	}
1116 }
1117 
1118 static int bch2_nocow_write_convert_one_unwritten(struct btree_trans *trans,
1119 						  struct btree_iter *iter,
1120 						  struct bkey_i *orig,
1121 						  struct bkey_s_c k,
1122 						  u64 new_i_size)
1123 {
1124 	if (!bch2_extents_match(bkey_i_to_s_c(orig), k)) {
1125 		/* trace this */
1126 		return 0;
1127 	}
1128 
1129 	struct bkey_i *new = bch2_bkey_make_mut_noupdate(trans, k);
1130 	int ret = PTR_ERR_OR_ZERO(new);
1131 	if (ret)
1132 		return ret;
1133 
1134 	bch2_cut_front(bkey_start_pos(&orig->k), new);
1135 	bch2_cut_back(orig->k.p, new);
1136 
1137 	struct bkey_ptrs ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
1138 	bkey_for_each_ptr(ptrs, ptr)
1139 		ptr->unwritten = 0;
1140 
1141 	/*
1142 	 * Note that we're not calling bch2_subvol_get_snapshot() in this path -
1143 	 * that was done when we kicked off the write, and here it's important
1144 	 * that we update the extent that we wrote to - even if a snapshot has
1145 	 * since been created. The write is still outstanding, so we're ok
1146 	 * w.r.t. snapshot atomicity:
1147 	 */
1148 	return  bch2_extent_update_i_size_sectors(trans, iter,
1149 					min(new->k.p.offset << 9, new_i_size), 0) ?:
1150 		bch2_trans_update(trans, iter, new,
1151 				  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
1152 }
1153 
1154 static void bch2_nocow_write_convert_unwritten(struct bch_write_op *op)
1155 {
1156 	struct bch_fs *c = op->c;
1157 	struct btree_trans *trans = bch2_trans_get(c);
1158 
1159 	for_each_keylist_key(&op->insert_keys, orig) {
1160 		int ret = for_each_btree_key_upto_commit(trans, iter, BTREE_ID_extents,
1161 				     bkey_start_pos(&orig->k), orig->k.p,
1162 				     BTREE_ITER_INTENT, k,
1163 				     NULL, NULL, BCH_TRANS_COMMIT_no_enospc, ({
1164 			bch2_nocow_write_convert_one_unwritten(trans, &iter, orig, k, op->new_i_size);
1165 		}));
1166 
1167 		if (ret && !bch2_err_matches(ret, EROFS)) {
1168 			struct bkey_i *insert = bch2_keylist_front(&op->insert_keys);
1169 
1170 			bch_err_inum_offset_ratelimited(c,
1171 				insert->k.p.inode, insert->k.p.offset << 9,
1172 				"write error while doing btree update: %s",
1173 				bch2_err_str(ret));
1174 		}
1175 
1176 		if (ret) {
1177 			op->error = ret;
1178 			break;
1179 		}
1180 	}
1181 
1182 	bch2_trans_put(trans);
1183 }
1184 
1185 static void __bch2_nocow_write_done(struct bch_write_op *op)
1186 {
1187 	bch2_nocow_write_unlock(op);
1188 
1189 	if (unlikely(op->flags & BCH_WRITE_IO_ERROR)) {
1190 		op->error = -EIO;
1191 	} else if (unlikely(op->flags & BCH_WRITE_CONVERT_UNWRITTEN))
1192 		bch2_nocow_write_convert_unwritten(op);
1193 }
1194 
1195 static CLOSURE_CALLBACK(bch2_nocow_write_done)
1196 {
1197 	closure_type(op, struct bch_write_op, cl);
1198 
1199 	__bch2_nocow_write_done(op);
1200 	bch2_write_done(cl);
1201 }
1202 
1203 struct bucket_to_lock {
1204 	struct bpos		b;
1205 	unsigned		gen;
1206 	struct nocow_lock_bucket *l;
1207 };
1208 
1209 static void bch2_nocow_write(struct bch_write_op *op)
1210 {
1211 	struct bch_fs *c = op->c;
1212 	struct btree_trans *trans;
1213 	struct btree_iter iter;
1214 	struct bkey_s_c k;
1215 	DARRAY_PREALLOCATED(struct bucket_to_lock, 3) buckets;
1216 	u32 snapshot;
1217 	struct bucket_to_lock *stale_at;
1218 	int ret;
1219 
1220 	if (op->flags & BCH_WRITE_MOVE)
1221 		return;
1222 
1223 	darray_init(&buckets);
1224 	trans = bch2_trans_get(c);
1225 retry:
1226 	bch2_trans_begin(trans);
1227 
1228 	ret = bch2_subvolume_get_snapshot(trans, op->subvol, &snapshot);
1229 	if (unlikely(ret))
1230 		goto err;
1231 
1232 	bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
1233 			     SPOS(op->pos.inode, op->pos.offset, snapshot),
1234 			     BTREE_ITER_SLOTS);
1235 	while (1) {
1236 		struct bio *bio = &op->wbio.bio;
1237 
1238 		buckets.nr = 0;
1239 
1240 		k = bch2_btree_iter_peek_slot(&iter);
1241 		ret = bkey_err(k);
1242 		if (ret)
1243 			break;
1244 
1245 		/* fall back to normal cow write path? */
1246 		if (unlikely(k.k->p.snapshot != snapshot ||
1247 			     !bch2_extent_is_writeable(op, k)))
1248 			break;
1249 
1250 		if (bch2_keylist_realloc(&op->insert_keys,
1251 					 op->inline_keys,
1252 					 ARRAY_SIZE(op->inline_keys),
1253 					 k.k->u64s))
1254 			break;
1255 
1256 		/* Get iorefs before dropping btree locks: */
1257 		struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1258 		bkey_for_each_ptr(ptrs, ptr) {
1259 			struct bpos b = PTR_BUCKET_POS(c, ptr);
1260 			struct nocow_lock_bucket *l =
1261 				bucket_nocow_lock(&c->nocow_locks, bucket_to_u64(b));
1262 			prefetch(l);
1263 
1264 			if (unlikely(!bch2_dev_get_ioref(bch_dev_bkey_exists(c, ptr->dev), WRITE)))
1265 				goto err_get_ioref;
1266 
1267 			/* XXX allocating memory with btree locks held - rare */
1268 			darray_push_gfp(&buckets, ((struct bucket_to_lock) {
1269 						   .b = b, .gen = ptr->gen, .l = l,
1270 						   }), GFP_KERNEL|__GFP_NOFAIL);
1271 
1272 			if (ptr->unwritten)
1273 				op->flags |= BCH_WRITE_CONVERT_UNWRITTEN;
1274 		}
1275 
1276 		/* Unlock before taking nocow locks, doing IO: */
1277 		bkey_reassemble(op->insert_keys.top, k);
1278 		bch2_trans_unlock(trans);
1279 
1280 		bch2_cut_front(op->pos, op->insert_keys.top);
1281 		if (op->flags & BCH_WRITE_CONVERT_UNWRITTEN)
1282 			bch2_cut_back(POS(op->pos.inode, op->pos.offset + bio_sectors(bio)), op->insert_keys.top);
1283 
1284 		darray_for_each(buckets, i) {
1285 			struct bch_dev *ca = bch_dev_bkey_exists(c, i->b.inode);
1286 
1287 			__bch2_bucket_nocow_lock(&c->nocow_locks, i->l,
1288 						 bucket_to_u64(i->b),
1289 						 BUCKET_NOCOW_LOCK_UPDATE);
1290 
1291 			rcu_read_lock();
1292 			bool stale = gen_after(*bucket_gen(ca, i->b.offset), i->gen);
1293 			rcu_read_unlock();
1294 
1295 			if (unlikely(stale)) {
1296 				stale_at = i;
1297 				goto err_bucket_stale;
1298 			}
1299 		}
1300 
1301 		bio = &op->wbio.bio;
1302 		if (k.k->p.offset < op->pos.offset + bio_sectors(bio)) {
1303 			bio = bio_split(bio, k.k->p.offset - op->pos.offset,
1304 					GFP_KERNEL, &c->bio_write);
1305 			wbio_init(bio)->put_bio = true;
1306 			bio->bi_opf = op->wbio.bio.bi_opf;
1307 		} else {
1308 			op->flags |= BCH_WRITE_DONE;
1309 		}
1310 
1311 		op->pos.offset += bio_sectors(bio);
1312 		op->written += bio_sectors(bio);
1313 
1314 		bio->bi_end_io	= bch2_write_endio;
1315 		bio->bi_private	= &op->cl;
1316 		bio->bi_opf |= REQ_OP_WRITE;
1317 		closure_get(&op->cl);
1318 		bch2_submit_wbio_replicas(to_wbio(bio), c, BCH_DATA_user,
1319 					  op->insert_keys.top, true);
1320 
1321 		bch2_keylist_push(&op->insert_keys);
1322 		if (op->flags & BCH_WRITE_DONE)
1323 			break;
1324 		bch2_btree_iter_advance(&iter);
1325 	}
1326 out:
1327 	bch2_trans_iter_exit(trans, &iter);
1328 err:
1329 	if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1330 		goto retry;
1331 
1332 	if (ret) {
1333 		bch_err_inum_offset_ratelimited(c,
1334 			op->pos.inode, op->pos.offset << 9,
1335 			"%s: btree lookup error %s", __func__, bch2_err_str(ret));
1336 		op->error = ret;
1337 		op->flags |= BCH_WRITE_DONE;
1338 	}
1339 
1340 	bch2_trans_put(trans);
1341 	darray_exit(&buckets);
1342 
1343 	/* fallback to cow write path? */
1344 	if (!(op->flags & BCH_WRITE_DONE)) {
1345 		closure_sync(&op->cl);
1346 		__bch2_nocow_write_done(op);
1347 		op->insert_keys.top = op->insert_keys.keys;
1348 	} else if (op->flags & BCH_WRITE_SYNC) {
1349 		closure_sync(&op->cl);
1350 		bch2_nocow_write_done(&op->cl.work);
1351 	} else {
1352 		/*
1353 		 * XXX
1354 		 * needs to run out of process context because ei_quota_lock is
1355 		 * a mutex
1356 		 */
1357 		continue_at(&op->cl, bch2_nocow_write_done, index_update_wq(op));
1358 	}
1359 	return;
1360 err_get_ioref:
1361 	darray_for_each(buckets, i)
1362 		percpu_ref_put(&bch_dev_bkey_exists(c, i->b.inode)->io_ref);
1363 
1364 	/* Fall back to COW path: */
1365 	goto out;
1366 err_bucket_stale:
1367 	darray_for_each(buckets, i) {
1368 		bch2_bucket_nocow_unlock(&c->nocow_locks, i->b, BUCKET_NOCOW_LOCK_UPDATE);
1369 		if (i == stale_at)
1370 			break;
1371 	}
1372 
1373 	/* We can retry this: */
1374 	ret = -BCH_ERR_transaction_restart;
1375 	goto err_get_ioref;
1376 }
1377 
1378 static void __bch2_write(struct bch_write_op *op)
1379 {
1380 	struct bch_fs *c = op->c;
1381 	struct write_point *wp = NULL;
1382 	struct bio *bio = NULL;
1383 	unsigned nofs_flags;
1384 	int ret;
1385 
1386 	nofs_flags = memalloc_nofs_save();
1387 
1388 	if (unlikely(op->opts.nocow && c->opts.nocow_enabled)) {
1389 		bch2_nocow_write(op);
1390 		if (op->flags & BCH_WRITE_DONE)
1391 			goto out_nofs_restore;
1392 	}
1393 again:
1394 	memset(&op->failed, 0, sizeof(op->failed));
1395 
1396 	do {
1397 		struct bkey_i *key_to_write;
1398 		unsigned key_to_write_offset = op->insert_keys.top_p -
1399 			op->insert_keys.keys_p;
1400 
1401 		/* +1 for possible cache device: */
1402 		if (op->open_buckets.nr + op->nr_replicas + 1 >
1403 		    ARRAY_SIZE(op->open_buckets.v))
1404 			break;
1405 
1406 		if (bch2_keylist_realloc(&op->insert_keys,
1407 					op->inline_keys,
1408 					ARRAY_SIZE(op->inline_keys),
1409 					BKEY_EXTENT_U64s_MAX))
1410 			break;
1411 
1412 		/*
1413 		 * The copygc thread is now global, which means it's no longer
1414 		 * freeing up space on specific disks, which means that
1415 		 * allocations for specific disks may hang arbitrarily long:
1416 		 */
1417 		ret = bch2_trans_do(c, NULL, NULL, 0,
1418 			bch2_alloc_sectors_start_trans(trans,
1419 				op->target,
1420 				op->opts.erasure_code && !(op->flags & BCH_WRITE_CACHED),
1421 				op->write_point,
1422 				&op->devs_have,
1423 				op->nr_replicas,
1424 				op->nr_replicas_required,
1425 				op->watermark,
1426 				op->flags,
1427 				(op->flags & (BCH_WRITE_ALLOC_NOWAIT|
1428 					      BCH_WRITE_ONLY_SPECIFIED_DEVS))
1429 				? NULL : &op->cl, &wp));
1430 		if (unlikely(ret)) {
1431 			if (bch2_err_matches(ret, BCH_ERR_operation_blocked))
1432 				break;
1433 
1434 			goto err;
1435 		}
1436 
1437 		EBUG_ON(!wp);
1438 
1439 		bch2_open_bucket_get(c, wp, &op->open_buckets);
1440 		ret = bch2_write_extent(op, wp, &bio);
1441 
1442 		bch2_alloc_sectors_done_inlined(c, wp);
1443 err:
1444 		if (ret <= 0) {
1445 			op->flags |= BCH_WRITE_DONE;
1446 
1447 			if (ret < 0) {
1448 				if (!(op->flags & BCH_WRITE_ALLOC_NOWAIT))
1449 					bch_err_inum_offset_ratelimited(c,
1450 						op->pos.inode,
1451 						op->pos.offset << 9,
1452 						"%s(): error: %s", __func__, bch2_err_str(ret));
1453 				op->error = ret;
1454 				break;
1455 			}
1456 		}
1457 
1458 		bio->bi_end_io	= bch2_write_endio;
1459 		bio->bi_private	= &op->cl;
1460 		bio->bi_opf |= REQ_OP_WRITE;
1461 
1462 		closure_get(bio->bi_private);
1463 
1464 		key_to_write = (void *) (op->insert_keys.keys_p +
1465 					 key_to_write_offset);
1466 
1467 		bch2_submit_wbio_replicas(to_wbio(bio), c, BCH_DATA_user,
1468 					  key_to_write, false);
1469 	} while (ret);
1470 
1471 	/*
1472 	 * Sync or no?
1473 	 *
1474 	 * If we're running asynchronously, wne may still want to block
1475 	 * synchronously here if we weren't able to submit all of the IO at
1476 	 * once, as that signals backpressure to the caller.
1477 	 */
1478 	if ((op->flags & BCH_WRITE_SYNC) ||
1479 	    (!(op->flags & BCH_WRITE_DONE) &&
1480 	     !(op->flags & BCH_WRITE_IN_WORKER))) {
1481 		closure_sync(&op->cl);
1482 		__bch2_write_index(op);
1483 
1484 		if (!(op->flags & BCH_WRITE_DONE))
1485 			goto again;
1486 		bch2_write_done(&op->cl);
1487 	} else {
1488 		bch2_write_queue(op, wp);
1489 		continue_at(&op->cl, bch2_write_index, NULL);
1490 	}
1491 out_nofs_restore:
1492 	memalloc_nofs_restore(nofs_flags);
1493 }
1494 
1495 static void bch2_write_data_inline(struct bch_write_op *op, unsigned data_len)
1496 {
1497 	struct bio *bio = &op->wbio.bio;
1498 	struct bvec_iter iter;
1499 	struct bkey_i_inline_data *id;
1500 	unsigned sectors;
1501 	int ret;
1502 
1503 	op->flags |= BCH_WRITE_WROTE_DATA_INLINE;
1504 	op->flags |= BCH_WRITE_DONE;
1505 
1506 	bch2_check_set_feature(op->c, BCH_FEATURE_inline_data);
1507 
1508 	ret = bch2_keylist_realloc(&op->insert_keys, op->inline_keys,
1509 				   ARRAY_SIZE(op->inline_keys),
1510 				   BKEY_U64s + DIV_ROUND_UP(data_len, 8));
1511 	if (ret) {
1512 		op->error = ret;
1513 		goto err;
1514 	}
1515 
1516 	sectors = bio_sectors(bio);
1517 	op->pos.offset += sectors;
1518 
1519 	id = bkey_inline_data_init(op->insert_keys.top);
1520 	id->k.p		= op->pos;
1521 	id->k.version	= op->version;
1522 	id->k.size	= sectors;
1523 
1524 	iter = bio->bi_iter;
1525 	iter.bi_size = data_len;
1526 	memcpy_from_bio(id->v.data, bio, iter);
1527 
1528 	while (data_len & 7)
1529 		id->v.data[data_len++] = '\0';
1530 	set_bkey_val_bytes(&id->k, data_len);
1531 	bch2_keylist_push(&op->insert_keys);
1532 
1533 	__bch2_write_index(op);
1534 err:
1535 	bch2_write_done(&op->cl);
1536 }
1537 
1538 /**
1539  * bch2_write() - handle a write to a cache device or flash only volume
1540  * @cl:		&bch_write_op->cl
1541  *
1542  * This is the starting point for any data to end up in a cache device; it could
1543  * be from a normal write, or a writeback write, or a write to a flash only
1544  * volume - it's also used by the moving garbage collector to compact data in
1545  * mostly empty buckets.
1546  *
1547  * It first writes the data to the cache, creating a list of keys to be inserted
1548  * (if the data won't fit in a single open bucket, there will be multiple keys);
1549  * after the data is written it calls bch_journal, and after the keys have been
1550  * added to the next journal write they're inserted into the btree.
1551  *
1552  * If op->discard is true, instead of inserting the data it invalidates the
1553  * region of the cache represented by op->bio and op->inode.
1554  */
1555 CLOSURE_CALLBACK(bch2_write)
1556 {
1557 	closure_type(op, struct bch_write_op, cl);
1558 	struct bio *bio = &op->wbio.bio;
1559 	struct bch_fs *c = op->c;
1560 	unsigned data_len;
1561 
1562 	EBUG_ON(op->cl.parent);
1563 	BUG_ON(!op->nr_replicas);
1564 	BUG_ON(!op->write_point.v);
1565 	BUG_ON(bkey_eq(op->pos, POS_MAX));
1566 
1567 	op->nr_replicas_required = min_t(unsigned, op->nr_replicas_required, op->nr_replicas);
1568 	op->start_time = local_clock();
1569 	bch2_keylist_init(&op->insert_keys, op->inline_keys);
1570 	wbio_init(bio)->put_bio = false;
1571 
1572 	if (bio->bi_iter.bi_size & (c->opts.block_size - 1)) {
1573 		bch_err_inum_offset_ratelimited(c,
1574 			op->pos.inode,
1575 			op->pos.offset << 9,
1576 			"misaligned write");
1577 		op->error = -EIO;
1578 		goto err;
1579 	}
1580 
1581 	if (c->opts.nochanges) {
1582 		op->error = -BCH_ERR_erofs_no_writes;
1583 		goto err;
1584 	}
1585 
1586 	if (!(op->flags & BCH_WRITE_MOVE) &&
1587 	    !bch2_write_ref_tryget(c, BCH_WRITE_REF_write)) {
1588 		op->error = -BCH_ERR_erofs_no_writes;
1589 		goto err;
1590 	}
1591 
1592 	this_cpu_add(c->counters[BCH_COUNTER_io_write], bio_sectors(bio));
1593 	bch2_increment_clock(c, bio_sectors(bio), WRITE);
1594 
1595 	data_len = min_t(u64, bio->bi_iter.bi_size,
1596 			 op->new_i_size - (op->pos.offset << 9));
1597 
1598 	if (c->opts.inline_data &&
1599 	    data_len <= min(block_bytes(c) / 2, 1024U)) {
1600 		bch2_write_data_inline(op, data_len);
1601 		return;
1602 	}
1603 
1604 	__bch2_write(op);
1605 	return;
1606 err:
1607 	bch2_disk_reservation_put(c, &op->res);
1608 
1609 	closure_debug_destroy(&op->cl);
1610 	if (op->end_io)
1611 		op->end_io(op);
1612 }
1613 
1614 static const char * const bch2_write_flags[] = {
1615 #define x(f)	#f,
1616 	BCH_WRITE_FLAGS()
1617 #undef x
1618 	NULL
1619 };
1620 
1621 void bch2_write_op_to_text(struct printbuf *out, struct bch_write_op *op)
1622 {
1623 	prt_str(out, "pos: ");
1624 	bch2_bpos_to_text(out, op->pos);
1625 	prt_newline(out);
1626 	printbuf_indent_add(out, 2);
1627 
1628 	prt_str(out, "started: ");
1629 	bch2_pr_time_units(out, local_clock() - op->start_time);
1630 	prt_newline(out);
1631 
1632 	prt_str(out, "flags: ");
1633 	prt_bitflags(out, bch2_write_flags, op->flags);
1634 	prt_newline(out);
1635 
1636 	prt_printf(out, "ref: %u", closure_nr_remaining(&op->cl));
1637 	prt_newline(out);
1638 
1639 	printbuf_indent_sub(out, 2);
1640 }
1641 
1642 void bch2_fs_io_write_exit(struct bch_fs *c)
1643 {
1644 	mempool_exit(&c->bio_bounce_pages);
1645 	bioset_exit(&c->bio_write);
1646 }
1647 
1648 int bch2_fs_io_write_init(struct bch_fs *c)
1649 {
1650 	if (bioset_init(&c->bio_write, 1, offsetof(struct bch_write_bio, bio),
1651 			BIOSET_NEED_BVECS))
1652 		return -BCH_ERR_ENOMEM_bio_write_init;
1653 
1654 	if (mempool_init_page_pool(&c->bio_bounce_pages,
1655 				   max_t(unsigned,
1656 					 c->opts.btree_node_size,
1657 					 c->opts.encoded_extent_max) /
1658 				   PAGE_SIZE, 0))
1659 		return -BCH_ERR_ENOMEM_bio_bounce_pages_init;
1660 
1661 	return 0;
1662 }
1663