xref: /linux/fs/bcachefs/fs-io-buffered.c (revision 0d6ccfe6b319d56da63b7d7cfbcecd92780a680d)
1 // SPDX-License-Identifier: GPL-2.0
2 #ifndef NO_BCACHEFS_FS
3 
4 #include "bcachefs.h"
5 #include "alloc_foreground.h"
6 #include "bkey_buf.h"
7 #include "fs-io.h"
8 #include "fs-io-buffered.h"
9 #include "fs-io-direct.h"
10 #include "fs-io-pagecache.h"
11 #include "io_read.h"
12 #include "io_write.h"
13 
14 #include <linux/backing-dev.h>
15 #include <linux/pagemap.h>
16 #include <linux/writeback.h>
17 
18 static inline bool bio_full(struct bio *bio, unsigned len)
19 {
20 	if (bio->bi_vcnt >= bio->bi_max_vecs)
21 		return true;
22 	if (bio->bi_iter.bi_size > UINT_MAX - len)
23 		return true;
24 	return false;
25 }
26 
27 /* readpage(s): */
28 
29 static void bch2_readpages_end_io(struct bio *bio)
30 {
31 	struct folio_iter fi;
32 
33 	bio_for_each_folio_all(fi, bio)
34 		folio_end_read(fi.folio, bio->bi_status == BLK_STS_OK);
35 
36 	bio_put(bio);
37 }
38 
39 struct readpages_iter {
40 	struct address_space	*mapping;
41 	unsigned		idx;
42 	folios			folios;
43 };
44 
45 static int readpages_iter_init(struct readpages_iter *iter,
46 			       struct readahead_control *ractl)
47 {
48 	struct folio *folio;
49 
50 	*iter = (struct readpages_iter) { ractl->mapping };
51 
52 	while ((folio = __readahead_folio(ractl))) {
53 		if (!bch2_folio_create(folio, GFP_KERNEL) ||
54 		    darray_push(&iter->folios, folio)) {
55 			bch2_folio_release(folio);
56 			ractl->_nr_pages += folio_nr_pages(folio);
57 			ractl->_index -= folio_nr_pages(folio);
58 			return iter->folios.nr ? 0 : -ENOMEM;
59 		}
60 
61 		folio_put(folio);
62 	}
63 
64 	return 0;
65 }
66 
67 static inline struct folio *readpage_iter_peek(struct readpages_iter *iter)
68 {
69 	if (iter->idx >= iter->folios.nr)
70 		return NULL;
71 	return iter->folios.data[iter->idx];
72 }
73 
74 static inline void readpage_iter_advance(struct readpages_iter *iter)
75 {
76 	iter->idx++;
77 }
78 
79 static bool extent_partial_reads_expensive(struct bkey_s_c k)
80 {
81 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
82 	struct bch_extent_crc_unpacked crc;
83 	const union bch_extent_entry *i;
84 
85 	bkey_for_each_crc(k.k, ptrs, crc, i)
86 		if (crc.csum_type || crc.compression_type)
87 			return true;
88 	return false;
89 }
90 
91 static int readpage_bio_extend(struct btree_trans *trans,
92 			       struct readpages_iter *iter,
93 			       struct bio *bio,
94 			       unsigned sectors_this_extent,
95 			       bool get_more)
96 {
97 	/* Don't hold btree locks while allocating memory: */
98 	bch2_trans_unlock(trans);
99 
100 	while (bio_sectors(bio) < sectors_this_extent &&
101 	       bio->bi_vcnt < bio->bi_max_vecs) {
102 		struct folio *folio = readpage_iter_peek(iter);
103 		int ret;
104 
105 		if (folio) {
106 			readpage_iter_advance(iter);
107 		} else {
108 			pgoff_t folio_offset = bio_end_sector(bio) >> PAGE_SECTORS_SHIFT;
109 
110 			if (!get_more)
111 				break;
112 
113 			folio = xa_load(&iter->mapping->i_pages, folio_offset);
114 			if (folio && !xa_is_value(folio))
115 				break;
116 
117 			folio = filemap_alloc_folio(readahead_gfp_mask(iter->mapping), 0);
118 			if (!folio)
119 				break;
120 
121 			if (!__bch2_folio_create(folio, GFP_KERNEL)) {
122 				folio_put(folio);
123 				break;
124 			}
125 
126 			ret = filemap_add_folio(iter->mapping, folio, folio_offset, GFP_KERNEL);
127 			if (ret) {
128 				__bch2_folio_release(folio);
129 				folio_put(folio);
130 				break;
131 			}
132 
133 			folio_put(folio);
134 		}
135 
136 		BUG_ON(folio_sector(folio) != bio_end_sector(bio));
137 
138 		BUG_ON(!bio_add_folio(bio, folio, folio_size(folio), 0));
139 	}
140 
141 	return bch2_trans_relock(trans);
142 }
143 
144 static void bchfs_read(struct btree_trans *trans,
145 		       struct bch_read_bio *rbio,
146 		       subvol_inum inum,
147 		       struct readpages_iter *readpages_iter)
148 {
149 	struct bch_fs *c = trans->c;
150 	struct btree_iter iter;
151 	struct bkey_buf sk;
152 	int flags = BCH_READ_RETRY_IF_STALE|
153 		BCH_READ_MAY_PROMOTE;
154 	u32 snapshot;
155 	int ret = 0;
156 
157 	rbio->c = c;
158 	rbio->start_time = local_clock();
159 	rbio->subvol = inum.subvol;
160 
161 	bch2_bkey_buf_init(&sk);
162 retry:
163 	bch2_trans_begin(trans);
164 	iter = (struct btree_iter) { NULL };
165 
166 	ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
167 	if (ret)
168 		goto err;
169 
170 	bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
171 			     SPOS(inum.inum, rbio->bio.bi_iter.bi_sector, snapshot),
172 			     BTREE_ITER_slots);
173 	while (1) {
174 		struct bkey_s_c k;
175 		unsigned bytes, sectors, offset_into_extent;
176 		enum btree_id data_btree = BTREE_ID_extents;
177 
178 		/*
179 		 * read_extent -> io_time_reset may cause a transaction restart
180 		 * without returning an error, we need to check for that here:
181 		 */
182 		ret = bch2_trans_relock(trans);
183 		if (ret)
184 			break;
185 
186 		bch2_btree_iter_set_pos(&iter,
187 				POS(inum.inum, rbio->bio.bi_iter.bi_sector));
188 
189 		k = bch2_btree_iter_peek_slot(&iter);
190 		ret = bkey_err(k);
191 		if (ret)
192 			break;
193 
194 		offset_into_extent = iter.pos.offset -
195 			bkey_start_offset(k.k);
196 		sectors = k.k->size - offset_into_extent;
197 
198 		bch2_bkey_buf_reassemble(&sk, c, k);
199 
200 		ret = bch2_read_indirect_extent(trans, &data_btree,
201 					&offset_into_extent, &sk);
202 		if (ret)
203 			break;
204 
205 		k = bkey_i_to_s_c(sk.k);
206 
207 		sectors = min(sectors, k.k->size - offset_into_extent);
208 
209 		if (readpages_iter) {
210 			ret = readpage_bio_extend(trans, readpages_iter, &rbio->bio, sectors,
211 						  extent_partial_reads_expensive(k));
212 			if (ret)
213 				break;
214 		}
215 
216 		bytes = min(sectors, bio_sectors(&rbio->bio)) << 9;
217 		swap(rbio->bio.bi_iter.bi_size, bytes);
218 
219 		if (rbio->bio.bi_iter.bi_size == bytes)
220 			flags |= BCH_READ_LAST_FRAGMENT;
221 
222 		bch2_bio_page_state_set(&rbio->bio, k);
223 
224 		bch2_read_extent(trans, rbio, iter.pos,
225 				 data_btree, k, offset_into_extent, flags);
226 
227 		if (flags & BCH_READ_LAST_FRAGMENT)
228 			break;
229 
230 		swap(rbio->bio.bi_iter.bi_size, bytes);
231 		bio_advance(&rbio->bio, bytes);
232 
233 		ret = btree_trans_too_many_iters(trans);
234 		if (ret)
235 			break;
236 	}
237 err:
238 	bch2_trans_iter_exit(trans, &iter);
239 
240 	if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
241 		goto retry;
242 
243 	if (ret) {
244 		bch_err_inum_offset_ratelimited(c,
245 				iter.pos.inode,
246 				iter.pos.offset << 9,
247 				"read error %i from btree lookup", ret);
248 		rbio->bio.bi_status = BLK_STS_IOERR;
249 		bio_endio(&rbio->bio);
250 	}
251 
252 	bch2_bkey_buf_exit(&sk, c);
253 }
254 
255 void bch2_readahead(struct readahead_control *ractl)
256 {
257 	struct bch_inode_info *inode = to_bch_ei(ractl->mapping->host);
258 	struct bch_fs *c = inode->v.i_sb->s_fs_info;
259 	struct bch_io_opts opts;
260 	struct folio *folio;
261 	struct readpages_iter readpages_iter;
262 
263 	bch2_inode_opts_get(&opts, c, &inode->ei_inode);
264 
265 	int ret = readpages_iter_init(&readpages_iter, ractl);
266 	if (ret)
267 		return;
268 
269 	bch2_pagecache_add_get(inode);
270 
271 	struct btree_trans *trans = bch2_trans_get(c);
272 	while ((folio = readpage_iter_peek(&readpages_iter))) {
273 		unsigned n = min_t(unsigned,
274 				   readpages_iter.folios.nr -
275 				   readpages_iter.idx,
276 				   BIO_MAX_VECS);
277 		struct bch_read_bio *rbio =
278 			rbio_init(bio_alloc_bioset(NULL, n, REQ_OP_READ,
279 						   GFP_KERNEL, &c->bio_read),
280 				  opts);
281 
282 		readpage_iter_advance(&readpages_iter);
283 
284 		rbio->bio.bi_iter.bi_sector = folio_sector(folio);
285 		rbio->bio.bi_end_io = bch2_readpages_end_io;
286 		BUG_ON(!bio_add_folio(&rbio->bio, folio, folio_size(folio), 0));
287 
288 		bchfs_read(trans, rbio, inode_inum(inode),
289 			   &readpages_iter);
290 		bch2_trans_unlock(trans);
291 	}
292 	bch2_trans_put(trans);
293 
294 	bch2_pagecache_add_put(inode);
295 
296 	darray_exit(&readpages_iter.folios);
297 }
298 
299 static void bch2_read_single_folio_end_io(struct bio *bio)
300 {
301 	complete(bio->bi_private);
302 }
303 
304 int bch2_read_single_folio(struct folio *folio, struct address_space *mapping)
305 {
306 	struct bch_inode_info *inode = to_bch_ei(mapping->host);
307 	struct bch_fs *c = inode->v.i_sb->s_fs_info;
308 	struct bch_read_bio *rbio;
309 	struct bch_io_opts opts;
310 	int ret;
311 	DECLARE_COMPLETION_ONSTACK(done);
312 
313 	if (!bch2_folio_create(folio, GFP_KERNEL))
314 		return -ENOMEM;
315 
316 	bch2_inode_opts_get(&opts, c, &inode->ei_inode);
317 
318 	rbio = rbio_init(bio_alloc_bioset(NULL, 1, REQ_OP_READ, GFP_KERNEL, &c->bio_read),
319 			 opts);
320 	rbio->bio.bi_private = &done;
321 	rbio->bio.bi_end_io = bch2_read_single_folio_end_io;
322 
323 	rbio->bio.bi_opf = REQ_OP_READ|REQ_SYNC;
324 	rbio->bio.bi_iter.bi_sector = folio_sector(folio);
325 	BUG_ON(!bio_add_folio(&rbio->bio, folio, folio_size(folio), 0));
326 
327 	bch2_trans_run(c, (bchfs_read(trans, rbio, inode_inum(inode), NULL), 0));
328 	wait_for_completion(&done);
329 
330 	ret = blk_status_to_errno(rbio->bio.bi_status);
331 	bio_put(&rbio->bio);
332 
333 	if (ret < 0)
334 		return ret;
335 
336 	folio_mark_uptodate(folio);
337 	return 0;
338 }
339 
340 int bch2_read_folio(struct file *file, struct folio *folio)
341 {
342 	int ret;
343 
344 	ret = bch2_read_single_folio(folio, folio->mapping);
345 	folio_unlock(folio);
346 	return bch2_err_class(ret);
347 }
348 
349 /* writepages: */
350 
351 struct bch_writepage_io {
352 	struct bch_inode_info		*inode;
353 
354 	/* must be last: */
355 	struct bch_write_op		op;
356 };
357 
358 struct bch_writepage_state {
359 	struct bch_writepage_io	*io;
360 	struct bch_io_opts	opts;
361 	struct bch_folio_sector	*tmp;
362 	unsigned		tmp_sectors;
363 };
364 
365 static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
366 								  struct bch_inode_info *inode)
367 {
368 	struct bch_writepage_state ret = { 0 };
369 
370 	bch2_inode_opts_get(&ret.opts, c, &inode->ei_inode);
371 	return ret;
372 }
373 
374 /*
375  * Determine when a writepage io is full. We have to limit writepage bios to a
376  * single page per bvec (i.e. 1MB with 4k pages) because that is the limit to
377  * what the bounce path in bch2_write_extent() can handle. In theory we could
378  * loosen this restriction for non-bounce I/O, but we don't have that context
379  * here. Ideally, we can up this limit and make it configurable in the future
380  * when the bounce path can be enhanced to accommodate larger source bios.
381  */
382 static inline bool bch_io_full(struct bch_writepage_io *io, unsigned len)
383 {
384 	struct bio *bio = &io->op.wbio.bio;
385 	return bio_full(bio, len) ||
386 		(bio->bi_iter.bi_size + len > BIO_MAX_VECS * PAGE_SIZE);
387 }
388 
389 static void bch2_writepage_io_done(struct bch_write_op *op)
390 {
391 	struct bch_writepage_io *io =
392 		container_of(op, struct bch_writepage_io, op);
393 	struct bch_fs *c = io->op.c;
394 	struct bio *bio = &io->op.wbio.bio;
395 	struct folio_iter fi;
396 	unsigned i;
397 
398 	if (io->op.error) {
399 		set_bit(EI_INODE_ERROR, &io->inode->ei_flags);
400 
401 		bio_for_each_folio_all(fi, bio) {
402 			struct bch_folio *s;
403 
404 			mapping_set_error(fi.folio->mapping, -EIO);
405 
406 			s = __bch2_folio(fi.folio);
407 			spin_lock(&s->lock);
408 			for (i = 0; i < folio_sectors(fi.folio); i++)
409 				s->s[i].nr_replicas = 0;
410 			spin_unlock(&s->lock);
411 		}
412 	}
413 
414 	if (io->op.flags & BCH_WRITE_WROTE_DATA_INLINE) {
415 		bio_for_each_folio_all(fi, bio) {
416 			struct bch_folio *s;
417 
418 			s = __bch2_folio(fi.folio);
419 			spin_lock(&s->lock);
420 			for (i = 0; i < folio_sectors(fi.folio); i++)
421 				s->s[i].nr_replicas = 0;
422 			spin_unlock(&s->lock);
423 		}
424 	}
425 
426 	/*
427 	 * racing with fallocate can cause us to add fewer sectors than
428 	 * expected - but we shouldn't add more sectors than expected:
429 	 */
430 	WARN_ON_ONCE(io->op.i_sectors_delta > 0);
431 
432 	/*
433 	 * (error (due to going RO) halfway through a page can screw that up
434 	 * slightly)
435 	 * XXX wtf?
436 	   BUG_ON(io->op.op.i_sectors_delta >= PAGE_SECTORS);
437 	 */
438 
439 	/*
440 	 * The writeback flag is effectively our ref on the inode -
441 	 * fixup i_blocks before calling folio_end_writeback:
442 	 */
443 	bch2_i_sectors_acct(c, io->inode, NULL, io->op.i_sectors_delta);
444 
445 	bio_for_each_folio_all(fi, bio) {
446 		struct bch_folio *s = __bch2_folio(fi.folio);
447 
448 		if (atomic_dec_and_test(&s->write_count))
449 			folio_end_writeback(fi.folio);
450 	}
451 
452 	bio_put(&io->op.wbio.bio);
453 }
454 
455 static void bch2_writepage_do_io(struct bch_writepage_state *w)
456 {
457 	struct bch_writepage_io *io = w->io;
458 
459 	w->io = NULL;
460 	closure_call(&io->op.cl, bch2_write, NULL, NULL);
461 }
462 
463 /*
464  * Get a bch_writepage_io and add @page to it - appending to an existing one if
465  * possible, else allocating a new one:
466  */
467 static void bch2_writepage_io_alloc(struct bch_fs *c,
468 				    struct writeback_control *wbc,
469 				    struct bch_writepage_state *w,
470 				    struct bch_inode_info *inode,
471 				    u64 sector,
472 				    unsigned nr_replicas)
473 {
474 	struct bch_write_op *op;
475 
476 	w->io = container_of(bio_alloc_bioset(NULL, BIO_MAX_VECS,
477 					      REQ_OP_WRITE,
478 					      GFP_KERNEL,
479 					      &c->writepage_bioset),
480 			     struct bch_writepage_io, op.wbio.bio);
481 
482 	w->io->inode		= inode;
483 	op			= &w->io->op;
484 	bch2_write_op_init(op, c, w->opts);
485 	op->target		= w->opts.foreground_target;
486 	op->nr_replicas		= nr_replicas;
487 	op->res.nr_replicas	= nr_replicas;
488 	op->write_point		= writepoint_hashed(inode->ei_last_dirtied);
489 	op->subvol		= inode->ei_subvol;
490 	op->pos			= POS(inode->v.i_ino, sector);
491 	op->end_io		= bch2_writepage_io_done;
492 	op->devs_need_flush	= &inode->ei_devs_need_flush;
493 	op->wbio.bio.bi_iter.bi_sector = sector;
494 	op->wbio.bio.bi_opf	= wbc_to_write_flags(wbc);
495 }
496 
497 static int __bch2_writepage(struct folio *folio,
498 			    struct writeback_control *wbc,
499 			    void *data)
500 {
501 	struct bch_inode_info *inode = to_bch_ei(folio->mapping->host);
502 	struct bch_fs *c = inode->v.i_sb->s_fs_info;
503 	struct bch_writepage_state *w = data;
504 	struct bch_folio *s;
505 	unsigned i, offset, f_sectors, nr_replicas_this_write = U32_MAX;
506 	loff_t i_size = i_size_read(&inode->v);
507 	int ret;
508 
509 	EBUG_ON(!folio_test_uptodate(folio));
510 
511 	/* Is the folio fully inside i_size? */
512 	if (folio_end_pos(folio) <= i_size)
513 		goto do_io;
514 
515 	/* Is the folio fully outside i_size? (truncate in progress) */
516 	if (folio_pos(folio) >= i_size) {
517 		folio_unlock(folio);
518 		return 0;
519 	}
520 
521 	/*
522 	 * The folio straddles i_size.  It must be zeroed out on each and every
523 	 * writepage invocation because it may be mmapped.  "A file is mapped
524 	 * in multiples of the folio size.  For a file that is not a multiple of
525 	 * the  folio size, the remaining memory is zeroed when mapped, and
526 	 * writes to that region are not written out to the file."
527 	 */
528 	folio_zero_segment(folio,
529 			   i_size - folio_pos(folio),
530 			   folio_size(folio));
531 do_io:
532 	f_sectors = folio_sectors(folio);
533 	s = bch2_folio(folio);
534 
535 	if (f_sectors > w->tmp_sectors) {
536 		kfree(w->tmp);
537 		w->tmp = kcalloc(f_sectors, sizeof(struct bch_folio_sector), __GFP_NOFAIL);
538 		w->tmp_sectors = f_sectors;
539 	}
540 
541 	/*
542 	 * Things get really hairy with errors during writeback:
543 	 */
544 	ret = bch2_get_folio_disk_reservation(c, inode, folio, false);
545 	BUG_ON(ret);
546 
547 	/* Before unlocking the page, get copy of reservations: */
548 	spin_lock(&s->lock);
549 	memcpy(w->tmp, s->s, sizeof(struct bch_folio_sector) * f_sectors);
550 
551 	for (i = 0; i < f_sectors; i++) {
552 		if (s->s[i].state < SECTOR_dirty)
553 			continue;
554 
555 		nr_replicas_this_write =
556 			min_t(unsigned, nr_replicas_this_write,
557 			      s->s[i].nr_replicas +
558 			      s->s[i].replicas_reserved);
559 	}
560 
561 	for (i = 0; i < f_sectors; i++) {
562 		if (s->s[i].state < SECTOR_dirty)
563 			continue;
564 
565 		s->s[i].nr_replicas = w->opts.compression
566 			? 0 : nr_replicas_this_write;
567 
568 		s->s[i].replicas_reserved = 0;
569 		bch2_folio_sector_set(folio, s, i, SECTOR_allocated);
570 	}
571 	spin_unlock(&s->lock);
572 
573 	BUG_ON(atomic_read(&s->write_count));
574 	atomic_set(&s->write_count, 1);
575 
576 	BUG_ON(folio_test_writeback(folio));
577 	folio_start_writeback(folio);
578 
579 	folio_unlock(folio);
580 
581 	offset = 0;
582 	while (1) {
583 		unsigned sectors = 0, dirty_sectors = 0, reserved_sectors = 0;
584 		u64 sector;
585 
586 		while (offset < f_sectors &&
587 		       w->tmp[offset].state < SECTOR_dirty)
588 			offset++;
589 
590 		if (offset == f_sectors)
591 			break;
592 
593 		while (offset + sectors < f_sectors &&
594 		       w->tmp[offset + sectors].state >= SECTOR_dirty) {
595 			reserved_sectors += w->tmp[offset + sectors].replicas_reserved;
596 			dirty_sectors += w->tmp[offset + sectors].state == SECTOR_dirty;
597 			sectors++;
598 		}
599 		BUG_ON(!sectors);
600 
601 		sector = folio_sector(folio) + offset;
602 
603 		if (w->io &&
604 		    (w->io->op.res.nr_replicas != nr_replicas_this_write ||
605 		     bch_io_full(w->io, sectors << 9) ||
606 		     bio_end_sector(&w->io->op.wbio.bio) != sector))
607 			bch2_writepage_do_io(w);
608 
609 		if (!w->io)
610 			bch2_writepage_io_alloc(c, wbc, w, inode, sector,
611 						nr_replicas_this_write);
612 
613 		atomic_inc(&s->write_count);
614 
615 		BUG_ON(inode != w->io->inode);
616 		BUG_ON(!bio_add_folio(&w->io->op.wbio.bio, folio,
617 				     sectors << 9, offset << 9));
618 
619 		/* Check for writing past i_size: */
620 		WARN_ONCE((bio_end_sector(&w->io->op.wbio.bio) << 9) >
621 			  round_up(i_size, block_bytes(c)) &&
622 			  !test_bit(BCH_FS_emergency_ro, &c->flags),
623 			  "writing past i_size: %llu > %llu (unrounded %llu)\n",
624 			  bio_end_sector(&w->io->op.wbio.bio) << 9,
625 			  round_up(i_size, block_bytes(c)),
626 			  i_size);
627 
628 		w->io->op.res.sectors += reserved_sectors;
629 		w->io->op.i_sectors_delta -= dirty_sectors;
630 		w->io->op.new_i_size = i_size;
631 
632 		offset += sectors;
633 	}
634 
635 	if (atomic_dec_and_test(&s->write_count))
636 		folio_end_writeback(folio);
637 
638 	return 0;
639 }
640 
641 int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
642 {
643 	struct bch_fs *c = mapping->host->i_sb->s_fs_info;
644 	struct bch_writepage_state w =
645 		bch_writepage_state_init(c, to_bch_ei(mapping->host));
646 	struct blk_plug plug;
647 	int ret;
648 
649 	blk_start_plug(&plug);
650 	ret = write_cache_pages(mapping, wbc, __bch2_writepage, &w);
651 	if (w.io)
652 		bch2_writepage_do_io(&w);
653 	blk_finish_plug(&plug);
654 	kfree(w.tmp);
655 	return bch2_err_class(ret);
656 }
657 
658 /* buffered writes: */
659 
660 int bch2_write_begin(struct file *file, struct address_space *mapping,
661 		     loff_t pos, unsigned len,
662 		     struct page **pagep, void **fsdata)
663 {
664 	struct bch_inode_info *inode = to_bch_ei(mapping->host);
665 	struct bch_fs *c = inode->v.i_sb->s_fs_info;
666 	struct bch2_folio_reservation *res;
667 	struct folio *folio;
668 	unsigned offset;
669 	int ret = -ENOMEM;
670 
671 	res = kmalloc(sizeof(*res), GFP_KERNEL);
672 	if (!res)
673 		return -ENOMEM;
674 
675 	bch2_folio_reservation_init(c, inode, res);
676 	*fsdata = res;
677 
678 	bch2_pagecache_add_get(inode);
679 
680 	folio = __filemap_get_folio(mapping, pos >> PAGE_SHIFT,
681 				    FGP_WRITEBEGIN | fgf_set_order(len),
682 				    mapping_gfp_mask(mapping));
683 	if (IS_ERR_OR_NULL(folio))
684 		goto err_unlock;
685 
686 	offset = pos - folio_pos(folio);
687 	len = min_t(size_t, len, folio_end_pos(folio) - pos);
688 
689 	if (folio_test_uptodate(folio))
690 		goto out;
691 
692 	/* If we're writing entire folio, don't need to read it in first: */
693 	if (!offset && len == folio_size(folio))
694 		goto out;
695 
696 	if (!offset && pos + len >= inode->v.i_size) {
697 		folio_zero_segment(folio, len, folio_size(folio));
698 		flush_dcache_folio(folio);
699 		goto out;
700 	}
701 
702 	if (folio_pos(folio) >= inode->v.i_size) {
703 		folio_zero_segments(folio, 0, offset, offset + len, folio_size(folio));
704 		flush_dcache_folio(folio);
705 		goto out;
706 	}
707 readpage:
708 	ret = bch2_read_single_folio(folio, mapping);
709 	if (ret)
710 		goto err;
711 out:
712 	ret = bch2_folio_set(c, inode_inum(inode), &folio, 1);
713 	if (ret)
714 		goto err;
715 
716 	ret = bch2_folio_reservation_get(c, inode, folio, res, offset, len);
717 	if (ret) {
718 		if (!folio_test_uptodate(folio)) {
719 			/*
720 			 * If the folio hasn't been read in, we won't know if we
721 			 * actually need a reservation - we don't actually need
722 			 * to read here, we just need to check if the folio is
723 			 * fully backed by uncompressed data:
724 			 */
725 			goto readpage;
726 		}
727 
728 		goto err;
729 	}
730 
731 	*pagep = &folio->page;
732 	return 0;
733 err:
734 	folio_unlock(folio);
735 	folio_put(folio);
736 	*pagep = NULL;
737 err_unlock:
738 	bch2_pagecache_add_put(inode);
739 	kfree(res);
740 	*fsdata = NULL;
741 	return bch2_err_class(ret);
742 }
743 
744 int bch2_write_end(struct file *file, struct address_space *mapping,
745 		   loff_t pos, unsigned len, unsigned copied,
746 		   struct page *page, void *fsdata)
747 {
748 	struct bch_inode_info *inode = to_bch_ei(mapping->host);
749 	struct bch_fs *c = inode->v.i_sb->s_fs_info;
750 	struct bch2_folio_reservation *res = fsdata;
751 	struct folio *folio = page_folio(page);
752 	unsigned offset = pos - folio_pos(folio);
753 
754 	lockdep_assert_held(&inode->v.i_rwsem);
755 	BUG_ON(offset + copied > folio_size(folio));
756 
757 	if (unlikely(copied < len && !folio_test_uptodate(folio))) {
758 		/*
759 		 * The folio needs to be read in, but that would destroy
760 		 * our partial write - simplest thing is to just force
761 		 * userspace to redo the write:
762 		 */
763 		folio_zero_range(folio, 0, folio_size(folio));
764 		flush_dcache_folio(folio);
765 		copied = 0;
766 	}
767 
768 	spin_lock(&inode->v.i_lock);
769 	if (pos + copied > inode->v.i_size)
770 		i_size_write(&inode->v, pos + copied);
771 	spin_unlock(&inode->v.i_lock);
772 
773 	if (copied) {
774 		if (!folio_test_uptodate(folio))
775 			folio_mark_uptodate(folio);
776 
777 		bch2_set_folio_dirty(c, inode, folio, res, offset, copied);
778 
779 		inode->ei_last_dirtied = (unsigned long) current;
780 	}
781 
782 	folio_unlock(folio);
783 	folio_put(folio);
784 	bch2_pagecache_add_put(inode);
785 
786 	bch2_folio_reservation_put(c, inode, res);
787 	kfree(res);
788 
789 	return copied;
790 }
791 
792 static noinline void folios_trunc(folios *fs, struct folio **fi)
793 {
794 	while (fs->data + fs->nr > fi) {
795 		struct folio *f = darray_pop(fs);
796 
797 		folio_unlock(f);
798 		folio_put(f);
799 	}
800 }
801 
802 static int __bch2_buffered_write(struct bch_inode_info *inode,
803 				 struct address_space *mapping,
804 				 struct iov_iter *iter,
805 				 loff_t pos, unsigned len,
806 				 bool inode_locked)
807 {
808 	struct bch_fs *c = inode->v.i_sb->s_fs_info;
809 	struct bch2_folio_reservation res;
810 	folios fs;
811 	struct folio *f;
812 	unsigned copied = 0, f_offset, f_copied;
813 	u64 end = pos + len, f_pos, f_len;
814 	loff_t last_folio_pos = inode->v.i_size;
815 	int ret = 0;
816 
817 	BUG_ON(!len);
818 
819 	bch2_folio_reservation_init(c, inode, &res);
820 	darray_init(&fs);
821 
822 	ret = bch2_filemap_get_contig_folios_d(mapping, pos, end,
823 					       FGP_WRITEBEGIN | fgf_set_order(len),
824 					       mapping_gfp_mask(mapping), &fs);
825 	if (ret)
826 		goto out;
827 
828 	BUG_ON(!fs.nr);
829 
830 	/*
831 	 * If we're not using the inode lock, we need to lock all the folios for
832 	 * atomiticity of writes vs. other writes:
833 	 */
834 	if (!inode_locked && folio_end_pos(darray_last(fs)) < end) {
835 		ret = -BCH_ERR_need_inode_lock;
836 		goto out;
837 	}
838 
839 	f = darray_first(fs);
840 	if (pos != folio_pos(f) && !folio_test_uptodate(f)) {
841 		ret = bch2_read_single_folio(f, mapping);
842 		if (ret)
843 			goto out;
844 	}
845 
846 	f = darray_last(fs);
847 	end = min(end, folio_end_pos(f));
848 	last_folio_pos = folio_pos(f);
849 	if (end != folio_end_pos(f) && !folio_test_uptodate(f)) {
850 		if (end >= inode->v.i_size) {
851 			folio_zero_range(f, 0, folio_size(f));
852 		} else {
853 			ret = bch2_read_single_folio(f, mapping);
854 			if (ret)
855 				goto out;
856 		}
857 	}
858 
859 	ret = bch2_folio_set(c, inode_inum(inode), fs.data, fs.nr);
860 	if (ret)
861 		goto out;
862 
863 	f_pos = pos;
864 	f_offset = pos - folio_pos(darray_first(fs));
865 	darray_for_each(fs, fi) {
866 		ssize_t f_reserved;
867 
868 		f = *fi;
869 		f_len = min(end, folio_end_pos(f)) - f_pos;
870 		f_reserved = bch2_folio_reservation_get_partial(c, inode, f, &res, f_offset, f_len);
871 
872 		if (unlikely(f_reserved != f_len)) {
873 			if (f_reserved < 0) {
874 				if (f == darray_first(fs)) {
875 					ret = f_reserved;
876 					goto out;
877 				}
878 
879 				folios_trunc(&fs, fi);
880 				end = min(end, folio_end_pos(darray_last(fs)));
881 			} else {
882 				folios_trunc(&fs, fi + 1);
883 				end = f_pos + f_reserved;
884 			}
885 
886 			break;
887 		}
888 
889 		f_pos = folio_end_pos(f);
890 		f_offset = 0;
891 	}
892 
893 	if (mapping_writably_mapped(mapping))
894 		darray_for_each(fs, fi)
895 			flush_dcache_folio(*fi);
896 
897 	f_pos = pos;
898 	f_offset = pos - folio_pos(darray_first(fs));
899 	darray_for_each(fs, fi) {
900 		f = *fi;
901 		f_len = min(end, folio_end_pos(f)) - f_pos;
902 		f_copied = copy_folio_from_iter_atomic(f, f_offset, f_len, iter);
903 		if (!f_copied) {
904 			folios_trunc(&fs, fi);
905 			break;
906 		}
907 
908 		if (!folio_test_uptodate(f) &&
909 		    f_copied != folio_size(f) &&
910 		    pos + copied + f_copied < inode->v.i_size) {
911 			iov_iter_revert(iter, f_copied);
912 			folio_zero_range(f, 0, folio_size(f));
913 			folios_trunc(&fs, fi);
914 			break;
915 		}
916 
917 		flush_dcache_folio(f);
918 		copied += f_copied;
919 
920 		if (f_copied != f_len) {
921 			folios_trunc(&fs, fi + 1);
922 			break;
923 		}
924 
925 		f_pos = folio_end_pos(f);
926 		f_offset = 0;
927 	}
928 
929 	if (!copied)
930 		goto out;
931 
932 	end = pos + copied;
933 
934 	spin_lock(&inode->v.i_lock);
935 	if (end > inode->v.i_size) {
936 		BUG_ON(!inode_locked);
937 		i_size_write(&inode->v, end);
938 	}
939 	spin_unlock(&inode->v.i_lock);
940 
941 	f_pos = pos;
942 	f_offset = pos - folio_pos(darray_first(fs));
943 	darray_for_each(fs, fi) {
944 		f = *fi;
945 		f_len = min(end, folio_end_pos(f)) - f_pos;
946 
947 		if (!folio_test_uptodate(f))
948 			folio_mark_uptodate(f);
949 
950 		bch2_set_folio_dirty(c, inode, f, &res, f_offset, f_len);
951 
952 		f_pos = folio_end_pos(f);
953 		f_offset = 0;
954 	}
955 
956 	inode->ei_last_dirtied = (unsigned long) current;
957 out:
958 	darray_for_each(fs, fi) {
959 		folio_unlock(*fi);
960 		folio_put(*fi);
961 	}
962 
963 	/*
964 	 * If the last folio added to the mapping starts beyond current EOF, we
965 	 * performed a short write but left around at least one post-EOF folio.
966 	 * Clean up the mapping before we return.
967 	 */
968 	if (last_folio_pos >= inode->v.i_size)
969 		truncate_pagecache(&inode->v, inode->v.i_size);
970 
971 	darray_exit(&fs);
972 	bch2_folio_reservation_put(c, inode, &res);
973 
974 	return copied ?: ret;
975 }
976 
977 static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter)
978 {
979 	struct file *file = iocb->ki_filp;
980 	struct address_space *mapping = file->f_mapping;
981 	struct bch_inode_info *inode = file_bch_inode(file);
982 	loff_t pos;
983 	bool inode_locked = false;
984 	ssize_t written = 0, written2 = 0, ret = 0;
985 
986 	/*
987 	 * We don't take the inode lock unless i_size will be changing. Folio
988 	 * locks provide exclusion with other writes, and the pagecache add lock
989 	 * provides exclusion with truncate and hole punching.
990 	 *
991 	 * There is one nasty corner case where atomicity would be broken
992 	 * without great care: when copying data from userspace to the page
993 	 * cache, we do that with faults disable - a page fault would recurse
994 	 * back into the filesystem, taking filesystem locks again, and
995 	 * deadlock; so it's done with faults disabled, and we fault in the user
996 	 * buffer when we aren't holding locks.
997 	 *
998 	 * If we do part of the write, but we then race and in the userspace
999 	 * buffer have been evicted and are no longer resident, then we have to
1000 	 * drop our folio locks to re-fault them in, breaking write atomicity.
1001 	 *
1002 	 * To fix this, we restart the write from the start, if we weren't
1003 	 * holding the inode lock.
1004 	 *
1005 	 * There is another wrinkle after that; if we restart the write from the
1006 	 * start, and then get an unrecoverable error, we _cannot_ claim to
1007 	 * userspace that we did not write data we actually did - so we must
1008 	 * track (written2) the most we ever wrote.
1009 	 */
1010 
1011 	if ((iocb->ki_flags & IOCB_APPEND) ||
1012 	    (iocb->ki_pos + iov_iter_count(iter) > i_size_read(&inode->v))) {
1013 		inode_lock(&inode->v);
1014 		inode_locked = true;
1015 	}
1016 
1017 	ret = generic_write_checks(iocb, iter);
1018 	if (ret <= 0)
1019 		goto unlock;
1020 
1021 	ret = file_remove_privs_flags(file, !inode_locked ? IOCB_NOWAIT : 0);
1022 	if (ret) {
1023 		if (!inode_locked) {
1024 			inode_lock(&inode->v);
1025 			inode_locked = true;
1026 			ret = file_remove_privs_flags(file, 0);
1027 		}
1028 		if (ret)
1029 			goto unlock;
1030 	}
1031 
1032 	ret = file_update_time(file);
1033 	if (ret)
1034 		goto unlock;
1035 
1036 	pos = iocb->ki_pos;
1037 
1038 	bch2_pagecache_add_get(inode);
1039 
1040 	if (!inode_locked &&
1041 	    (iocb->ki_pos + iov_iter_count(iter) > i_size_read(&inode->v)))
1042 		goto get_inode_lock;
1043 
1044 	do {
1045 		unsigned offset = pos & (PAGE_SIZE - 1);
1046 		unsigned bytes = iov_iter_count(iter);
1047 again:
1048 		/*
1049 		 * Bring in the user page that we will copy from _first_.
1050 		 * Otherwise there's a nasty deadlock on copying from the
1051 		 * same page as we're writing to, without it being marked
1052 		 * up-to-date.
1053 		 *
1054 		 * Not only is this an optimisation, but it is also required
1055 		 * to check that the address is actually valid, when atomic
1056 		 * usercopies are used, below.
1057 		 */
1058 		if (unlikely(fault_in_iov_iter_readable(iter, bytes))) {
1059 			bytes = min_t(unsigned long, iov_iter_count(iter),
1060 				      PAGE_SIZE - offset);
1061 
1062 			if (unlikely(fault_in_iov_iter_readable(iter, bytes))) {
1063 				ret = -EFAULT;
1064 				break;
1065 			}
1066 		}
1067 
1068 		if (unlikely(bytes != iov_iter_count(iter) && !inode_locked))
1069 			goto get_inode_lock;
1070 
1071 		if (unlikely(fatal_signal_pending(current))) {
1072 			ret = -EINTR;
1073 			break;
1074 		}
1075 
1076 		ret = __bch2_buffered_write(inode, mapping, iter, pos, bytes, inode_locked);
1077 		if (ret == -BCH_ERR_need_inode_lock)
1078 			goto get_inode_lock;
1079 		if (unlikely(ret < 0))
1080 			break;
1081 
1082 		cond_resched();
1083 
1084 		if (unlikely(ret == 0)) {
1085 			/*
1086 			 * If we were unable to copy any data at all, we must
1087 			 * fall back to a single segment length write.
1088 			 *
1089 			 * If we didn't fallback here, we could livelock
1090 			 * because not all segments in the iov can be copied at
1091 			 * once without a pagefault.
1092 			 */
1093 			bytes = min_t(unsigned long, PAGE_SIZE - offset,
1094 				      iov_iter_single_seg_count(iter));
1095 			goto again;
1096 		}
1097 		pos += ret;
1098 		written += ret;
1099 		written2 = max(written, written2);
1100 
1101 		if (ret != bytes && !inode_locked)
1102 			goto get_inode_lock;
1103 		ret = 0;
1104 
1105 		balance_dirty_pages_ratelimited(mapping);
1106 
1107 		if (0) {
1108 get_inode_lock:
1109 			bch2_pagecache_add_put(inode);
1110 			inode_lock(&inode->v);
1111 			inode_locked = true;
1112 			bch2_pagecache_add_get(inode);
1113 
1114 			iov_iter_revert(iter, written);
1115 			pos -= written;
1116 			written = 0;
1117 			ret = 0;
1118 		}
1119 	} while (iov_iter_count(iter));
1120 	bch2_pagecache_add_put(inode);
1121 unlock:
1122 	if (inode_locked)
1123 		inode_unlock(&inode->v);
1124 
1125 	iocb->ki_pos += written;
1126 
1127 	ret = max(written, written2) ?: ret;
1128 	if (ret > 0)
1129 		ret = generic_write_sync(iocb, ret);
1130 	return ret;
1131 }
1132 
1133 ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *iter)
1134 {
1135 	ssize_t ret = iocb->ki_flags & IOCB_DIRECT
1136 		? bch2_direct_write(iocb, iter)
1137 		: bch2_buffered_write(iocb, iter);
1138 
1139 	return bch2_err_class(ret);
1140 }
1141 
1142 void bch2_fs_fs_io_buffered_exit(struct bch_fs *c)
1143 {
1144 	bioset_exit(&c->writepage_bioset);
1145 }
1146 
1147 int bch2_fs_fs_io_buffered_init(struct bch_fs *c)
1148 {
1149 	if (bioset_init(&c->writepage_bioset,
1150 			4, offsetof(struct bch_writepage_io, op.wbio.bio),
1151 			BIOSET_NEED_BVECS))
1152 		return -BCH_ERR_ENOMEM_writepage_bioset_init;
1153 
1154 	return 0;
1155 }
1156 
1157 #endif /* NO_BCACHEFS_FS */
1158