Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2010 Red Hat, Inc.
4 * Copyright (c) 2016-2018 Christoph Hellwig.
5 */
6#include <linux/module.h>
7#include <linux/compiler.h>
8#include <linux/fs.h>
9#include <linux/iomap.h>
10#include <linux/pagemap.h>
11#include <linux/uio.h>
12#include <linux/buffer_head.h>
13#include <linux/dax.h>
14#include <linux/writeback.h>
15#include <linux/swap.h>
16#include <linux/bio.h>
17#include <linux/sched/signal.h>
18#include <linux/migrate.h>
19
20#include "../internal.h"
21
22static struct iomap_page *
23iomap_page_create(struct inode *inode, struct page *page)
24{
25 struct iomap_page *iop = to_iomap_page(page);
26
27 if (iop || i_blocksize(inode) == PAGE_SIZE)
28 return iop;
29
30 iop = kmalloc(sizeof(*iop), GFP_NOFS | __GFP_NOFAIL);
31 atomic_set(&iop->read_count, 0);
32 atomic_set(&iop->write_count, 0);
33 bitmap_zero(iop->uptodate, PAGE_SIZE / SECTOR_SIZE);
34
35 /*
36 * migrate_page_move_mapping() assumes that pages with private data have
37 * their count elevated by 1.
38 */
39 get_page(page);
40 set_page_private(page, (unsigned long)iop);
41 SetPagePrivate(page);
42 return iop;
43}
44
45static void
46iomap_page_release(struct page *page)
47{
48 struct iomap_page *iop = to_iomap_page(page);
49
50 if (!iop)
51 return;
52 WARN_ON_ONCE(atomic_read(&iop->read_count));
53 WARN_ON_ONCE(atomic_read(&iop->write_count));
54 ClearPagePrivate(page);
55 set_page_private(page, 0);
56 put_page(page);
57 kfree(iop);
58}
59
60/*
61 * Calculate the range inside the page that we actually need to read.
62 */
63static void
64iomap_adjust_read_range(struct inode *inode, struct iomap_page *iop,
65 loff_t *pos, loff_t length, unsigned *offp, unsigned *lenp)
66{
67 loff_t orig_pos = *pos;
68 loff_t isize = i_size_read(inode);
69 unsigned block_bits = inode->i_blkbits;
70 unsigned block_size = (1 << block_bits);
71 unsigned poff = offset_in_page(*pos);
72 unsigned plen = min_t(loff_t, PAGE_SIZE - poff, length);
73 unsigned first = poff >> block_bits;
74 unsigned last = (poff + plen - 1) >> block_bits;
75
76 /*
77 * If the block size is smaller than the page size we need to check the
78 * per-block uptodate status and adjust the offset and length if needed
79 * to avoid reading in already uptodate ranges.
80 */
81 if (iop) {
82 unsigned int i;
83
84 /* move forward for each leading block marked uptodate */
85 for (i = first; i <= last; i++) {
86 if (!test_bit(i, iop->uptodate))
87 break;
88 *pos += block_size;
89 poff += block_size;
90 plen -= block_size;
91 first++;
92 }
93
94 /* truncate len if we find any trailing uptodate block(s) */
95 for ( ; i <= last; i++) {
96 if (test_bit(i, iop->uptodate)) {
97 plen -= (last - i + 1) * block_size;
98 last = i - 1;
99 break;
100 }
101 }
102 }
103
104 /*
105 * If the extent spans the block that contains the i_size we need to
106 * handle both halves separately so that we properly zero data in the
107 * page cache for blocks that are entirely outside of i_size.
108 */
109 if (orig_pos <= isize && orig_pos + length > isize) {
110 unsigned end = offset_in_page(isize - 1) >> block_bits;
111
112 if (first <= end && last > end)
113 plen -= (last - end) * block_size;
114 }
115
116 *offp = poff;
117 *lenp = plen;
118}
119
120static void
121iomap_set_range_uptodate(struct page *page, unsigned off, unsigned len)
122{
123 struct iomap_page *iop = to_iomap_page(page);
124 struct inode *inode = page->mapping->host;
125 unsigned first = off >> inode->i_blkbits;
126 unsigned last = (off + len - 1) >> inode->i_blkbits;
127 unsigned int i;
128 bool uptodate = true;
129
130 if (iop) {
131 for (i = 0; i < PAGE_SIZE / i_blocksize(inode); i++) {
132 if (i >= first && i <= last)
133 set_bit(i, iop->uptodate);
134 else if (!test_bit(i, iop->uptodate))
135 uptodate = false;
136 }
137 }
138
139 if (uptodate && !PageError(page))
140 SetPageUptodate(page);
141}
142
143static void
144iomap_read_finish(struct iomap_page *iop, struct page *page)
145{
146 if (!iop || atomic_dec_and_test(&iop->read_count))
147 unlock_page(page);
148}
149
150static void
151iomap_read_page_end_io(struct bio_vec *bvec, int error)
152{
153 struct page *page = bvec->bv_page;
154 struct iomap_page *iop = to_iomap_page(page);
155
156 if (unlikely(error)) {
157 ClearPageUptodate(page);
158 SetPageError(page);
159 } else {
160 iomap_set_range_uptodate(page, bvec->bv_offset, bvec->bv_len);
161 }
162
163 iomap_read_finish(iop, page);
164}
165
166static void
167iomap_read_end_io(struct bio *bio)
168{
169 int error = blk_status_to_errno(bio->bi_status);
170 struct bio_vec *bvec;
171 struct bvec_iter_all iter_all;
172
173 bio_for_each_segment_all(bvec, bio, iter_all)
174 iomap_read_page_end_io(bvec, error);
175 bio_put(bio);
176}
177
178struct iomap_readpage_ctx {
179 struct page *cur_page;
180 bool cur_page_in_bio;
181 bool is_readahead;
182 struct bio *bio;
183 struct list_head *pages;
184};
185
186static void
187iomap_read_inline_data(struct inode *inode, struct page *page,
188 struct iomap *iomap)
189{
190 size_t size = i_size_read(inode);
191 void *addr;
192
193 if (PageUptodate(page))
194 return;
195
196 BUG_ON(page->index);
197 BUG_ON(size > PAGE_SIZE - offset_in_page(iomap->inline_data));
198
199 addr = kmap_atomic(page);
200 memcpy(addr, iomap->inline_data, size);
201 memset(addr + size, 0, PAGE_SIZE - size);
202 kunmap_atomic(addr);
203 SetPageUptodate(page);
204}
205
206static loff_t
207iomap_readpage_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
208 struct iomap *iomap)
209{
210 struct iomap_readpage_ctx *ctx = data;
211 struct page *page = ctx->cur_page;
212 struct iomap_page *iop = iomap_page_create(inode, page);
213 bool same_page = false, is_contig = false;
214 loff_t orig_pos = pos;
215 unsigned poff, plen;
216 sector_t sector;
217
218 if (iomap->type == IOMAP_INLINE) {
219 WARN_ON_ONCE(pos);
220 iomap_read_inline_data(inode, page, iomap);
221 return PAGE_SIZE;
222 }
223
224 /* zero post-eof blocks as the page may be mapped */
225 iomap_adjust_read_range(inode, iop, &pos, length, &poff, &plen);
226 if (plen == 0)
227 goto done;
228
229 if (iomap->type != IOMAP_MAPPED || pos >= i_size_read(inode)) {
230 zero_user(page, poff, plen);
231 iomap_set_range_uptodate(page, poff, plen);
232 goto done;
233 }
234
235 ctx->cur_page_in_bio = true;
236
237 /*
238 * Try to merge into a previous segment if we can.
239 */
240 sector = iomap_sector(iomap, pos);
241 if (ctx->bio && bio_end_sector(ctx->bio) == sector)
242 is_contig = true;
243
244 if (is_contig &&
245 __bio_try_merge_page(ctx->bio, page, plen, poff, &same_page)) {
246 if (!same_page && iop)
247 atomic_inc(&iop->read_count);
248 goto done;
249 }
250
251 /*
252 * If we start a new segment we need to increase the read count, and we
253 * need to do so before submitting any previous full bio to make sure
254 * that we don't prematurely unlock the page.
255 */
256 if (iop)
257 atomic_inc(&iop->read_count);
258
259 if (!ctx->bio || !is_contig || bio_full(ctx->bio, plen)) {
260 gfp_t gfp = mapping_gfp_constraint(page->mapping, GFP_KERNEL);
261 int nr_vecs = (length + PAGE_SIZE - 1) >> PAGE_SHIFT;
262
263 if (ctx->bio)
264 submit_bio(ctx->bio);
265
266 if (ctx->is_readahead) /* same as readahead_gfp_mask */
267 gfp |= __GFP_NORETRY | __GFP_NOWARN;
268 ctx->bio = bio_alloc(gfp, min(BIO_MAX_PAGES, nr_vecs));
269 ctx->bio->bi_opf = REQ_OP_READ;
270 if (ctx->is_readahead)
271 ctx->bio->bi_opf |= REQ_RAHEAD;
272 ctx->bio->bi_iter.bi_sector = sector;
273 bio_set_dev(ctx->bio, iomap->bdev);
274 ctx->bio->bi_end_io = iomap_read_end_io;
275 }
276
277 bio_add_page(ctx->bio, page, plen, poff);
278done:
279 /*
280 * Move the caller beyond our range so that it keeps making progress.
281 * For that we have to include any leading non-uptodate ranges, but
282 * we can skip trailing ones as they will be handled in the next
283 * iteration.
284 */
285 return pos - orig_pos + plen;
286}
287
288int
289iomap_readpage(struct page *page, const struct iomap_ops *ops)
290{
291 struct iomap_readpage_ctx ctx = { .cur_page = page };
292 struct inode *inode = page->mapping->host;
293 unsigned poff;
294 loff_t ret;
295
296 for (poff = 0; poff < PAGE_SIZE; poff += ret) {
297 ret = iomap_apply(inode, page_offset(page) + poff,
298 PAGE_SIZE - poff, 0, ops, &ctx,
299 iomap_readpage_actor);
300 if (ret <= 0) {
301 WARN_ON_ONCE(ret == 0);
302 SetPageError(page);
303 break;
304 }
305 }
306
307 if (ctx.bio) {
308 submit_bio(ctx.bio);
309 WARN_ON_ONCE(!ctx.cur_page_in_bio);
310 } else {
311 WARN_ON_ONCE(ctx.cur_page_in_bio);
312 unlock_page(page);
313 }
314
315 /*
316 * Just like mpage_readpages and block_read_full_page we always
317 * return 0 and just mark the page as PageError on errors. This
318 * should be cleaned up all through the stack eventually.
319 */
320 return 0;
321}
322EXPORT_SYMBOL_GPL(iomap_readpage);
323
324static struct page *
325iomap_next_page(struct inode *inode, struct list_head *pages, loff_t pos,
326 loff_t length, loff_t *done)
327{
328 while (!list_empty(pages)) {
329 struct page *page = lru_to_page(pages);
330
331 if (page_offset(page) >= (u64)pos + length)
332 break;
333
334 list_del(&page->lru);
335 if (!add_to_page_cache_lru(page, inode->i_mapping, page->index,
336 GFP_NOFS))
337 return page;
338
339 /*
340 * If we already have a page in the page cache at index we are
341 * done. Upper layers don't care if it is uptodate after the
342 * readpages call itself as every page gets checked again once
343 * actually needed.
344 */
345 *done += PAGE_SIZE;
346 put_page(page);
347 }
348
349 return NULL;
350}
351
352static loff_t
353iomap_readpages_actor(struct inode *inode, loff_t pos, loff_t length,
354 void *data, struct iomap *iomap)
355{
356 struct iomap_readpage_ctx *ctx = data;
357 loff_t done, ret;
358
359 for (done = 0; done < length; done += ret) {
360 if (ctx->cur_page && offset_in_page(pos + done) == 0) {
361 if (!ctx->cur_page_in_bio)
362 unlock_page(ctx->cur_page);
363 put_page(ctx->cur_page);
364 ctx->cur_page = NULL;
365 }
366 if (!ctx->cur_page) {
367 ctx->cur_page = iomap_next_page(inode, ctx->pages,
368 pos, length, &done);
369 if (!ctx->cur_page)
370 break;
371 ctx->cur_page_in_bio = false;
372 }
373 ret = iomap_readpage_actor(inode, pos + done, length - done,
374 ctx, iomap);
375 }
376
377 return done;
378}
379
380int
381iomap_readpages(struct address_space *mapping, struct list_head *pages,
382 unsigned nr_pages, const struct iomap_ops *ops)
383{
384 struct iomap_readpage_ctx ctx = {
385 .pages = pages,
386 .is_readahead = true,
387 };
388 loff_t pos = page_offset(list_entry(pages->prev, struct page, lru));
389 loff_t last = page_offset(list_entry(pages->next, struct page, lru));
390 loff_t length = last - pos + PAGE_SIZE, ret = 0;
391
392 while (length > 0) {
393 ret = iomap_apply(mapping->host, pos, length, 0, ops,
394 &ctx, iomap_readpages_actor);
395 if (ret <= 0) {
396 WARN_ON_ONCE(ret == 0);
397 goto done;
398 }
399 pos += ret;
400 length -= ret;
401 }
402 ret = 0;
403done:
404 if (ctx.bio)
405 submit_bio(ctx.bio);
406 if (ctx.cur_page) {
407 if (!ctx.cur_page_in_bio)
408 unlock_page(ctx.cur_page);
409 put_page(ctx.cur_page);
410 }
411
412 /*
413 * Check that we didn't lose a page due to the arcance calling
414 * conventions..
415 */
416 WARN_ON_ONCE(!ret && !list_empty(ctx.pages));
417 return ret;
418}
419EXPORT_SYMBOL_GPL(iomap_readpages);
420
421/*
422 * iomap_is_partially_uptodate checks whether blocks within a page are
423 * uptodate or not.
424 *
425 * Returns true if all blocks which correspond to a file portion
426 * we want to read within the page are uptodate.
427 */
428int
429iomap_is_partially_uptodate(struct page *page, unsigned long from,
430 unsigned long count)
431{
432 struct iomap_page *iop = to_iomap_page(page);
433 struct inode *inode = page->mapping->host;
434 unsigned len, first, last;
435 unsigned i;
436
437 /* Limit range to one page */
438 len = min_t(unsigned, PAGE_SIZE - from, count);
439
440 /* First and last blocks in range within page */
441 first = from >> inode->i_blkbits;
442 last = (from + len - 1) >> inode->i_blkbits;
443
444 if (iop) {
445 for (i = first; i <= last; i++)
446 if (!test_bit(i, iop->uptodate))
447 return 0;
448 return 1;
449 }
450
451 return 0;
452}
453EXPORT_SYMBOL_GPL(iomap_is_partially_uptodate);
454
455int
456iomap_releasepage(struct page *page, gfp_t gfp_mask)
457{
458 /*
459 * mm accommodates an old ext3 case where clean pages might not have had
460 * the dirty bit cleared. Thus, it can send actual dirty pages to
461 * ->releasepage() via shrink_active_list(), skip those here.
462 */
463 if (PageDirty(page) || PageWriteback(page))
464 return 0;
465 iomap_page_release(page);
466 return 1;
467}
468EXPORT_SYMBOL_GPL(iomap_releasepage);
469
470void
471iomap_invalidatepage(struct page *page, unsigned int offset, unsigned int len)
472{
473 /*
474 * If we are invalidating the entire page, clear the dirty state from it
475 * and release it to avoid unnecessary buildup of the LRU.
476 */
477 if (offset == 0 && len == PAGE_SIZE) {
478 WARN_ON_ONCE(PageWriteback(page));
479 cancel_dirty_page(page);
480 iomap_page_release(page);
481 }
482}
483EXPORT_SYMBOL_GPL(iomap_invalidatepage);
484
485#ifdef CONFIG_MIGRATION
486int
487iomap_migrate_page(struct address_space *mapping, struct page *newpage,
488 struct page *page, enum migrate_mode mode)
489{
490 int ret;
491
492 ret = migrate_page_move_mapping(mapping, newpage, page, 0);
493 if (ret != MIGRATEPAGE_SUCCESS)
494 return ret;
495
496 if (page_has_private(page)) {
497 ClearPagePrivate(page);
498 get_page(newpage);
499 set_page_private(newpage, page_private(page));
500 set_page_private(page, 0);
501 put_page(page);
502 SetPagePrivate(newpage);
503 }
504
505 if (mode != MIGRATE_SYNC_NO_COPY)
506 migrate_page_copy(newpage, page);
507 else
508 migrate_page_states(newpage, page);
509 return MIGRATEPAGE_SUCCESS;
510}
511EXPORT_SYMBOL_GPL(iomap_migrate_page);
512#endif /* CONFIG_MIGRATION */
513
514static void
515iomap_write_failed(struct inode *inode, loff_t pos, unsigned len)
516{
517 loff_t i_size = i_size_read(inode);
518
519 /*
520 * Only truncate newly allocated pages beyoned EOF, even if the
521 * write started inside the existing inode size.
522 */
523 if (pos + len > i_size)
524 truncate_pagecache_range(inode, max(pos, i_size), pos + len);
525}
526
527static int
528iomap_read_page_sync(struct inode *inode, loff_t block_start, struct page *page,
529 unsigned poff, unsigned plen, unsigned from, unsigned to,
530 struct iomap *iomap)
531{
532 struct bio_vec bvec;
533 struct bio bio;
534
535 if (iomap->type != IOMAP_MAPPED || block_start >= i_size_read(inode)) {
536 zero_user_segments(page, poff, from, to, poff + plen);
537 iomap_set_range_uptodate(page, poff, plen);
538 return 0;
539 }
540
541 bio_init(&bio, &bvec, 1);
542 bio.bi_opf = REQ_OP_READ;
543 bio.bi_iter.bi_sector = iomap_sector(iomap, block_start);
544 bio_set_dev(&bio, iomap->bdev);
545 __bio_add_page(&bio, page, plen, poff);
546 return submit_bio_wait(&bio);
547}
548
549static int
550__iomap_write_begin(struct inode *inode, loff_t pos, unsigned len,
551 struct page *page, struct iomap *iomap)
552{
553 struct iomap_page *iop = iomap_page_create(inode, page);
554 loff_t block_size = i_blocksize(inode);
555 loff_t block_start = pos & ~(block_size - 1);
556 loff_t block_end = (pos + len + block_size - 1) & ~(block_size - 1);
557 unsigned from = offset_in_page(pos), to = from + len, poff, plen;
558 int status = 0;
559
560 if (PageUptodate(page))
561 return 0;
562
563 do {
564 iomap_adjust_read_range(inode, iop, &block_start,
565 block_end - block_start, &poff, &plen);
566 if (plen == 0)
567 break;
568
569 if ((from > poff && from < poff + plen) ||
570 (to > poff && to < poff + plen)) {
571 status = iomap_read_page_sync(inode, block_start, page,
572 poff, plen, from, to, iomap);
573 if (status)
574 break;
575 }
576
577 } while ((block_start += plen) < block_end);
578
579 return status;
580}
581
582static int
583iomap_write_begin(struct inode *inode, loff_t pos, unsigned len, unsigned flags,
584 struct page **pagep, struct iomap *iomap)
585{
586 const struct iomap_page_ops *page_ops = iomap->page_ops;
587 pgoff_t index = pos >> PAGE_SHIFT;
588 struct page *page;
589 int status = 0;
590
591 BUG_ON(pos + len > iomap->offset + iomap->length);
592
593 if (fatal_signal_pending(current))
594 return -EINTR;
595
596 if (page_ops && page_ops->page_prepare) {
597 status = page_ops->page_prepare(inode, pos, len, iomap);
598 if (status)
599 return status;
600 }
601
602 page = grab_cache_page_write_begin(inode->i_mapping, index, flags);
603 if (!page) {
604 status = -ENOMEM;
605 goto out_no_page;
606 }
607
608 if (iomap->type == IOMAP_INLINE)
609 iomap_read_inline_data(inode, page, iomap);
610 else if (iomap->flags & IOMAP_F_BUFFER_HEAD)
611 status = __block_write_begin_int(page, pos, len, NULL, iomap);
612 else
613 status = __iomap_write_begin(inode, pos, len, page, iomap);
614
615 if (unlikely(status))
616 goto out_unlock;
617
618 *pagep = page;
619 return 0;
620
621out_unlock:
622 unlock_page(page);
623 put_page(page);
624 iomap_write_failed(inode, pos, len);
625
626out_no_page:
627 if (page_ops && page_ops->page_done)
628 page_ops->page_done(inode, pos, 0, NULL, iomap);
629 return status;
630}
631
632int
633iomap_set_page_dirty(struct page *page)
634{
635 struct address_space *mapping = page_mapping(page);
636 int newly_dirty;
637
638 if (unlikely(!mapping))
639 return !TestSetPageDirty(page);
640
641 /*
642 * Lock out page->mem_cgroup migration to keep PageDirty
643 * synchronized with per-memcg dirty page counters.
644 */
645 lock_page_memcg(page);
646 newly_dirty = !TestSetPageDirty(page);
647 if (newly_dirty)
648 __set_page_dirty(page, mapping, 0);
649 unlock_page_memcg(page);
650
651 if (newly_dirty)
652 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
653 return newly_dirty;
654}
655EXPORT_SYMBOL_GPL(iomap_set_page_dirty);
656
657static int
658__iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
659 unsigned copied, struct page *page, struct iomap *iomap)
660{
661 flush_dcache_page(page);
662
663 /*
664 * The blocks that were entirely written will now be uptodate, so we
665 * don't have to worry about a readpage reading them and overwriting a
666 * partial write. However if we have encountered a short write and only
667 * partially written into a block, it will not be marked uptodate, so a
668 * readpage might come in and destroy our partial write.
669 *
670 * Do the simplest thing, and just treat any short write to a non
671 * uptodate page as a zero-length write, and force the caller to redo
672 * the whole thing.
673 */
674 if (unlikely(copied < len && !PageUptodate(page)))
675 return 0;
676 iomap_set_range_uptodate(page, offset_in_page(pos), len);
677 iomap_set_page_dirty(page);
678 return copied;
679}
680
681static int
682iomap_write_end_inline(struct inode *inode, struct page *page,
683 struct iomap *iomap, loff_t pos, unsigned copied)
684{
685 void *addr;
686
687 WARN_ON_ONCE(!PageUptodate(page));
688 BUG_ON(pos + copied > PAGE_SIZE - offset_in_page(iomap->inline_data));
689
690 addr = kmap_atomic(page);
691 memcpy(iomap->inline_data + pos, addr + pos, copied);
692 kunmap_atomic(addr);
693
694 mark_inode_dirty(inode);
695 return copied;
696}
697
698static int
699iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
700 unsigned copied, struct page *page, struct iomap *iomap)
701{
702 const struct iomap_page_ops *page_ops = iomap->page_ops;
703 loff_t old_size = inode->i_size;
704 int ret;
705
706 if (iomap->type == IOMAP_INLINE) {
707 ret = iomap_write_end_inline(inode, page, iomap, pos, copied);
708 } else if (iomap->flags & IOMAP_F_BUFFER_HEAD) {
709 ret = block_write_end(NULL, inode->i_mapping, pos, len, copied,
710 page, NULL);
711 } else {
712 ret = __iomap_write_end(inode, pos, len, copied, page, iomap);
713 }
714
715 /*
716 * Update the in-memory inode size after copying the data into the page
717 * cache. It's up to the file system to write the updated size to disk,
718 * preferably after I/O completion so that no stale data is exposed.
719 */
720 if (pos + ret > old_size) {
721 i_size_write(inode, pos + ret);
722 iomap->flags |= IOMAP_F_SIZE_CHANGED;
723 }
724 unlock_page(page);
725
726 if (old_size < pos)
727 pagecache_isize_extended(inode, old_size, pos);
728 if (page_ops && page_ops->page_done)
729 page_ops->page_done(inode, pos, ret, page, iomap);
730 put_page(page);
731
732 if (ret < len)
733 iomap_write_failed(inode, pos, len);
734 return ret;
735}
736
737static loff_t
738iomap_write_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
739 struct iomap *iomap)
740{
741 struct iov_iter *i = data;
742 long status = 0;
743 ssize_t written = 0;
744 unsigned int flags = AOP_FLAG_NOFS;
745
746 do {
747 struct page *page;
748 unsigned long offset; /* Offset into pagecache page */
749 unsigned long bytes; /* Bytes to write to page */
750 size_t copied; /* Bytes copied from user */
751
752 offset = offset_in_page(pos);
753 bytes = min_t(unsigned long, PAGE_SIZE - offset,
754 iov_iter_count(i));
755again:
756 if (bytes > length)
757 bytes = length;
758
759 /*
760 * Bring in the user page that we will copy from _first_.
761 * Otherwise there's a nasty deadlock on copying from the
762 * same page as we're writing to, without it being marked
763 * up-to-date.
764 *
765 * Not only is this an optimisation, but it is also required
766 * to check that the address is actually valid, when atomic
767 * usercopies are used, below.
768 */
769 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
770 status = -EFAULT;
771 break;
772 }
773
774 status = iomap_write_begin(inode, pos, bytes, flags, &page,
775 iomap);
776 if (unlikely(status))
777 break;
778
779 if (mapping_writably_mapped(inode->i_mapping))
780 flush_dcache_page(page);
781
782 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
783
784 flush_dcache_page(page);
785
786 status = iomap_write_end(inode, pos, bytes, copied, page,
787 iomap);
788 if (unlikely(status < 0))
789 break;
790 copied = status;
791
792 cond_resched();
793
794 iov_iter_advance(i, copied);
795 if (unlikely(copied == 0)) {
796 /*
797 * If we were unable to copy any data at all, we must
798 * fall back to a single segment length write.
799 *
800 * If we didn't fallback here, we could livelock
801 * because not all segments in the iov can be copied at
802 * once without a pagefault.
803 */
804 bytes = min_t(unsigned long, PAGE_SIZE - offset,
805 iov_iter_single_seg_count(i));
806 goto again;
807 }
808 pos += copied;
809 written += copied;
810 length -= copied;
811
812 balance_dirty_pages_ratelimited(inode->i_mapping);
813 } while (iov_iter_count(i) && length);
814
815 return written ? written : status;
816}
817
818ssize_t
819iomap_file_buffered_write(struct kiocb *iocb, struct iov_iter *iter,
820 const struct iomap_ops *ops)
821{
822 struct inode *inode = iocb->ki_filp->f_mapping->host;
823 loff_t pos = iocb->ki_pos, ret = 0, written = 0;
824
825 while (iov_iter_count(iter)) {
826 ret = iomap_apply(inode, pos, iov_iter_count(iter),
827 IOMAP_WRITE, ops, iter, iomap_write_actor);
828 if (ret <= 0)
829 break;
830 pos += ret;
831 written += ret;
832 }
833
834 return written ? written : ret;
835}
836EXPORT_SYMBOL_GPL(iomap_file_buffered_write);
837
838static struct page *
839__iomap_read_page(struct inode *inode, loff_t offset)
840{
841 struct address_space *mapping = inode->i_mapping;
842 struct page *page;
843
844 page = read_mapping_page(mapping, offset >> PAGE_SHIFT, NULL);
845 if (IS_ERR(page))
846 return page;
847 if (!PageUptodate(page)) {
848 put_page(page);
849 return ERR_PTR(-EIO);
850 }
851 return page;
852}
853
854static loff_t
855iomap_dirty_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
856 struct iomap *iomap)
857{
858 long status = 0;
859 ssize_t written = 0;
860
861 do {
862 struct page *page, *rpage;
863 unsigned long offset; /* Offset into pagecache page */
864 unsigned long bytes; /* Bytes to write to page */
865
866 offset = offset_in_page(pos);
867 bytes = min_t(loff_t, PAGE_SIZE - offset, length);
868
869 rpage = __iomap_read_page(inode, pos);
870 if (IS_ERR(rpage))
871 return PTR_ERR(rpage);
872
873 status = iomap_write_begin(inode, pos, bytes,
874 AOP_FLAG_NOFS, &page, iomap);
875 put_page(rpage);
876 if (unlikely(status))
877 return status;
878
879 WARN_ON_ONCE(!PageUptodate(page));
880
881 status = iomap_write_end(inode, pos, bytes, bytes, page, iomap);
882 if (unlikely(status <= 0)) {
883 if (WARN_ON_ONCE(status == 0))
884 return -EIO;
885 return status;
886 }
887
888 cond_resched();
889
890 pos += status;
891 written += status;
892 length -= status;
893
894 balance_dirty_pages_ratelimited(inode->i_mapping);
895 } while (length);
896
897 return written;
898}
899
900int
901iomap_file_dirty(struct inode *inode, loff_t pos, loff_t len,
902 const struct iomap_ops *ops)
903{
904 loff_t ret;
905
906 while (len) {
907 ret = iomap_apply(inode, pos, len, IOMAP_WRITE, ops, NULL,
908 iomap_dirty_actor);
909 if (ret <= 0)
910 return ret;
911 pos += ret;
912 len -= ret;
913 }
914
915 return 0;
916}
917EXPORT_SYMBOL_GPL(iomap_file_dirty);
918
919static int iomap_zero(struct inode *inode, loff_t pos, unsigned offset,
920 unsigned bytes, struct iomap *iomap)
921{
922 struct page *page;
923 int status;
924
925 status = iomap_write_begin(inode, pos, bytes, AOP_FLAG_NOFS, &page,
926 iomap);
927 if (status)
928 return status;
929
930 zero_user(page, offset, bytes);
931 mark_page_accessed(page);
932
933 return iomap_write_end(inode, pos, bytes, bytes, page, iomap);
934}
935
936static int iomap_dax_zero(loff_t pos, unsigned offset, unsigned bytes,
937 struct iomap *iomap)
938{
939 return __dax_zero_page_range(iomap->bdev, iomap->dax_dev,
940 iomap_sector(iomap, pos & PAGE_MASK), offset, bytes);
941}
942
943static loff_t
944iomap_zero_range_actor(struct inode *inode, loff_t pos, loff_t count,
945 void *data, struct iomap *iomap)
946{
947 bool *did_zero = data;
948 loff_t written = 0;
949 int status;
950
951 /* already zeroed? we're done. */
952 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
953 return count;
954
955 do {
956 unsigned offset, bytes;
957
958 offset = offset_in_page(pos);
959 bytes = min_t(loff_t, PAGE_SIZE - offset, count);
960
961 if (IS_DAX(inode))
962 status = iomap_dax_zero(pos, offset, bytes, iomap);
963 else
964 status = iomap_zero(inode, pos, offset, bytes, iomap);
965 if (status < 0)
966 return status;
967
968 pos += bytes;
969 count -= bytes;
970 written += bytes;
971 if (did_zero)
972 *did_zero = true;
973 } while (count > 0);
974
975 return written;
976}
977
978int
979iomap_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
980 const struct iomap_ops *ops)
981{
982 loff_t ret;
983
984 while (len > 0) {
985 ret = iomap_apply(inode, pos, len, IOMAP_ZERO,
986 ops, did_zero, iomap_zero_range_actor);
987 if (ret <= 0)
988 return ret;
989
990 pos += ret;
991 len -= ret;
992 }
993
994 return 0;
995}
996EXPORT_SYMBOL_GPL(iomap_zero_range);
997
998int
999iomap_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
1000 const struct iomap_ops *ops)
1001{
1002 unsigned int blocksize = i_blocksize(inode);
1003 unsigned int off = pos & (blocksize - 1);
1004
1005 /* Block boundary? Nothing to do */
1006 if (!off)
1007 return 0;
1008 return iomap_zero_range(inode, pos, blocksize - off, did_zero, ops);
1009}
1010EXPORT_SYMBOL_GPL(iomap_truncate_page);
1011
1012static loff_t
1013iomap_page_mkwrite_actor(struct inode *inode, loff_t pos, loff_t length,
1014 void *data, struct iomap *iomap)
1015{
1016 struct page *page = data;
1017 int ret;
1018
1019 if (iomap->flags & IOMAP_F_BUFFER_HEAD) {
1020 ret = __block_write_begin_int(page, pos, length, NULL, iomap);
1021 if (ret)
1022 return ret;
1023 block_commit_write(page, 0, length);
1024 } else {
1025 WARN_ON_ONCE(!PageUptodate(page));
1026 iomap_page_create(inode, page);
1027 set_page_dirty(page);
1028 }
1029
1030 return length;
1031}
1032
1033vm_fault_t iomap_page_mkwrite(struct vm_fault *vmf, const struct iomap_ops *ops)
1034{
1035 struct page *page = vmf->page;
1036 struct inode *inode = file_inode(vmf->vma->vm_file);
1037 unsigned long length;
1038 loff_t offset, size;
1039 ssize_t ret;
1040
1041 lock_page(page);
1042 size = i_size_read(inode);
1043 if ((page->mapping != inode->i_mapping) ||
1044 (page_offset(page) > size)) {
1045 /* We overload EFAULT to mean page got truncated */
1046 ret = -EFAULT;
1047 goto out_unlock;
1048 }
1049
1050 /* page is wholly or partially inside EOF */
1051 if (((page->index + 1) << PAGE_SHIFT) > size)
1052 length = offset_in_page(size);
1053 else
1054 length = PAGE_SIZE;
1055
1056 offset = page_offset(page);
1057 while (length > 0) {
1058 ret = iomap_apply(inode, offset, length,
1059 IOMAP_WRITE | IOMAP_FAULT, ops, page,
1060 iomap_page_mkwrite_actor);
1061 if (unlikely(ret <= 0))
1062 goto out_unlock;
1063 offset += ret;
1064 length -= ret;
1065 }
1066
1067 wait_for_stable_page(page);
1068 return VM_FAULT_LOCKED;
1069out_unlock:
1070 unlock_page(page);
1071 return block_page_mkwrite_return(ret);
1072}
1073EXPORT_SYMBOL_GPL(iomap_page_mkwrite);
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2010 Red Hat, Inc.
4 * Copyright (C) 2016-2023 Christoph Hellwig.
5 */
6#include <linux/module.h>
7#include <linux/compiler.h>
8#include <linux/fs.h>
9#include <linux/iomap.h>
10#include <linux/pagemap.h>
11#include <linux/uio.h>
12#include <linux/buffer_head.h>
13#include <linux/dax.h>
14#include <linux/writeback.h>
15#include <linux/list_sort.h>
16#include <linux/swap.h>
17#include <linux/bio.h>
18#include <linux/sched/signal.h>
19#include <linux/migrate.h>
20#include "trace.h"
21
22#include "../internal.h"
23
24#define IOEND_BATCH_SIZE 4096
25
26/*
27 * Structure allocated for each folio to track per-block uptodate, dirty state
28 * and I/O completions.
29 */
30struct iomap_folio_state {
31 spinlock_t state_lock;
32 unsigned int read_bytes_pending;
33 atomic_t write_bytes_pending;
34
35 /*
36 * Each block has two bits in this bitmap:
37 * Bits [0..blocks_per_folio) has the uptodate status.
38 * Bits [b_p_f...(2*b_p_f)) has the dirty status.
39 */
40 unsigned long state[];
41};
42
43static struct bio_set iomap_ioend_bioset;
44
45static inline bool ifs_is_fully_uptodate(struct folio *folio,
46 struct iomap_folio_state *ifs)
47{
48 struct inode *inode = folio->mapping->host;
49
50 return bitmap_full(ifs->state, i_blocks_per_folio(inode, folio));
51}
52
53static inline bool ifs_block_is_uptodate(struct iomap_folio_state *ifs,
54 unsigned int block)
55{
56 return test_bit(block, ifs->state);
57}
58
59static bool ifs_set_range_uptodate(struct folio *folio,
60 struct iomap_folio_state *ifs, size_t off, size_t len)
61{
62 struct inode *inode = folio->mapping->host;
63 unsigned int first_blk = off >> inode->i_blkbits;
64 unsigned int last_blk = (off + len - 1) >> inode->i_blkbits;
65 unsigned int nr_blks = last_blk - first_blk + 1;
66
67 bitmap_set(ifs->state, first_blk, nr_blks);
68 return ifs_is_fully_uptodate(folio, ifs);
69}
70
71static void iomap_set_range_uptodate(struct folio *folio, size_t off,
72 size_t len)
73{
74 struct iomap_folio_state *ifs = folio->private;
75 unsigned long flags;
76 bool uptodate = true;
77
78 if (ifs) {
79 spin_lock_irqsave(&ifs->state_lock, flags);
80 uptodate = ifs_set_range_uptodate(folio, ifs, off, len);
81 spin_unlock_irqrestore(&ifs->state_lock, flags);
82 }
83
84 if (uptodate)
85 folio_mark_uptodate(folio);
86}
87
88static inline bool ifs_block_is_dirty(struct folio *folio,
89 struct iomap_folio_state *ifs, int block)
90{
91 struct inode *inode = folio->mapping->host;
92 unsigned int blks_per_folio = i_blocks_per_folio(inode, folio);
93
94 return test_bit(block + blks_per_folio, ifs->state);
95}
96
97static unsigned ifs_find_dirty_range(struct folio *folio,
98 struct iomap_folio_state *ifs, u64 *range_start, u64 range_end)
99{
100 struct inode *inode = folio->mapping->host;
101 unsigned start_blk =
102 offset_in_folio(folio, *range_start) >> inode->i_blkbits;
103 unsigned end_blk = min_not_zero(
104 offset_in_folio(folio, range_end) >> inode->i_blkbits,
105 i_blocks_per_folio(inode, folio));
106 unsigned nblks = 1;
107
108 while (!ifs_block_is_dirty(folio, ifs, start_blk))
109 if (++start_blk == end_blk)
110 return 0;
111
112 while (start_blk + nblks < end_blk) {
113 if (!ifs_block_is_dirty(folio, ifs, start_blk + nblks))
114 break;
115 nblks++;
116 }
117
118 *range_start = folio_pos(folio) + (start_blk << inode->i_blkbits);
119 return nblks << inode->i_blkbits;
120}
121
122static unsigned iomap_find_dirty_range(struct folio *folio, u64 *range_start,
123 u64 range_end)
124{
125 struct iomap_folio_state *ifs = folio->private;
126
127 if (*range_start >= range_end)
128 return 0;
129
130 if (ifs)
131 return ifs_find_dirty_range(folio, ifs, range_start, range_end);
132 return range_end - *range_start;
133}
134
135static void ifs_clear_range_dirty(struct folio *folio,
136 struct iomap_folio_state *ifs, size_t off, size_t len)
137{
138 struct inode *inode = folio->mapping->host;
139 unsigned int blks_per_folio = i_blocks_per_folio(inode, folio);
140 unsigned int first_blk = (off >> inode->i_blkbits);
141 unsigned int last_blk = (off + len - 1) >> inode->i_blkbits;
142 unsigned int nr_blks = last_blk - first_blk + 1;
143 unsigned long flags;
144
145 spin_lock_irqsave(&ifs->state_lock, flags);
146 bitmap_clear(ifs->state, first_blk + blks_per_folio, nr_blks);
147 spin_unlock_irqrestore(&ifs->state_lock, flags);
148}
149
150static void iomap_clear_range_dirty(struct folio *folio, size_t off, size_t len)
151{
152 struct iomap_folio_state *ifs = folio->private;
153
154 if (ifs)
155 ifs_clear_range_dirty(folio, ifs, off, len);
156}
157
158static void ifs_set_range_dirty(struct folio *folio,
159 struct iomap_folio_state *ifs, size_t off, size_t len)
160{
161 struct inode *inode = folio->mapping->host;
162 unsigned int blks_per_folio = i_blocks_per_folio(inode, folio);
163 unsigned int first_blk = (off >> inode->i_blkbits);
164 unsigned int last_blk = (off + len - 1) >> inode->i_blkbits;
165 unsigned int nr_blks = last_blk - first_blk + 1;
166 unsigned long flags;
167
168 spin_lock_irqsave(&ifs->state_lock, flags);
169 bitmap_set(ifs->state, first_blk + blks_per_folio, nr_blks);
170 spin_unlock_irqrestore(&ifs->state_lock, flags);
171}
172
173static void iomap_set_range_dirty(struct folio *folio, size_t off, size_t len)
174{
175 struct iomap_folio_state *ifs = folio->private;
176
177 if (ifs)
178 ifs_set_range_dirty(folio, ifs, off, len);
179}
180
181static struct iomap_folio_state *ifs_alloc(struct inode *inode,
182 struct folio *folio, unsigned int flags)
183{
184 struct iomap_folio_state *ifs = folio->private;
185 unsigned int nr_blocks = i_blocks_per_folio(inode, folio);
186 gfp_t gfp;
187
188 if (ifs || nr_blocks <= 1)
189 return ifs;
190
191 if (flags & IOMAP_NOWAIT)
192 gfp = GFP_NOWAIT;
193 else
194 gfp = GFP_NOFS | __GFP_NOFAIL;
195
196 /*
197 * ifs->state tracks two sets of state flags when the
198 * filesystem block size is smaller than the folio size.
199 * The first state tracks per-block uptodate and the
200 * second tracks per-block dirty state.
201 */
202 ifs = kzalloc(struct_size(ifs, state,
203 BITS_TO_LONGS(2 * nr_blocks)), gfp);
204 if (!ifs)
205 return ifs;
206
207 spin_lock_init(&ifs->state_lock);
208 if (folio_test_uptodate(folio))
209 bitmap_set(ifs->state, 0, nr_blocks);
210 if (folio_test_dirty(folio))
211 bitmap_set(ifs->state, nr_blocks, nr_blocks);
212 folio_attach_private(folio, ifs);
213
214 return ifs;
215}
216
217static void ifs_free(struct folio *folio)
218{
219 struct iomap_folio_state *ifs = folio_detach_private(folio);
220
221 if (!ifs)
222 return;
223 WARN_ON_ONCE(ifs->read_bytes_pending != 0);
224 WARN_ON_ONCE(atomic_read(&ifs->write_bytes_pending));
225 WARN_ON_ONCE(ifs_is_fully_uptodate(folio, ifs) !=
226 folio_test_uptodate(folio));
227 kfree(ifs);
228}
229
230/*
231 * Calculate the range inside the folio that we actually need to read.
232 */
233static void iomap_adjust_read_range(struct inode *inode, struct folio *folio,
234 loff_t *pos, loff_t length, size_t *offp, size_t *lenp)
235{
236 struct iomap_folio_state *ifs = folio->private;
237 loff_t orig_pos = *pos;
238 loff_t isize = i_size_read(inode);
239 unsigned block_bits = inode->i_blkbits;
240 unsigned block_size = (1 << block_bits);
241 size_t poff = offset_in_folio(folio, *pos);
242 size_t plen = min_t(loff_t, folio_size(folio) - poff, length);
243 size_t orig_plen = plen;
244 unsigned first = poff >> block_bits;
245 unsigned last = (poff + plen - 1) >> block_bits;
246
247 /*
248 * If the block size is smaller than the page size, we need to check the
249 * per-block uptodate status and adjust the offset and length if needed
250 * to avoid reading in already uptodate ranges.
251 */
252 if (ifs) {
253 unsigned int i;
254
255 /* move forward for each leading block marked uptodate */
256 for (i = first; i <= last; i++) {
257 if (!ifs_block_is_uptodate(ifs, i))
258 break;
259 *pos += block_size;
260 poff += block_size;
261 plen -= block_size;
262 first++;
263 }
264
265 /* truncate len if we find any trailing uptodate block(s) */
266 for ( ; i <= last; i++) {
267 if (ifs_block_is_uptodate(ifs, i)) {
268 plen -= (last - i + 1) * block_size;
269 last = i - 1;
270 break;
271 }
272 }
273 }
274
275 /*
276 * If the extent spans the block that contains the i_size, we need to
277 * handle both halves separately so that we properly zero data in the
278 * page cache for blocks that are entirely outside of i_size.
279 */
280 if (orig_pos <= isize && orig_pos + orig_plen > isize) {
281 unsigned end = offset_in_folio(folio, isize - 1) >> block_bits;
282
283 if (first <= end && last > end)
284 plen -= (last - end) * block_size;
285 }
286
287 *offp = poff;
288 *lenp = plen;
289}
290
291static void iomap_finish_folio_read(struct folio *folio, size_t off,
292 size_t len, int error)
293{
294 struct iomap_folio_state *ifs = folio->private;
295 bool uptodate = !error;
296 bool finished = true;
297
298 if (ifs) {
299 unsigned long flags;
300
301 spin_lock_irqsave(&ifs->state_lock, flags);
302 if (!error)
303 uptodate = ifs_set_range_uptodate(folio, ifs, off, len);
304 ifs->read_bytes_pending -= len;
305 finished = !ifs->read_bytes_pending;
306 spin_unlock_irqrestore(&ifs->state_lock, flags);
307 }
308
309 if (finished)
310 folio_end_read(folio, uptodate);
311}
312
313static void iomap_read_end_io(struct bio *bio)
314{
315 int error = blk_status_to_errno(bio->bi_status);
316 struct folio_iter fi;
317
318 bio_for_each_folio_all(fi, bio)
319 iomap_finish_folio_read(fi.folio, fi.offset, fi.length, error);
320 bio_put(bio);
321}
322
323struct iomap_readpage_ctx {
324 struct folio *cur_folio;
325 bool cur_folio_in_bio;
326 struct bio *bio;
327 struct readahead_control *rac;
328};
329
330/**
331 * iomap_read_inline_data - copy inline data into the page cache
332 * @iter: iteration structure
333 * @folio: folio to copy to
334 *
335 * Copy the inline data in @iter into @folio and zero out the rest of the folio.
336 * Only a single IOMAP_INLINE extent is allowed at the end of each file.
337 * Returns zero for success to complete the read, or the usual negative errno.
338 */
339static int iomap_read_inline_data(const struct iomap_iter *iter,
340 struct folio *folio)
341{
342 const struct iomap *iomap = iomap_iter_srcmap(iter);
343 size_t size = i_size_read(iter->inode) - iomap->offset;
344 size_t offset = offset_in_folio(folio, iomap->offset);
345
346 if (folio_test_uptodate(folio))
347 return 0;
348
349 if (WARN_ON_ONCE(size > iomap->length))
350 return -EIO;
351 if (offset > 0)
352 ifs_alloc(iter->inode, folio, iter->flags);
353
354 folio_fill_tail(folio, offset, iomap->inline_data, size);
355 iomap_set_range_uptodate(folio, offset, folio_size(folio) - offset);
356 return 0;
357}
358
359static inline bool iomap_block_needs_zeroing(const struct iomap_iter *iter,
360 loff_t pos)
361{
362 const struct iomap *srcmap = iomap_iter_srcmap(iter);
363
364 return srcmap->type != IOMAP_MAPPED ||
365 (srcmap->flags & IOMAP_F_NEW) ||
366 pos >= i_size_read(iter->inode);
367}
368
369static loff_t iomap_readpage_iter(const struct iomap_iter *iter,
370 struct iomap_readpage_ctx *ctx, loff_t offset)
371{
372 const struct iomap *iomap = &iter->iomap;
373 loff_t pos = iter->pos + offset;
374 loff_t length = iomap_length(iter) - offset;
375 struct folio *folio = ctx->cur_folio;
376 struct iomap_folio_state *ifs;
377 loff_t orig_pos = pos;
378 size_t poff, plen;
379 sector_t sector;
380
381 if (iomap->type == IOMAP_INLINE)
382 return iomap_read_inline_data(iter, folio);
383
384 /* zero post-eof blocks as the page may be mapped */
385 ifs = ifs_alloc(iter->inode, folio, iter->flags);
386 iomap_adjust_read_range(iter->inode, folio, &pos, length, &poff, &plen);
387 if (plen == 0)
388 goto done;
389
390 if (iomap_block_needs_zeroing(iter, pos)) {
391 folio_zero_range(folio, poff, plen);
392 iomap_set_range_uptodate(folio, poff, plen);
393 goto done;
394 }
395
396 ctx->cur_folio_in_bio = true;
397 if (ifs) {
398 spin_lock_irq(&ifs->state_lock);
399 ifs->read_bytes_pending += plen;
400 spin_unlock_irq(&ifs->state_lock);
401 }
402
403 sector = iomap_sector(iomap, pos);
404 if (!ctx->bio ||
405 bio_end_sector(ctx->bio) != sector ||
406 !bio_add_folio(ctx->bio, folio, plen, poff)) {
407 gfp_t gfp = mapping_gfp_constraint(folio->mapping, GFP_KERNEL);
408 gfp_t orig_gfp = gfp;
409 unsigned int nr_vecs = DIV_ROUND_UP(length, PAGE_SIZE);
410
411 if (ctx->bio)
412 submit_bio(ctx->bio);
413
414 if (ctx->rac) /* same as readahead_gfp_mask */
415 gfp |= __GFP_NORETRY | __GFP_NOWARN;
416 ctx->bio = bio_alloc(iomap->bdev, bio_max_segs(nr_vecs),
417 REQ_OP_READ, gfp);
418 /*
419 * If the bio_alloc fails, try it again for a single page to
420 * avoid having to deal with partial page reads. This emulates
421 * what do_mpage_read_folio does.
422 */
423 if (!ctx->bio) {
424 ctx->bio = bio_alloc(iomap->bdev, 1, REQ_OP_READ,
425 orig_gfp);
426 }
427 if (ctx->rac)
428 ctx->bio->bi_opf |= REQ_RAHEAD;
429 ctx->bio->bi_iter.bi_sector = sector;
430 ctx->bio->bi_end_io = iomap_read_end_io;
431 bio_add_folio_nofail(ctx->bio, folio, plen, poff);
432 }
433
434done:
435 /*
436 * Move the caller beyond our range so that it keeps making progress.
437 * For that, we have to include any leading non-uptodate ranges, but
438 * we can skip trailing ones as they will be handled in the next
439 * iteration.
440 */
441 return pos - orig_pos + plen;
442}
443
444static loff_t iomap_read_folio_iter(const struct iomap_iter *iter,
445 struct iomap_readpage_ctx *ctx)
446{
447 struct folio *folio = ctx->cur_folio;
448 size_t offset = offset_in_folio(folio, iter->pos);
449 loff_t length = min_t(loff_t, folio_size(folio) - offset,
450 iomap_length(iter));
451 loff_t done, ret;
452
453 for (done = 0; done < length; done += ret) {
454 ret = iomap_readpage_iter(iter, ctx, done);
455 if (ret <= 0)
456 return ret;
457 }
458
459 return done;
460}
461
462int iomap_read_folio(struct folio *folio, const struct iomap_ops *ops)
463{
464 struct iomap_iter iter = {
465 .inode = folio->mapping->host,
466 .pos = folio_pos(folio),
467 .len = folio_size(folio),
468 };
469 struct iomap_readpage_ctx ctx = {
470 .cur_folio = folio,
471 };
472 int ret;
473
474 trace_iomap_readpage(iter.inode, 1);
475
476 while ((ret = iomap_iter(&iter, ops)) > 0)
477 iter.processed = iomap_read_folio_iter(&iter, &ctx);
478
479 if (ctx.bio) {
480 submit_bio(ctx.bio);
481 WARN_ON_ONCE(!ctx.cur_folio_in_bio);
482 } else {
483 WARN_ON_ONCE(ctx.cur_folio_in_bio);
484 folio_unlock(folio);
485 }
486
487 /*
488 * Just like mpage_readahead and block_read_full_folio, we always
489 * return 0 and just set the folio error flag on errors. This
490 * should be cleaned up throughout the stack eventually.
491 */
492 return 0;
493}
494EXPORT_SYMBOL_GPL(iomap_read_folio);
495
496static loff_t iomap_readahead_iter(const struct iomap_iter *iter,
497 struct iomap_readpage_ctx *ctx)
498{
499 loff_t length = iomap_length(iter);
500 loff_t done, ret;
501
502 for (done = 0; done < length; done += ret) {
503 if (ctx->cur_folio &&
504 offset_in_folio(ctx->cur_folio, iter->pos + done) == 0) {
505 if (!ctx->cur_folio_in_bio)
506 folio_unlock(ctx->cur_folio);
507 ctx->cur_folio = NULL;
508 }
509 if (!ctx->cur_folio) {
510 ctx->cur_folio = readahead_folio(ctx->rac);
511 ctx->cur_folio_in_bio = false;
512 }
513 ret = iomap_readpage_iter(iter, ctx, done);
514 if (ret <= 0)
515 return ret;
516 }
517
518 return done;
519}
520
521/**
522 * iomap_readahead - Attempt to read pages from a file.
523 * @rac: Describes the pages to be read.
524 * @ops: The operations vector for the filesystem.
525 *
526 * This function is for filesystems to call to implement their readahead
527 * address_space operation.
528 *
529 * Context: The @ops callbacks may submit I/O (eg to read the addresses of
530 * blocks from disc), and may wait for it. The caller may be trying to
531 * access a different page, and so sleeping excessively should be avoided.
532 * It may allocate memory, but should avoid costly allocations. This
533 * function is called with memalloc_nofs set, so allocations will not cause
534 * the filesystem to be reentered.
535 */
536void iomap_readahead(struct readahead_control *rac, const struct iomap_ops *ops)
537{
538 struct iomap_iter iter = {
539 .inode = rac->mapping->host,
540 .pos = readahead_pos(rac),
541 .len = readahead_length(rac),
542 };
543 struct iomap_readpage_ctx ctx = {
544 .rac = rac,
545 };
546
547 trace_iomap_readahead(rac->mapping->host, readahead_count(rac));
548
549 while (iomap_iter(&iter, ops) > 0)
550 iter.processed = iomap_readahead_iter(&iter, &ctx);
551
552 if (ctx.bio)
553 submit_bio(ctx.bio);
554 if (ctx.cur_folio) {
555 if (!ctx.cur_folio_in_bio)
556 folio_unlock(ctx.cur_folio);
557 }
558}
559EXPORT_SYMBOL_GPL(iomap_readahead);
560
561/*
562 * iomap_is_partially_uptodate checks whether blocks within a folio are
563 * uptodate or not.
564 *
565 * Returns true if all blocks which correspond to the specified part
566 * of the folio are uptodate.
567 */
568bool iomap_is_partially_uptodate(struct folio *folio, size_t from, size_t count)
569{
570 struct iomap_folio_state *ifs = folio->private;
571 struct inode *inode = folio->mapping->host;
572 unsigned first, last, i;
573
574 if (!ifs)
575 return false;
576
577 /* Caller's range may extend past the end of this folio */
578 count = min(folio_size(folio) - from, count);
579
580 /* First and last blocks in range within folio */
581 first = from >> inode->i_blkbits;
582 last = (from + count - 1) >> inode->i_blkbits;
583
584 for (i = first; i <= last; i++)
585 if (!ifs_block_is_uptodate(ifs, i))
586 return false;
587 return true;
588}
589EXPORT_SYMBOL_GPL(iomap_is_partially_uptodate);
590
591/**
592 * iomap_get_folio - get a folio reference for writing
593 * @iter: iteration structure
594 * @pos: start offset of write
595 * @len: Suggested size of folio to create.
596 *
597 * Returns a locked reference to the folio at @pos, or an error pointer if the
598 * folio could not be obtained.
599 */
600struct folio *iomap_get_folio(struct iomap_iter *iter, loff_t pos, size_t len)
601{
602 fgf_t fgp = FGP_WRITEBEGIN | FGP_NOFS;
603
604 if (iter->flags & IOMAP_NOWAIT)
605 fgp |= FGP_NOWAIT;
606 fgp |= fgf_set_order(len);
607
608 return __filemap_get_folio(iter->inode->i_mapping, pos >> PAGE_SHIFT,
609 fgp, mapping_gfp_mask(iter->inode->i_mapping));
610}
611EXPORT_SYMBOL_GPL(iomap_get_folio);
612
613bool iomap_release_folio(struct folio *folio, gfp_t gfp_flags)
614{
615 trace_iomap_release_folio(folio->mapping->host, folio_pos(folio),
616 folio_size(folio));
617
618 /*
619 * If the folio is dirty, we refuse to release our metadata because
620 * it may be partially dirty. Once we track per-block dirty state,
621 * we can release the metadata if every block is dirty.
622 */
623 if (folio_test_dirty(folio))
624 return false;
625 ifs_free(folio);
626 return true;
627}
628EXPORT_SYMBOL_GPL(iomap_release_folio);
629
630void iomap_invalidate_folio(struct folio *folio, size_t offset, size_t len)
631{
632 trace_iomap_invalidate_folio(folio->mapping->host,
633 folio_pos(folio) + offset, len);
634
635 /*
636 * If we're invalidating the entire folio, clear the dirty state
637 * from it and release it to avoid unnecessary buildup of the LRU.
638 */
639 if (offset == 0 && len == folio_size(folio)) {
640 WARN_ON_ONCE(folio_test_writeback(folio));
641 folio_cancel_dirty(folio);
642 ifs_free(folio);
643 }
644}
645EXPORT_SYMBOL_GPL(iomap_invalidate_folio);
646
647bool iomap_dirty_folio(struct address_space *mapping, struct folio *folio)
648{
649 struct inode *inode = mapping->host;
650 size_t len = folio_size(folio);
651
652 ifs_alloc(inode, folio, 0);
653 iomap_set_range_dirty(folio, 0, len);
654 return filemap_dirty_folio(mapping, folio);
655}
656EXPORT_SYMBOL_GPL(iomap_dirty_folio);
657
658static void
659iomap_write_failed(struct inode *inode, loff_t pos, unsigned len)
660{
661 loff_t i_size = i_size_read(inode);
662
663 /*
664 * Only truncate newly allocated pages beyoned EOF, even if the
665 * write started inside the existing inode size.
666 */
667 if (pos + len > i_size)
668 truncate_pagecache_range(inode, max(pos, i_size),
669 pos + len - 1);
670}
671
672static int iomap_read_folio_sync(loff_t block_start, struct folio *folio,
673 size_t poff, size_t plen, const struct iomap *iomap)
674{
675 struct bio_vec bvec;
676 struct bio bio;
677
678 bio_init(&bio, iomap->bdev, &bvec, 1, REQ_OP_READ);
679 bio.bi_iter.bi_sector = iomap_sector(iomap, block_start);
680 bio_add_folio_nofail(&bio, folio, plen, poff);
681 return submit_bio_wait(&bio);
682}
683
684static int __iomap_write_begin(const struct iomap_iter *iter, loff_t pos,
685 size_t len, struct folio *folio)
686{
687 const struct iomap *srcmap = iomap_iter_srcmap(iter);
688 struct iomap_folio_state *ifs;
689 loff_t block_size = i_blocksize(iter->inode);
690 loff_t block_start = round_down(pos, block_size);
691 loff_t block_end = round_up(pos + len, block_size);
692 unsigned int nr_blocks = i_blocks_per_folio(iter->inode, folio);
693 size_t from = offset_in_folio(folio, pos), to = from + len;
694 size_t poff, plen;
695
696 /*
697 * If the write or zeroing completely overlaps the current folio, then
698 * entire folio will be dirtied so there is no need for
699 * per-block state tracking structures to be attached to this folio.
700 * For the unshare case, we must read in the ondisk contents because we
701 * are not changing pagecache contents.
702 */
703 if (!(iter->flags & IOMAP_UNSHARE) && pos <= folio_pos(folio) &&
704 pos + len >= folio_pos(folio) + folio_size(folio))
705 return 0;
706
707 ifs = ifs_alloc(iter->inode, folio, iter->flags);
708 if ((iter->flags & IOMAP_NOWAIT) && !ifs && nr_blocks > 1)
709 return -EAGAIN;
710
711 if (folio_test_uptodate(folio))
712 return 0;
713
714 do {
715 iomap_adjust_read_range(iter->inode, folio, &block_start,
716 block_end - block_start, &poff, &plen);
717 if (plen == 0)
718 break;
719
720 if (!(iter->flags & IOMAP_UNSHARE) &&
721 (from <= poff || from >= poff + plen) &&
722 (to <= poff || to >= poff + plen))
723 continue;
724
725 if (iomap_block_needs_zeroing(iter, block_start)) {
726 if (WARN_ON_ONCE(iter->flags & IOMAP_UNSHARE))
727 return -EIO;
728 folio_zero_segments(folio, poff, from, to, poff + plen);
729 } else {
730 int status;
731
732 if (iter->flags & IOMAP_NOWAIT)
733 return -EAGAIN;
734
735 status = iomap_read_folio_sync(block_start, folio,
736 poff, plen, srcmap);
737 if (status)
738 return status;
739 }
740 iomap_set_range_uptodate(folio, poff, plen);
741 } while ((block_start += plen) < block_end);
742
743 return 0;
744}
745
746static struct folio *__iomap_get_folio(struct iomap_iter *iter, loff_t pos,
747 size_t len)
748{
749 const struct iomap_folio_ops *folio_ops = iter->iomap.folio_ops;
750
751 if (folio_ops && folio_ops->get_folio)
752 return folio_ops->get_folio(iter, pos, len);
753 else
754 return iomap_get_folio(iter, pos, len);
755}
756
757static void __iomap_put_folio(struct iomap_iter *iter, loff_t pos, size_t ret,
758 struct folio *folio)
759{
760 const struct iomap_folio_ops *folio_ops = iter->iomap.folio_ops;
761
762 if (folio_ops && folio_ops->put_folio) {
763 folio_ops->put_folio(iter->inode, pos, ret, folio);
764 } else {
765 folio_unlock(folio);
766 folio_put(folio);
767 }
768}
769
770static int iomap_write_begin_inline(const struct iomap_iter *iter,
771 struct folio *folio)
772{
773 /* needs more work for the tailpacking case; disable for now */
774 if (WARN_ON_ONCE(iomap_iter_srcmap(iter)->offset != 0))
775 return -EIO;
776 return iomap_read_inline_data(iter, folio);
777}
778
779static int iomap_write_begin(struct iomap_iter *iter, loff_t pos,
780 size_t len, struct folio **foliop)
781{
782 const struct iomap_folio_ops *folio_ops = iter->iomap.folio_ops;
783 const struct iomap *srcmap = iomap_iter_srcmap(iter);
784 struct folio *folio;
785 int status = 0;
786
787 BUG_ON(pos + len > iter->iomap.offset + iter->iomap.length);
788 if (srcmap != &iter->iomap)
789 BUG_ON(pos + len > srcmap->offset + srcmap->length);
790
791 if (fatal_signal_pending(current))
792 return -EINTR;
793
794 if (!mapping_large_folio_support(iter->inode->i_mapping))
795 len = min_t(size_t, len, PAGE_SIZE - offset_in_page(pos));
796
797 folio = __iomap_get_folio(iter, pos, len);
798 if (IS_ERR(folio))
799 return PTR_ERR(folio);
800
801 /*
802 * Now we have a locked folio, before we do anything with it we need to
803 * check that the iomap we have cached is not stale. The inode extent
804 * mapping can change due to concurrent IO in flight (e.g.
805 * IOMAP_UNWRITTEN state can change and memory reclaim could have
806 * reclaimed a previously partially written page at this index after IO
807 * completion before this write reaches this file offset) and hence we
808 * could do the wrong thing here (zero a page range incorrectly or fail
809 * to zero) and corrupt data.
810 */
811 if (folio_ops && folio_ops->iomap_valid) {
812 bool iomap_valid = folio_ops->iomap_valid(iter->inode,
813 &iter->iomap);
814 if (!iomap_valid) {
815 iter->iomap.flags |= IOMAP_F_STALE;
816 status = 0;
817 goto out_unlock;
818 }
819 }
820
821 if (pos + len > folio_pos(folio) + folio_size(folio))
822 len = folio_pos(folio) + folio_size(folio) - pos;
823
824 if (srcmap->type == IOMAP_INLINE)
825 status = iomap_write_begin_inline(iter, folio);
826 else if (srcmap->flags & IOMAP_F_BUFFER_HEAD)
827 status = __block_write_begin_int(folio, pos, len, NULL, srcmap);
828 else
829 status = __iomap_write_begin(iter, pos, len, folio);
830
831 if (unlikely(status))
832 goto out_unlock;
833
834 *foliop = folio;
835 return 0;
836
837out_unlock:
838 __iomap_put_folio(iter, pos, 0, folio);
839
840 return status;
841}
842
843static bool __iomap_write_end(struct inode *inode, loff_t pos, size_t len,
844 size_t copied, struct folio *folio)
845{
846 flush_dcache_folio(folio);
847
848 /*
849 * The blocks that were entirely written will now be uptodate, so we
850 * don't have to worry about a read_folio reading them and overwriting a
851 * partial write. However, if we've encountered a short write and only
852 * partially written into a block, it will not be marked uptodate, so a
853 * read_folio might come in and destroy our partial write.
854 *
855 * Do the simplest thing and just treat any short write to a
856 * non-uptodate page as a zero-length write, and force the caller to
857 * redo the whole thing.
858 */
859 if (unlikely(copied < len && !folio_test_uptodate(folio)))
860 return false;
861 iomap_set_range_uptodate(folio, offset_in_folio(folio, pos), len);
862 iomap_set_range_dirty(folio, offset_in_folio(folio, pos), copied);
863 filemap_dirty_folio(inode->i_mapping, folio);
864 return true;
865}
866
867static void iomap_write_end_inline(const struct iomap_iter *iter,
868 struct folio *folio, loff_t pos, size_t copied)
869{
870 const struct iomap *iomap = &iter->iomap;
871 void *addr;
872
873 WARN_ON_ONCE(!folio_test_uptodate(folio));
874 BUG_ON(!iomap_inline_data_valid(iomap));
875
876 flush_dcache_folio(folio);
877 addr = kmap_local_folio(folio, pos);
878 memcpy(iomap_inline_data(iomap, pos), addr, copied);
879 kunmap_local(addr);
880
881 mark_inode_dirty(iter->inode);
882}
883
884/*
885 * Returns true if all copied bytes have been written to the pagecache,
886 * otherwise return false.
887 */
888static bool iomap_write_end(struct iomap_iter *iter, loff_t pos, size_t len,
889 size_t copied, struct folio *folio)
890{
891 const struct iomap *srcmap = iomap_iter_srcmap(iter);
892
893 if (srcmap->type == IOMAP_INLINE) {
894 iomap_write_end_inline(iter, folio, pos, copied);
895 return true;
896 }
897
898 if (srcmap->flags & IOMAP_F_BUFFER_HEAD) {
899 size_t bh_written;
900
901 bh_written = block_write_end(NULL, iter->inode->i_mapping, pos,
902 len, copied, folio, NULL);
903 WARN_ON_ONCE(bh_written != copied && bh_written != 0);
904 return bh_written == copied;
905 }
906
907 return __iomap_write_end(iter->inode, pos, len, copied, folio);
908}
909
910static loff_t iomap_write_iter(struct iomap_iter *iter, struct iov_iter *i)
911{
912 loff_t length = iomap_length(iter);
913 loff_t pos = iter->pos;
914 ssize_t total_written = 0;
915 long status = 0;
916 struct address_space *mapping = iter->inode->i_mapping;
917 size_t chunk = mapping_max_folio_size(mapping);
918 unsigned int bdp_flags = (iter->flags & IOMAP_NOWAIT) ? BDP_ASYNC : 0;
919
920 do {
921 struct folio *folio;
922 loff_t old_size;
923 size_t offset; /* Offset into folio */
924 size_t bytes; /* Bytes to write to folio */
925 size_t copied; /* Bytes copied from user */
926 size_t written; /* Bytes have been written */
927
928 bytes = iov_iter_count(i);
929retry:
930 offset = pos & (chunk - 1);
931 bytes = min(chunk - offset, bytes);
932 status = balance_dirty_pages_ratelimited_flags(mapping,
933 bdp_flags);
934 if (unlikely(status))
935 break;
936
937 if (bytes > length)
938 bytes = length;
939
940 /*
941 * Bring in the user page that we'll copy from _first_.
942 * Otherwise there's a nasty deadlock on copying from the
943 * same page as we're writing to, without it being marked
944 * up-to-date.
945 *
946 * For async buffered writes the assumption is that the user
947 * page has already been faulted in. This can be optimized by
948 * faulting the user page.
949 */
950 if (unlikely(fault_in_iov_iter_readable(i, bytes) == bytes)) {
951 status = -EFAULT;
952 break;
953 }
954
955 status = iomap_write_begin(iter, pos, bytes, &folio);
956 if (unlikely(status)) {
957 iomap_write_failed(iter->inode, pos, bytes);
958 break;
959 }
960 if (iter->iomap.flags & IOMAP_F_STALE)
961 break;
962
963 offset = offset_in_folio(folio, pos);
964 if (bytes > folio_size(folio) - offset)
965 bytes = folio_size(folio) - offset;
966
967 if (mapping_writably_mapped(mapping))
968 flush_dcache_folio(folio);
969
970 copied = copy_folio_from_iter_atomic(folio, offset, bytes, i);
971 written = iomap_write_end(iter, pos, bytes, copied, folio) ?
972 copied : 0;
973
974 /*
975 * Update the in-memory inode size after copying the data into
976 * the page cache. It's up to the file system to write the
977 * updated size to disk, preferably after I/O completion so that
978 * no stale data is exposed. Only once that's done can we
979 * unlock and release the folio.
980 */
981 old_size = iter->inode->i_size;
982 if (pos + written > old_size) {
983 i_size_write(iter->inode, pos + written);
984 iter->iomap.flags |= IOMAP_F_SIZE_CHANGED;
985 }
986 __iomap_put_folio(iter, pos, written, folio);
987
988 if (old_size < pos)
989 pagecache_isize_extended(iter->inode, old_size, pos);
990
991 cond_resched();
992 if (unlikely(written == 0)) {
993 /*
994 * A short copy made iomap_write_end() reject the
995 * thing entirely. Might be memory poisoning
996 * halfway through, might be a race with munmap,
997 * might be severe memory pressure.
998 */
999 iomap_write_failed(iter->inode, pos, bytes);
1000 iov_iter_revert(i, copied);
1001
1002 if (chunk > PAGE_SIZE)
1003 chunk /= 2;
1004 if (copied) {
1005 bytes = copied;
1006 goto retry;
1007 }
1008 } else {
1009 pos += written;
1010 total_written += written;
1011 length -= written;
1012 }
1013 } while (iov_iter_count(i) && length);
1014
1015 if (status == -EAGAIN) {
1016 iov_iter_revert(i, total_written);
1017 return -EAGAIN;
1018 }
1019 return total_written ? total_written : status;
1020}
1021
1022ssize_t
1023iomap_file_buffered_write(struct kiocb *iocb, struct iov_iter *i,
1024 const struct iomap_ops *ops, void *private)
1025{
1026 struct iomap_iter iter = {
1027 .inode = iocb->ki_filp->f_mapping->host,
1028 .pos = iocb->ki_pos,
1029 .len = iov_iter_count(i),
1030 .flags = IOMAP_WRITE,
1031 .private = private,
1032 };
1033 ssize_t ret;
1034
1035 if (iocb->ki_flags & IOCB_NOWAIT)
1036 iter.flags |= IOMAP_NOWAIT;
1037
1038 while ((ret = iomap_iter(&iter, ops)) > 0)
1039 iter.processed = iomap_write_iter(&iter, i);
1040
1041 if (unlikely(iter.pos == iocb->ki_pos))
1042 return ret;
1043 ret = iter.pos - iocb->ki_pos;
1044 iocb->ki_pos = iter.pos;
1045 return ret;
1046}
1047EXPORT_SYMBOL_GPL(iomap_file_buffered_write);
1048
1049static void iomap_write_delalloc_ifs_punch(struct inode *inode,
1050 struct folio *folio, loff_t start_byte, loff_t end_byte,
1051 struct iomap *iomap, iomap_punch_t punch)
1052{
1053 unsigned int first_blk, last_blk, i;
1054 loff_t last_byte;
1055 u8 blkbits = inode->i_blkbits;
1056 struct iomap_folio_state *ifs;
1057
1058 /*
1059 * When we have per-block dirty tracking, there can be
1060 * blocks within a folio which are marked uptodate
1061 * but not dirty. In that case it is necessary to punch
1062 * out such blocks to avoid leaking any delalloc blocks.
1063 */
1064 ifs = folio->private;
1065 if (!ifs)
1066 return;
1067
1068 last_byte = min_t(loff_t, end_byte - 1,
1069 folio_pos(folio) + folio_size(folio) - 1);
1070 first_blk = offset_in_folio(folio, start_byte) >> blkbits;
1071 last_blk = offset_in_folio(folio, last_byte) >> blkbits;
1072 for (i = first_blk; i <= last_blk; i++) {
1073 if (!ifs_block_is_dirty(folio, ifs, i))
1074 punch(inode, folio_pos(folio) + (i << blkbits),
1075 1 << blkbits, iomap);
1076 }
1077}
1078
1079static void iomap_write_delalloc_punch(struct inode *inode, struct folio *folio,
1080 loff_t *punch_start_byte, loff_t start_byte, loff_t end_byte,
1081 struct iomap *iomap, iomap_punch_t punch)
1082{
1083 if (!folio_test_dirty(folio))
1084 return;
1085
1086 /* if dirty, punch up to offset */
1087 if (start_byte > *punch_start_byte) {
1088 punch(inode, *punch_start_byte, start_byte - *punch_start_byte,
1089 iomap);
1090 }
1091
1092 /* Punch non-dirty blocks within folio */
1093 iomap_write_delalloc_ifs_punch(inode, folio, start_byte, end_byte,
1094 iomap, punch);
1095
1096 /*
1097 * Make sure the next punch start is correctly bound to
1098 * the end of this data range, not the end of the folio.
1099 */
1100 *punch_start_byte = min_t(loff_t, end_byte,
1101 folio_pos(folio) + folio_size(folio));
1102}
1103
1104/*
1105 * Scan the data range passed to us for dirty page cache folios. If we find a
1106 * dirty folio, punch out the preceding range and update the offset from which
1107 * the next punch will start from.
1108 *
1109 * We can punch out storage reservations under clean pages because they either
1110 * contain data that has been written back - in which case the delalloc punch
1111 * over that range is a no-op - or they have been read faults in which case they
1112 * contain zeroes and we can remove the delalloc backing range and any new
1113 * writes to those pages will do the normal hole filling operation...
1114 *
1115 * This makes the logic simple: we only need to keep the delalloc extents only
1116 * over the dirty ranges of the page cache.
1117 *
1118 * This function uses [start_byte, end_byte) intervals (i.e. open ended) to
1119 * simplify range iterations.
1120 */
1121static void iomap_write_delalloc_scan(struct inode *inode,
1122 loff_t *punch_start_byte, loff_t start_byte, loff_t end_byte,
1123 struct iomap *iomap, iomap_punch_t punch)
1124{
1125 while (start_byte < end_byte) {
1126 struct folio *folio;
1127
1128 /* grab locked page */
1129 folio = filemap_lock_folio(inode->i_mapping,
1130 start_byte >> PAGE_SHIFT);
1131 if (IS_ERR(folio)) {
1132 start_byte = ALIGN_DOWN(start_byte, PAGE_SIZE) +
1133 PAGE_SIZE;
1134 continue;
1135 }
1136
1137 iomap_write_delalloc_punch(inode, folio, punch_start_byte,
1138 start_byte, end_byte, iomap, punch);
1139
1140 /* move offset to start of next folio in range */
1141 start_byte = folio_pos(folio) + folio_size(folio);
1142 folio_unlock(folio);
1143 folio_put(folio);
1144 }
1145}
1146
1147/*
1148 * When a short write occurs, the filesystem might need to use ->iomap_end
1149 * to remove space reservations created in ->iomap_begin.
1150 *
1151 * For filesystems that use delayed allocation, there can be dirty pages over
1152 * the delalloc extent outside the range of a short write but still within the
1153 * delalloc extent allocated for this iomap if the write raced with page
1154 * faults.
1155 *
1156 * Punch out all the delalloc blocks in the range given except for those that
1157 * have dirty data still pending in the page cache - those are going to be
1158 * written and so must still retain the delalloc backing for writeback.
1159 *
1160 * The punch() callback *must* only punch delalloc extents in the range passed
1161 * to it. It must skip over all other types of extents in the range and leave
1162 * them completely unchanged. It must do this punch atomically with respect to
1163 * other extent modifications.
1164 *
1165 * The punch() callback may be called with a folio locked to prevent writeback
1166 * extent allocation racing at the edge of the range we are currently punching.
1167 * The locked folio may or may not cover the range being punched, so it is not
1168 * safe for the punch() callback to lock folios itself.
1169 *
1170 * Lock order is:
1171 *
1172 * inode->i_rwsem (shared or exclusive)
1173 * inode->i_mapping->invalidate_lock (exclusive)
1174 * folio_lock()
1175 * ->punch
1176 * internal filesystem allocation lock
1177 *
1178 * As we are scanning the page cache for data, we don't need to reimplement the
1179 * wheel - mapping_seek_hole_data() does exactly what we need to identify the
1180 * start and end of data ranges correctly even for sub-folio block sizes. This
1181 * byte range based iteration is especially convenient because it means we
1182 * don't have to care about variable size folios, nor where the start or end of
1183 * the data range lies within a folio, if they lie within the same folio or even
1184 * if there are multiple discontiguous data ranges within the folio.
1185 *
1186 * It should be noted that mapping_seek_hole_data() is not aware of EOF, and so
1187 * can return data ranges that exist in the cache beyond EOF. e.g. a page fault
1188 * spanning EOF will initialise the post-EOF data to zeroes and mark it up to
1189 * date. A write page fault can then mark it dirty. If we then fail a write()
1190 * beyond EOF into that up to date cached range, we allocate a delalloc block
1191 * beyond EOF and then have to punch it out. Because the range is up to date,
1192 * mapping_seek_hole_data() will return it, and we will skip the punch because
1193 * the folio is dirty. THis is incorrect - we always need to punch out delalloc
1194 * beyond EOF in this case as writeback will never write back and covert that
1195 * delalloc block beyond EOF. Hence we limit the cached data scan range to EOF,
1196 * resulting in always punching out the range from the EOF to the end of the
1197 * range the iomap spans.
1198 *
1199 * Intervals are of the form [start_byte, end_byte) (i.e. open ended) because it
1200 * matches the intervals returned by mapping_seek_hole_data(). i.e. SEEK_DATA
1201 * returns the start of a data range (start_byte), and SEEK_HOLE(start_byte)
1202 * returns the end of the data range (data_end). Using closed intervals would
1203 * require sprinkling this code with magic "+ 1" and "- 1" arithmetic and expose
1204 * the code to subtle off-by-one bugs....
1205 */
1206void iomap_write_delalloc_release(struct inode *inode, loff_t start_byte,
1207 loff_t end_byte, unsigned flags, struct iomap *iomap,
1208 iomap_punch_t punch)
1209{
1210 loff_t punch_start_byte = start_byte;
1211 loff_t scan_end_byte = min(i_size_read(inode), end_byte);
1212
1213 /*
1214 * The caller must hold invalidate_lock to avoid races with page faults
1215 * re-instantiating folios and dirtying them via ->page_mkwrite whilst
1216 * we walk the cache and perform delalloc extent removal. Failing to do
1217 * this can leave dirty pages with no space reservation in the cache.
1218 */
1219 lockdep_assert_held_write(&inode->i_mapping->invalidate_lock);
1220
1221 while (start_byte < scan_end_byte) {
1222 loff_t data_end;
1223
1224 start_byte = mapping_seek_hole_data(inode->i_mapping,
1225 start_byte, scan_end_byte, SEEK_DATA);
1226 /*
1227 * If there is no more data to scan, all that is left is to
1228 * punch out the remaining range.
1229 *
1230 * Note that mapping_seek_hole_data is only supposed to return
1231 * either an offset or -ENXIO, so WARN on any other error as
1232 * that would be an API change without updating the callers.
1233 */
1234 if (start_byte == -ENXIO || start_byte == scan_end_byte)
1235 break;
1236 if (WARN_ON_ONCE(start_byte < 0))
1237 return;
1238 WARN_ON_ONCE(start_byte < punch_start_byte);
1239 WARN_ON_ONCE(start_byte > scan_end_byte);
1240
1241 /*
1242 * We find the end of this contiguous cached data range by
1243 * seeking from start_byte to the beginning of the next hole.
1244 */
1245 data_end = mapping_seek_hole_data(inode->i_mapping, start_byte,
1246 scan_end_byte, SEEK_HOLE);
1247 if (WARN_ON_ONCE(data_end < 0))
1248 return;
1249
1250 /*
1251 * If we race with post-direct I/O invalidation of the page cache,
1252 * there might be no data left at start_byte.
1253 */
1254 if (data_end == start_byte)
1255 continue;
1256
1257 WARN_ON_ONCE(data_end < start_byte);
1258 WARN_ON_ONCE(data_end > scan_end_byte);
1259
1260 iomap_write_delalloc_scan(inode, &punch_start_byte, start_byte,
1261 data_end, iomap, punch);
1262
1263 /* The next data search starts at the end of this one. */
1264 start_byte = data_end;
1265 }
1266
1267 if (punch_start_byte < end_byte)
1268 punch(inode, punch_start_byte, end_byte - punch_start_byte,
1269 iomap);
1270}
1271EXPORT_SYMBOL_GPL(iomap_write_delalloc_release);
1272
1273static loff_t iomap_unshare_iter(struct iomap_iter *iter)
1274{
1275 struct iomap *iomap = &iter->iomap;
1276 loff_t pos = iter->pos;
1277 loff_t length = iomap_length(iter);
1278 loff_t written = 0;
1279
1280 if (!iomap_want_unshare_iter(iter))
1281 return length;
1282
1283 do {
1284 struct folio *folio;
1285 int status;
1286 size_t offset;
1287 size_t bytes = min_t(u64, SIZE_MAX, length);
1288 bool ret;
1289
1290 status = iomap_write_begin(iter, pos, bytes, &folio);
1291 if (unlikely(status))
1292 return status;
1293 if (iomap->flags & IOMAP_F_STALE)
1294 break;
1295
1296 offset = offset_in_folio(folio, pos);
1297 if (bytes > folio_size(folio) - offset)
1298 bytes = folio_size(folio) - offset;
1299
1300 ret = iomap_write_end(iter, pos, bytes, bytes, folio);
1301 __iomap_put_folio(iter, pos, bytes, folio);
1302 if (WARN_ON_ONCE(!ret))
1303 return -EIO;
1304
1305 cond_resched();
1306
1307 pos += bytes;
1308 written += bytes;
1309 length -= bytes;
1310
1311 balance_dirty_pages_ratelimited(iter->inode->i_mapping);
1312 } while (length > 0);
1313
1314 return written;
1315}
1316
1317int
1318iomap_file_unshare(struct inode *inode, loff_t pos, loff_t len,
1319 const struct iomap_ops *ops)
1320{
1321 struct iomap_iter iter = {
1322 .inode = inode,
1323 .pos = pos,
1324 .flags = IOMAP_WRITE | IOMAP_UNSHARE,
1325 };
1326 loff_t size = i_size_read(inode);
1327 int ret;
1328
1329 if (pos < 0 || pos >= size)
1330 return 0;
1331
1332 iter.len = min(len, size - pos);
1333 while ((ret = iomap_iter(&iter, ops)) > 0)
1334 iter.processed = iomap_unshare_iter(&iter);
1335 return ret;
1336}
1337EXPORT_SYMBOL_GPL(iomap_file_unshare);
1338
1339/*
1340 * Flush the remaining range of the iter and mark the current mapping stale.
1341 * This is used when zero range sees an unwritten mapping that may have had
1342 * dirty pagecache over it.
1343 */
1344static inline int iomap_zero_iter_flush_and_stale(struct iomap_iter *i)
1345{
1346 struct address_space *mapping = i->inode->i_mapping;
1347 loff_t end = i->pos + i->len - 1;
1348
1349 i->iomap.flags |= IOMAP_F_STALE;
1350 return filemap_write_and_wait_range(mapping, i->pos, end);
1351}
1352
1353static loff_t iomap_zero_iter(struct iomap_iter *iter, bool *did_zero)
1354{
1355 loff_t pos = iter->pos;
1356 loff_t length = iomap_length(iter);
1357 loff_t written = 0;
1358
1359 do {
1360 struct folio *folio;
1361 int status;
1362 size_t offset;
1363 size_t bytes = min_t(u64, SIZE_MAX, length);
1364 bool ret;
1365
1366 status = iomap_write_begin(iter, pos, bytes, &folio);
1367 if (status)
1368 return status;
1369 if (iter->iomap.flags & IOMAP_F_STALE)
1370 break;
1371
1372 /* warn about zeroing folios beyond eof that won't write back */
1373 WARN_ON_ONCE(folio_pos(folio) > iter->inode->i_size);
1374 offset = offset_in_folio(folio, pos);
1375 if (bytes > folio_size(folio) - offset)
1376 bytes = folio_size(folio) - offset;
1377
1378 folio_zero_range(folio, offset, bytes);
1379 folio_mark_accessed(folio);
1380
1381 ret = iomap_write_end(iter, pos, bytes, bytes, folio);
1382 __iomap_put_folio(iter, pos, bytes, folio);
1383 if (WARN_ON_ONCE(!ret))
1384 return -EIO;
1385
1386 pos += bytes;
1387 length -= bytes;
1388 written += bytes;
1389 } while (length > 0);
1390
1391 if (did_zero)
1392 *did_zero = true;
1393 return written;
1394}
1395
1396int
1397iomap_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
1398 const struct iomap_ops *ops)
1399{
1400 struct iomap_iter iter = {
1401 .inode = inode,
1402 .pos = pos,
1403 .len = len,
1404 .flags = IOMAP_ZERO,
1405 };
1406 struct address_space *mapping = inode->i_mapping;
1407 unsigned int blocksize = i_blocksize(inode);
1408 unsigned int off = pos & (blocksize - 1);
1409 loff_t plen = min_t(loff_t, len, blocksize - off);
1410 int ret;
1411 bool range_dirty;
1412
1413 /*
1414 * Zero range can skip mappings that are zero on disk so long as
1415 * pagecache is clean. If pagecache was dirty prior to zero range, the
1416 * mapping converts on writeback completion and so must be zeroed.
1417 *
1418 * The simplest way to deal with this across a range is to flush
1419 * pagecache and process the updated mappings. To avoid excessive
1420 * flushing on partial eof zeroing, special case it to zero the
1421 * unaligned start portion if already dirty in pagecache.
1422 */
1423 if (off &&
1424 filemap_range_needs_writeback(mapping, pos, pos + plen - 1)) {
1425 iter.len = plen;
1426 while ((ret = iomap_iter(&iter, ops)) > 0)
1427 iter.processed = iomap_zero_iter(&iter, did_zero);
1428
1429 iter.len = len - (iter.pos - pos);
1430 if (ret || !iter.len)
1431 return ret;
1432 }
1433
1434 /*
1435 * To avoid an unconditional flush, check pagecache state and only flush
1436 * if dirty and the fs returns a mapping that might convert on
1437 * writeback.
1438 */
1439 range_dirty = filemap_range_needs_writeback(inode->i_mapping,
1440 iter.pos, iter.pos + iter.len - 1);
1441 while ((ret = iomap_iter(&iter, ops)) > 0) {
1442 const struct iomap *srcmap = iomap_iter_srcmap(&iter);
1443
1444 if (srcmap->type == IOMAP_HOLE ||
1445 srcmap->type == IOMAP_UNWRITTEN) {
1446 loff_t proc = iomap_length(&iter);
1447
1448 if (range_dirty) {
1449 range_dirty = false;
1450 proc = iomap_zero_iter_flush_and_stale(&iter);
1451 }
1452 iter.processed = proc;
1453 continue;
1454 }
1455
1456 iter.processed = iomap_zero_iter(&iter, did_zero);
1457 }
1458 return ret;
1459}
1460EXPORT_SYMBOL_GPL(iomap_zero_range);
1461
1462int
1463iomap_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
1464 const struct iomap_ops *ops)
1465{
1466 unsigned int blocksize = i_blocksize(inode);
1467 unsigned int off = pos & (blocksize - 1);
1468
1469 /* Block boundary? Nothing to do */
1470 if (!off)
1471 return 0;
1472 return iomap_zero_range(inode, pos, blocksize - off, did_zero, ops);
1473}
1474EXPORT_SYMBOL_GPL(iomap_truncate_page);
1475
1476static loff_t iomap_folio_mkwrite_iter(struct iomap_iter *iter,
1477 struct folio *folio)
1478{
1479 loff_t length = iomap_length(iter);
1480 int ret;
1481
1482 if (iter->iomap.flags & IOMAP_F_BUFFER_HEAD) {
1483 ret = __block_write_begin_int(folio, iter->pos, length, NULL,
1484 &iter->iomap);
1485 if (ret)
1486 return ret;
1487 block_commit_write(&folio->page, 0, length);
1488 } else {
1489 WARN_ON_ONCE(!folio_test_uptodate(folio));
1490 folio_mark_dirty(folio);
1491 }
1492
1493 return length;
1494}
1495
1496vm_fault_t iomap_page_mkwrite(struct vm_fault *vmf, const struct iomap_ops *ops)
1497{
1498 struct iomap_iter iter = {
1499 .inode = file_inode(vmf->vma->vm_file),
1500 .flags = IOMAP_WRITE | IOMAP_FAULT,
1501 };
1502 struct folio *folio = page_folio(vmf->page);
1503 ssize_t ret;
1504
1505 folio_lock(folio);
1506 ret = folio_mkwrite_check_truncate(folio, iter.inode);
1507 if (ret < 0)
1508 goto out_unlock;
1509 iter.pos = folio_pos(folio);
1510 iter.len = ret;
1511 while ((ret = iomap_iter(&iter, ops)) > 0)
1512 iter.processed = iomap_folio_mkwrite_iter(&iter, folio);
1513
1514 if (ret < 0)
1515 goto out_unlock;
1516 folio_wait_stable(folio);
1517 return VM_FAULT_LOCKED;
1518out_unlock:
1519 folio_unlock(folio);
1520 return vmf_fs_error(ret);
1521}
1522EXPORT_SYMBOL_GPL(iomap_page_mkwrite);
1523
1524static void iomap_finish_folio_write(struct inode *inode, struct folio *folio,
1525 size_t len)
1526{
1527 struct iomap_folio_state *ifs = folio->private;
1528
1529 WARN_ON_ONCE(i_blocks_per_folio(inode, folio) > 1 && !ifs);
1530 WARN_ON_ONCE(ifs && atomic_read(&ifs->write_bytes_pending) <= 0);
1531
1532 if (!ifs || atomic_sub_and_test(len, &ifs->write_bytes_pending))
1533 folio_end_writeback(folio);
1534}
1535
1536/*
1537 * We're now finished for good with this ioend structure. Update the page
1538 * state, release holds on bios, and finally free up memory. Do not use the
1539 * ioend after this.
1540 */
1541static u32
1542iomap_finish_ioend(struct iomap_ioend *ioend, int error)
1543{
1544 struct inode *inode = ioend->io_inode;
1545 struct bio *bio = &ioend->io_bio;
1546 struct folio_iter fi;
1547 u32 folio_count = 0;
1548
1549 if (error) {
1550 mapping_set_error(inode->i_mapping, error);
1551 if (!bio_flagged(bio, BIO_QUIET)) {
1552 pr_err_ratelimited(
1553"%s: writeback error on inode %lu, offset %lld, sector %llu",
1554 inode->i_sb->s_id, inode->i_ino,
1555 ioend->io_offset, ioend->io_sector);
1556 }
1557 }
1558
1559 /* walk all folios in bio, ending page IO on them */
1560 bio_for_each_folio_all(fi, bio) {
1561 iomap_finish_folio_write(inode, fi.folio, fi.length);
1562 folio_count++;
1563 }
1564
1565 bio_put(bio); /* frees the ioend */
1566 return folio_count;
1567}
1568
1569/*
1570 * Ioend completion routine for merged bios. This can only be called from task
1571 * contexts as merged ioends can be of unbound length. Hence we have to break up
1572 * the writeback completions into manageable chunks to avoid long scheduler
1573 * holdoffs. We aim to keep scheduler holdoffs down below 10ms so that we get
1574 * good batch processing throughput without creating adverse scheduler latency
1575 * conditions.
1576 */
1577void
1578iomap_finish_ioends(struct iomap_ioend *ioend, int error)
1579{
1580 struct list_head tmp;
1581 u32 completions;
1582
1583 might_sleep();
1584
1585 list_replace_init(&ioend->io_list, &tmp);
1586 completions = iomap_finish_ioend(ioend, error);
1587
1588 while (!list_empty(&tmp)) {
1589 if (completions > IOEND_BATCH_SIZE * 8) {
1590 cond_resched();
1591 completions = 0;
1592 }
1593 ioend = list_first_entry(&tmp, struct iomap_ioend, io_list);
1594 list_del_init(&ioend->io_list);
1595 completions += iomap_finish_ioend(ioend, error);
1596 }
1597}
1598EXPORT_SYMBOL_GPL(iomap_finish_ioends);
1599
1600/*
1601 * We can merge two adjacent ioends if they have the same set of work to do.
1602 */
1603static bool
1604iomap_ioend_can_merge(struct iomap_ioend *ioend, struct iomap_ioend *next)
1605{
1606 if (ioend->io_bio.bi_status != next->io_bio.bi_status)
1607 return false;
1608 if (next->io_flags & IOMAP_F_BOUNDARY)
1609 return false;
1610 if ((ioend->io_flags & IOMAP_F_SHARED) ^
1611 (next->io_flags & IOMAP_F_SHARED))
1612 return false;
1613 if ((ioend->io_type == IOMAP_UNWRITTEN) ^
1614 (next->io_type == IOMAP_UNWRITTEN))
1615 return false;
1616 if (ioend->io_offset + ioend->io_size != next->io_offset)
1617 return false;
1618 /*
1619 * Do not merge physically discontiguous ioends. The filesystem
1620 * completion functions will have to iterate the physical
1621 * discontiguities even if we merge the ioends at a logical level, so
1622 * we don't gain anything by merging physical discontiguities here.
1623 *
1624 * We cannot use bio->bi_iter.bi_sector here as it is modified during
1625 * submission so does not point to the start sector of the bio at
1626 * completion.
1627 */
1628 if (ioend->io_sector + (ioend->io_size >> 9) != next->io_sector)
1629 return false;
1630 return true;
1631}
1632
1633void
1634iomap_ioend_try_merge(struct iomap_ioend *ioend, struct list_head *more_ioends)
1635{
1636 struct iomap_ioend *next;
1637
1638 INIT_LIST_HEAD(&ioend->io_list);
1639
1640 while ((next = list_first_entry_or_null(more_ioends, struct iomap_ioend,
1641 io_list))) {
1642 if (!iomap_ioend_can_merge(ioend, next))
1643 break;
1644 list_move_tail(&next->io_list, &ioend->io_list);
1645 ioend->io_size += next->io_size;
1646 }
1647}
1648EXPORT_SYMBOL_GPL(iomap_ioend_try_merge);
1649
1650static int
1651iomap_ioend_compare(void *priv, const struct list_head *a,
1652 const struct list_head *b)
1653{
1654 struct iomap_ioend *ia = container_of(a, struct iomap_ioend, io_list);
1655 struct iomap_ioend *ib = container_of(b, struct iomap_ioend, io_list);
1656
1657 if (ia->io_offset < ib->io_offset)
1658 return -1;
1659 if (ia->io_offset > ib->io_offset)
1660 return 1;
1661 return 0;
1662}
1663
1664void
1665iomap_sort_ioends(struct list_head *ioend_list)
1666{
1667 list_sort(NULL, ioend_list, iomap_ioend_compare);
1668}
1669EXPORT_SYMBOL_GPL(iomap_sort_ioends);
1670
1671static void iomap_writepage_end_bio(struct bio *bio)
1672{
1673 iomap_finish_ioend(iomap_ioend_from_bio(bio),
1674 blk_status_to_errno(bio->bi_status));
1675}
1676
1677/*
1678 * Submit the final bio for an ioend.
1679 *
1680 * If @error is non-zero, it means that we have a situation where some part of
1681 * the submission process has failed after we've marked pages for writeback.
1682 * We cannot cancel ioend directly in that case, so call the bio end I/O handler
1683 * with the error status here to run the normal I/O completion handler to clear
1684 * the writeback bit and let the file system proess the errors.
1685 */
1686static int iomap_submit_ioend(struct iomap_writepage_ctx *wpc, int error)
1687{
1688 if (!wpc->ioend)
1689 return error;
1690
1691 /*
1692 * Let the file systems prepare the I/O submission and hook in an I/O
1693 * comletion handler. This also needs to happen in case after a
1694 * failure happened so that the file system end I/O handler gets called
1695 * to clean up.
1696 */
1697 if (wpc->ops->prepare_ioend)
1698 error = wpc->ops->prepare_ioend(wpc->ioend, error);
1699
1700 if (error) {
1701 wpc->ioend->io_bio.bi_status = errno_to_blk_status(error);
1702 bio_endio(&wpc->ioend->io_bio);
1703 } else {
1704 submit_bio(&wpc->ioend->io_bio);
1705 }
1706
1707 wpc->ioend = NULL;
1708 return error;
1709}
1710
1711static struct iomap_ioend *iomap_alloc_ioend(struct iomap_writepage_ctx *wpc,
1712 struct writeback_control *wbc, struct inode *inode, loff_t pos)
1713{
1714 struct iomap_ioend *ioend;
1715 struct bio *bio;
1716
1717 bio = bio_alloc_bioset(wpc->iomap.bdev, BIO_MAX_VECS,
1718 REQ_OP_WRITE | wbc_to_write_flags(wbc),
1719 GFP_NOFS, &iomap_ioend_bioset);
1720 bio->bi_iter.bi_sector = iomap_sector(&wpc->iomap, pos);
1721 bio->bi_end_io = iomap_writepage_end_bio;
1722 wbc_init_bio(wbc, bio);
1723 bio->bi_write_hint = inode->i_write_hint;
1724
1725 ioend = iomap_ioend_from_bio(bio);
1726 INIT_LIST_HEAD(&ioend->io_list);
1727 ioend->io_type = wpc->iomap.type;
1728 ioend->io_flags = wpc->iomap.flags;
1729 if (pos > wpc->iomap.offset)
1730 wpc->iomap.flags &= ~IOMAP_F_BOUNDARY;
1731 ioend->io_inode = inode;
1732 ioend->io_size = 0;
1733 ioend->io_offset = pos;
1734 ioend->io_sector = bio->bi_iter.bi_sector;
1735
1736 wpc->nr_folios = 0;
1737 return ioend;
1738}
1739
1740static bool iomap_can_add_to_ioend(struct iomap_writepage_ctx *wpc, loff_t pos)
1741{
1742 if (wpc->iomap.offset == pos && (wpc->iomap.flags & IOMAP_F_BOUNDARY))
1743 return false;
1744 if ((wpc->iomap.flags & IOMAP_F_SHARED) !=
1745 (wpc->ioend->io_flags & IOMAP_F_SHARED))
1746 return false;
1747 if (wpc->iomap.type != wpc->ioend->io_type)
1748 return false;
1749 if (pos != wpc->ioend->io_offset + wpc->ioend->io_size)
1750 return false;
1751 if (iomap_sector(&wpc->iomap, pos) !=
1752 bio_end_sector(&wpc->ioend->io_bio))
1753 return false;
1754 /*
1755 * Limit ioend bio chain lengths to minimise IO completion latency. This
1756 * also prevents long tight loops ending page writeback on all the
1757 * folios in the ioend.
1758 */
1759 if (wpc->nr_folios >= IOEND_BATCH_SIZE)
1760 return false;
1761 return true;
1762}
1763
1764/*
1765 * Test to see if we have an existing ioend structure that we could append to
1766 * first; otherwise finish off the current ioend and start another.
1767 *
1768 * If a new ioend is created and cached, the old ioend is submitted to the block
1769 * layer instantly. Batching optimisations are provided by higher level block
1770 * plugging.
1771 *
1772 * At the end of a writeback pass, there will be a cached ioend remaining on the
1773 * writepage context that the caller will need to submit.
1774 */
1775static int iomap_add_to_ioend(struct iomap_writepage_ctx *wpc,
1776 struct writeback_control *wbc, struct folio *folio,
1777 struct inode *inode, loff_t pos, loff_t end_pos,
1778 unsigned len)
1779{
1780 struct iomap_folio_state *ifs = folio->private;
1781 size_t poff = offset_in_folio(folio, pos);
1782 int error;
1783
1784 if (!wpc->ioend || !iomap_can_add_to_ioend(wpc, pos)) {
1785new_ioend:
1786 error = iomap_submit_ioend(wpc, 0);
1787 if (error)
1788 return error;
1789 wpc->ioend = iomap_alloc_ioend(wpc, wbc, inode, pos);
1790 }
1791
1792 if (!bio_add_folio(&wpc->ioend->io_bio, folio, len, poff))
1793 goto new_ioend;
1794
1795 if (ifs)
1796 atomic_add(len, &ifs->write_bytes_pending);
1797
1798 /*
1799 * Clamp io_offset and io_size to the incore EOF so that ondisk
1800 * file size updates in the ioend completion are byte-accurate.
1801 * This avoids recovering files with zeroed tail regions when
1802 * writeback races with appending writes:
1803 *
1804 * Thread 1: Thread 2:
1805 * ------------ -----------
1806 * write [A, A+B]
1807 * update inode size to A+B
1808 * submit I/O [A, A+BS]
1809 * write [A+B, A+B+C]
1810 * update inode size to A+B+C
1811 * <I/O completes, updates disk size to min(A+B+C, A+BS)>
1812 * <power failure>
1813 *
1814 * After reboot:
1815 * 1) with A+B+C < A+BS, the file has zero padding in range
1816 * [A+B, A+B+C]
1817 *
1818 * |< Block Size (BS) >|
1819 * |DDDDDDDDDDDD0000000000000|
1820 * ^ ^ ^
1821 * A A+B A+B+C
1822 * (EOF)
1823 *
1824 * 2) with A+B+C > A+BS, the file has zero padding in range
1825 * [A+B, A+BS]
1826 *
1827 * |< Block Size (BS) >|< Block Size (BS) >|
1828 * |DDDDDDDDDDDD0000000000000|00000000000000000000000000|
1829 * ^ ^ ^ ^
1830 * A A+B A+BS A+B+C
1831 * (EOF)
1832 *
1833 * D = Valid Data
1834 * 0 = Zero Padding
1835 *
1836 * Note that this defeats the ability to chain the ioends of
1837 * appending writes.
1838 */
1839 wpc->ioend->io_size += len;
1840 if (wpc->ioend->io_offset + wpc->ioend->io_size > end_pos)
1841 wpc->ioend->io_size = end_pos - wpc->ioend->io_offset;
1842
1843 wbc_account_cgroup_owner(wbc, folio, len);
1844 return 0;
1845}
1846
1847static int iomap_writepage_map_blocks(struct iomap_writepage_ctx *wpc,
1848 struct writeback_control *wbc, struct folio *folio,
1849 struct inode *inode, u64 pos, u64 end_pos,
1850 unsigned dirty_len, unsigned *count)
1851{
1852 int error;
1853
1854 do {
1855 unsigned map_len;
1856
1857 error = wpc->ops->map_blocks(wpc, inode, pos, dirty_len);
1858 if (error)
1859 break;
1860 trace_iomap_writepage_map(inode, pos, dirty_len, &wpc->iomap);
1861
1862 map_len = min_t(u64, dirty_len,
1863 wpc->iomap.offset + wpc->iomap.length - pos);
1864 WARN_ON_ONCE(!folio->private && map_len < dirty_len);
1865
1866 switch (wpc->iomap.type) {
1867 case IOMAP_INLINE:
1868 WARN_ON_ONCE(1);
1869 error = -EIO;
1870 break;
1871 case IOMAP_HOLE:
1872 break;
1873 default:
1874 error = iomap_add_to_ioend(wpc, wbc, folio, inode, pos,
1875 end_pos, map_len);
1876 if (!error)
1877 (*count)++;
1878 break;
1879 }
1880 dirty_len -= map_len;
1881 pos += map_len;
1882 } while (dirty_len && !error);
1883
1884 /*
1885 * We cannot cancel the ioend directly here on error. We may have
1886 * already set other pages under writeback and hence we have to run I/O
1887 * completion to mark the error state of the pages under writeback
1888 * appropriately.
1889 *
1890 * Just let the file system know what portion of the folio failed to
1891 * map.
1892 */
1893 if (error && wpc->ops->discard_folio)
1894 wpc->ops->discard_folio(folio, pos);
1895 return error;
1896}
1897
1898/*
1899 * Check interaction of the folio with the file end.
1900 *
1901 * If the folio is entirely beyond i_size, return false. If it straddles
1902 * i_size, adjust end_pos and zero all data beyond i_size.
1903 */
1904static bool iomap_writepage_handle_eof(struct folio *folio, struct inode *inode,
1905 u64 *end_pos)
1906{
1907 u64 isize = i_size_read(inode);
1908
1909 if (*end_pos > isize) {
1910 size_t poff = offset_in_folio(folio, isize);
1911 pgoff_t end_index = isize >> PAGE_SHIFT;
1912
1913 /*
1914 * If the folio is entirely ouside of i_size, skip it.
1915 *
1916 * This can happen due to a truncate operation that is in
1917 * progress and in that case truncate will finish it off once
1918 * we've dropped the folio lock.
1919 *
1920 * Note that the pgoff_t used for end_index is an unsigned long.
1921 * If the given offset is greater than 16TB on a 32-bit system,
1922 * then if we checked if the folio is fully outside i_size with
1923 * "if (folio->index >= end_index + 1)", "end_index + 1" would
1924 * overflow and evaluate to 0. Hence this folio would be
1925 * redirtied and written out repeatedly, which would result in
1926 * an infinite loop; the user program performing this operation
1927 * would hang. Instead, we can detect this situation by
1928 * checking if the folio is totally beyond i_size or if its
1929 * offset is just equal to the EOF.
1930 */
1931 if (folio->index > end_index ||
1932 (folio->index == end_index && poff == 0))
1933 return false;
1934
1935 /*
1936 * The folio straddles i_size.
1937 *
1938 * It must be zeroed out on each and every writepage invocation
1939 * because it may be mmapped:
1940 *
1941 * A file is mapped in multiples of the page size. For a
1942 * file that is not a multiple of the page size, the
1943 * remaining memory is zeroed when mapped, and writes to that
1944 * region are not written out to the file.
1945 *
1946 * Also adjust the end_pos to the end of file and skip writeback
1947 * for all blocks entirely beyond i_size.
1948 */
1949 folio_zero_segment(folio, poff, folio_size(folio));
1950 *end_pos = isize;
1951 }
1952
1953 return true;
1954}
1955
1956static int iomap_writepage_map(struct iomap_writepage_ctx *wpc,
1957 struct writeback_control *wbc, struct folio *folio)
1958{
1959 struct iomap_folio_state *ifs = folio->private;
1960 struct inode *inode = folio->mapping->host;
1961 u64 pos = folio_pos(folio);
1962 u64 end_pos = pos + folio_size(folio);
1963 u64 end_aligned = 0;
1964 unsigned count = 0;
1965 int error = 0;
1966 u32 rlen;
1967
1968 WARN_ON_ONCE(!folio_test_locked(folio));
1969 WARN_ON_ONCE(folio_test_dirty(folio));
1970 WARN_ON_ONCE(folio_test_writeback(folio));
1971
1972 trace_iomap_writepage(inode, pos, folio_size(folio));
1973
1974 if (!iomap_writepage_handle_eof(folio, inode, &end_pos)) {
1975 folio_unlock(folio);
1976 return 0;
1977 }
1978 WARN_ON_ONCE(end_pos <= pos);
1979
1980 if (i_blocks_per_folio(inode, folio) > 1) {
1981 if (!ifs) {
1982 ifs = ifs_alloc(inode, folio, 0);
1983 iomap_set_range_dirty(folio, 0, end_pos - pos);
1984 }
1985
1986 /*
1987 * Keep the I/O completion handler from clearing the writeback
1988 * bit until we have submitted all blocks by adding a bias to
1989 * ifs->write_bytes_pending, which is dropped after submitting
1990 * all blocks.
1991 */
1992 WARN_ON_ONCE(atomic_read(&ifs->write_bytes_pending) != 0);
1993 atomic_inc(&ifs->write_bytes_pending);
1994 }
1995
1996 /*
1997 * Set the writeback bit ASAP, as the I/O completion for the single
1998 * block per folio case happen hit as soon as we're submitting the bio.
1999 */
2000 folio_start_writeback(folio);
2001
2002 /*
2003 * Walk through the folio to find dirty areas to write back.
2004 */
2005 end_aligned = round_up(end_pos, i_blocksize(inode));
2006 while ((rlen = iomap_find_dirty_range(folio, &pos, end_aligned))) {
2007 error = iomap_writepage_map_blocks(wpc, wbc, folio, inode,
2008 pos, end_pos, rlen, &count);
2009 if (error)
2010 break;
2011 pos += rlen;
2012 }
2013
2014 if (count)
2015 wpc->nr_folios++;
2016
2017 /*
2018 * We can have dirty bits set past end of file in page_mkwrite path
2019 * while mapping the last partial folio. Hence it's better to clear
2020 * all the dirty bits in the folio here.
2021 */
2022 iomap_clear_range_dirty(folio, 0, folio_size(folio));
2023
2024 /*
2025 * Usually the writeback bit is cleared by the I/O completion handler.
2026 * But we may end up either not actually writing any blocks, or (when
2027 * there are multiple blocks in a folio) all I/O might have finished
2028 * already at this point. In that case we need to clear the writeback
2029 * bit ourselves right after unlocking the page.
2030 */
2031 folio_unlock(folio);
2032 if (ifs) {
2033 if (atomic_dec_and_test(&ifs->write_bytes_pending))
2034 folio_end_writeback(folio);
2035 } else {
2036 if (!count)
2037 folio_end_writeback(folio);
2038 }
2039 mapping_set_error(inode->i_mapping, error);
2040 return error;
2041}
2042
2043int
2044iomap_writepages(struct address_space *mapping, struct writeback_control *wbc,
2045 struct iomap_writepage_ctx *wpc,
2046 const struct iomap_writeback_ops *ops)
2047{
2048 struct folio *folio = NULL;
2049 int error;
2050
2051 /*
2052 * Writeback from reclaim context should never happen except in the case
2053 * of a VM regression so warn about it and refuse to write the data.
2054 */
2055 if (WARN_ON_ONCE((current->flags & (PF_MEMALLOC | PF_KSWAPD)) ==
2056 PF_MEMALLOC))
2057 return -EIO;
2058
2059 wpc->ops = ops;
2060 while ((folio = writeback_iter(mapping, wbc, folio, &error)))
2061 error = iomap_writepage_map(wpc, wbc, folio);
2062 return iomap_submit_ioend(wpc, error);
2063}
2064EXPORT_SYMBOL_GPL(iomap_writepages);
2065
2066static int __init iomap_buffered_init(void)
2067{
2068 return bioset_init(&iomap_ioend_bioset, 4 * (PAGE_SIZE / SECTOR_SIZE),
2069 offsetof(struct iomap_ioend, io_bio),
2070 BIOSET_NEED_BVECS);
2071}
2072fs_initcall(iomap_buffered_init);