Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * linux/fs/ext4/page-io.c
4 *
5 * This contains the new page_io functions for ext4
6 *
7 * Written by Theodore Ts'o, 2010.
8 */
9
10#include <linux/fs.h>
11#include <linux/time.h>
12#include <linux/highuid.h>
13#include <linux/pagemap.h>
14#include <linux/quotaops.h>
15#include <linux/string.h>
16#include <linux/buffer_head.h>
17#include <linux/writeback.h>
18#include <linux/pagevec.h>
19#include <linux/mpage.h>
20#include <linux/namei.h>
21#include <linux/uio.h>
22#include <linux/bio.h>
23#include <linux/workqueue.h>
24#include <linux/kernel.h>
25#include <linux/slab.h>
26#include <linux/mm.h>
27#include <linux/sched/mm.h>
28
29#include "ext4_jbd2.h"
30#include "xattr.h"
31#include "acl.h"
32
33static struct kmem_cache *io_end_cachep;
34static struct kmem_cache *io_end_vec_cachep;
35
36int __init ext4_init_pageio(void)
37{
38 io_end_cachep = KMEM_CACHE(ext4_io_end, SLAB_RECLAIM_ACCOUNT);
39 if (io_end_cachep == NULL)
40 return -ENOMEM;
41
42 io_end_vec_cachep = KMEM_CACHE(ext4_io_end_vec, 0);
43 if (io_end_vec_cachep == NULL) {
44 kmem_cache_destroy(io_end_cachep);
45 return -ENOMEM;
46 }
47 return 0;
48}
49
50void ext4_exit_pageio(void)
51{
52 kmem_cache_destroy(io_end_cachep);
53 kmem_cache_destroy(io_end_vec_cachep);
54}
55
56struct ext4_io_end_vec *ext4_alloc_io_end_vec(ext4_io_end_t *io_end)
57{
58 struct ext4_io_end_vec *io_end_vec;
59
60 io_end_vec = kmem_cache_zalloc(io_end_vec_cachep, GFP_NOFS);
61 if (!io_end_vec)
62 return ERR_PTR(-ENOMEM);
63 INIT_LIST_HEAD(&io_end_vec->list);
64 list_add_tail(&io_end_vec->list, &io_end->list_vec);
65 return io_end_vec;
66}
67
68static void ext4_free_io_end_vec(ext4_io_end_t *io_end)
69{
70 struct ext4_io_end_vec *io_end_vec, *tmp;
71
72 if (list_empty(&io_end->list_vec))
73 return;
74 list_for_each_entry_safe(io_end_vec, tmp, &io_end->list_vec, list) {
75 list_del(&io_end_vec->list);
76 kmem_cache_free(io_end_vec_cachep, io_end_vec);
77 }
78}
79
80struct ext4_io_end_vec *ext4_last_io_end_vec(ext4_io_end_t *io_end)
81{
82 BUG_ON(list_empty(&io_end->list_vec));
83 return list_last_entry(&io_end->list_vec, struct ext4_io_end_vec, list);
84}
85
86/*
87 * Print an buffer I/O error compatible with the fs/buffer.c. This
88 * provides compatibility with dmesg scrapers that look for a specific
89 * buffer I/O error message. We really need a unified error reporting
90 * structure to userspace ala Digital Unix's uerf system, but it's
91 * probably not going to happen in my lifetime, due to LKML politics...
92 */
93static void buffer_io_error(struct buffer_head *bh)
94{
95 printk_ratelimited(KERN_ERR "Buffer I/O error on device %pg, logical block %llu\n",
96 bh->b_bdev,
97 (unsigned long long)bh->b_blocknr);
98}
99
100static void ext4_finish_bio(struct bio *bio)
101{
102 struct bio_vec *bvec;
103 struct bvec_iter_all iter_all;
104
105 bio_for_each_segment_all(bvec, bio, iter_all) {
106 struct page *page = bvec->bv_page;
107 struct page *bounce_page = NULL;
108 struct buffer_head *bh, *head;
109 unsigned bio_start = bvec->bv_offset;
110 unsigned bio_end = bio_start + bvec->bv_len;
111 unsigned under_io = 0;
112 unsigned long flags;
113
114 if (fscrypt_is_bounce_page(page)) {
115 bounce_page = page;
116 page = fscrypt_pagecache_page(bounce_page);
117 }
118
119 if (bio->bi_status) {
120 SetPageError(page);
121 mapping_set_error(page->mapping, -EIO);
122 }
123 bh = head = page_buffers(page);
124 /*
125 * We check all buffers in the page under b_uptodate_lock
126 * to avoid races with other end io clearing async_write flags
127 */
128 spin_lock_irqsave(&head->b_uptodate_lock, flags);
129 do {
130 if (bh_offset(bh) < bio_start ||
131 bh_offset(bh) + bh->b_size > bio_end) {
132 if (buffer_async_write(bh))
133 under_io++;
134 continue;
135 }
136 clear_buffer_async_write(bh);
137 if (bio->bi_status) {
138 set_buffer_write_io_error(bh);
139 buffer_io_error(bh);
140 }
141 } while ((bh = bh->b_this_page) != head);
142 spin_unlock_irqrestore(&head->b_uptodate_lock, flags);
143 if (!under_io) {
144 fscrypt_free_bounce_page(bounce_page);
145 end_page_writeback(page);
146 }
147 }
148}
149
150static void ext4_release_io_end(ext4_io_end_t *io_end)
151{
152 struct bio *bio, *next_bio;
153
154 BUG_ON(!list_empty(&io_end->list));
155 BUG_ON(io_end->flag & EXT4_IO_END_UNWRITTEN);
156 WARN_ON(io_end->handle);
157
158 for (bio = io_end->bio; bio; bio = next_bio) {
159 next_bio = bio->bi_private;
160 ext4_finish_bio(bio);
161 bio_put(bio);
162 }
163 ext4_free_io_end_vec(io_end);
164 kmem_cache_free(io_end_cachep, io_end);
165}
166
167/*
168 * Check a range of space and convert unwritten extents to written. Note that
169 * we are protected from truncate touching same part of extent tree by the
170 * fact that truncate code waits for all DIO to finish (thus exclusion from
171 * direct IO is achieved) and also waits for PageWriteback bits. Thus we
172 * cannot get to ext4_ext_truncate() before all IOs overlapping that range are
173 * completed (happens from ext4_free_ioend()).
174 */
175static int ext4_end_io_end(ext4_io_end_t *io_end)
176{
177 struct inode *inode = io_end->inode;
178 handle_t *handle = io_end->handle;
179 int ret = 0;
180
181 ext4_debug("ext4_end_io_nolock: io_end 0x%p from inode %lu,list->next 0x%p,"
182 "list->prev 0x%p\n",
183 io_end, inode->i_ino, io_end->list.next, io_end->list.prev);
184
185 io_end->handle = NULL; /* Following call will use up the handle */
186 ret = ext4_convert_unwritten_io_end_vec(handle, io_end);
187 if (ret < 0 && !ext4_forced_shutdown(EXT4_SB(inode->i_sb))) {
188 ext4_msg(inode->i_sb, KERN_EMERG,
189 "failed to convert unwritten extents to written "
190 "extents -- potential data loss! "
191 "(inode %lu, error %d)", inode->i_ino, ret);
192 }
193 ext4_clear_io_unwritten_flag(io_end);
194 ext4_release_io_end(io_end);
195 return ret;
196}
197
198static void dump_completed_IO(struct inode *inode, struct list_head *head)
199{
200#ifdef EXT4FS_DEBUG
201 struct list_head *cur, *before, *after;
202 ext4_io_end_t *io_end, *io_end0, *io_end1;
203
204 if (list_empty(head))
205 return;
206
207 ext4_debug("Dump inode %lu completed io list\n", inode->i_ino);
208 list_for_each_entry(io_end, head, list) {
209 cur = &io_end->list;
210 before = cur->prev;
211 io_end0 = container_of(before, ext4_io_end_t, list);
212 after = cur->next;
213 io_end1 = container_of(after, ext4_io_end_t, list);
214
215 ext4_debug("io 0x%p from inode %lu,prev 0x%p,next 0x%p\n",
216 io_end, inode->i_ino, io_end0, io_end1);
217 }
218#endif
219}
220
221/* Add the io_end to per-inode completed end_io list. */
222static void ext4_add_complete_io(ext4_io_end_t *io_end)
223{
224 struct ext4_inode_info *ei = EXT4_I(io_end->inode);
225 struct ext4_sb_info *sbi = EXT4_SB(io_end->inode->i_sb);
226 struct workqueue_struct *wq;
227 unsigned long flags;
228
229 /* Only reserved conversions from writeback should enter here */
230 WARN_ON(!(io_end->flag & EXT4_IO_END_UNWRITTEN));
231 WARN_ON(!io_end->handle && sbi->s_journal);
232 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
233 wq = sbi->rsv_conversion_wq;
234 if (list_empty(&ei->i_rsv_conversion_list))
235 queue_work(wq, &ei->i_rsv_conversion_work);
236 list_add_tail(&io_end->list, &ei->i_rsv_conversion_list);
237 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
238}
239
240static int ext4_do_flush_completed_IO(struct inode *inode,
241 struct list_head *head)
242{
243 ext4_io_end_t *io_end;
244 struct list_head unwritten;
245 unsigned long flags;
246 struct ext4_inode_info *ei = EXT4_I(inode);
247 int err, ret = 0;
248
249 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
250 dump_completed_IO(inode, head);
251 list_replace_init(head, &unwritten);
252 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
253
254 while (!list_empty(&unwritten)) {
255 io_end = list_entry(unwritten.next, ext4_io_end_t, list);
256 BUG_ON(!(io_end->flag & EXT4_IO_END_UNWRITTEN));
257 list_del_init(&io_end->list);
258
259 err = ext4_end_io_end(io_end);
260 if (unlikely(!ret && err))
261 ret = err;
262 }
263 return ret;
264}
265
266/*
267 * work on completed IO, to convert unwritten extents to extents
268 */
269void ext4_end_io_rsv_work(struct work_struct *work)
270{
271 struct ext4_inode_info *ei = container_of(work, struct ext4_inode_info,
272 i_rsv_conversion_work);
273 ext4_do_flush_completed_IO(&ei->vfs_inode, &ei->i_rsv_conversion_list);
274}
275
276ext4_io_end_t *ext4_init_io_end(struct inode *inode, gfp_t flags)
277{
278 ext4_io_end_t *io_end = kmem_cache_zalloc(io_end_cachep, flags);
279
280 if (io_end) {
281 io_end->inode = inode;
282 INIT_LIST_HEAD(&io_end->list);
283 INIT_LIST_HEAD(&io_end->list_vec);
284 refcount_set(&io_end->count, 1);
285 }
286 return io_end;
287}
288
289void ext4_put_io_end_defer(ext4_io_end_t *io_end)
290{
291 if (refcount_dec_and_test(&io_end->count)) {
292 if (!(io_end->flag & EXT4_IO_END_UNWRITTEN) ||
293 list_empty(&io_end->list_vec)) {
294 ext4_release_io_end(io_end);
295 return;
296 }
297 ext4_add_complete_io(io_end);
298 }
299}
300
301int ext4_put_io_end(ext4_io_end_t *io_end)
302{
303 int err = 0;
304
305 if (refcount_dec_and_test(&io_end->count)) {
306 if (io_end->flag & EXT4_IO_END_UNWRITTEN) {
307 err = ext4_convert_unwritten_io_end_vec(io_end->handle,
308 io_end);
309 io_end->handle = NULL;
310 ext4_clear_io_unwritten_flag(io_end);
311 }
312 ext4_release_io_end(io_end);
313 }
314 return err;
315}
316
317ext4_io_end_t *ext4_get_io_end(ext4_io_end_t *io_end)
318{
319 refcount_inc(&io_end->count);
320 return io_end;
321}
322
323/* BIO completion function for page writeback */
324static void ext4_end_bio(struct bio *bio)
325{
326 ext4_io_end_t *io_end = bio->bi_private;
327 sector_t bi_sector = bio->bi_iter.bi_sector;
328
329 if (WARN_ONCE(!io_end, "io_end is NULL: %pg: sector %Lu len %u err %d\n",
330 bio->bi_bdev,
331 (long long) bio->bi_iter.bi_sector,
332 (unsigned) bio_sectors(bio),
333 bio->bi_status)) {
334 ext4_finish_bio(bio);
335 bio_put(bio);
336 return;
337 }
338 bio->bi_end_io = NULL;
339
340 if (bio->bi_status) {
341 struct inode *inode = io_end->inode;
342
343 ext4_warning(inode->i_sb, "I/O error %d writing to inode %lu "
344 "starting block %llu)",
345 bio->bi_status, inode->i_ino,
346 (unsigned long long)
347 bi_sector >> (inode->i_blkbits - 9));
348 mapping_set_error(inode->i_mapping,
349 blk_status_to_errno(bio->bi_status));
350 }
351
352 if (io_end->flag & EXT4_IO_END_UNWRITTEN) {
353 /*
354 * Link bio into list hanging from io_end. We have to do it
355 * atomically as bio completions can be racing against each
356 * other.
357 */
358 bio->bi_private = xchg(&io_end->bio, bio);
359 ext4_put_io_end_defer(io_end);
360 } else {
361 /*
362 * Drop io_end reference early. Inode can get freed once
363 * we finish the bio.
364 */
365 ext4_put_io_end_defer(io_end);
366 ext4_finish_bio(bio);
367 bio_put(bio);
368 }
369}
370
371void ext4_io_submit(struct ext4_io_submit *io)
372{
373 struct bio *bio = io->io_bio;
374
375 if (bio) {
376 if (io->io_wbc->sync_mode == WB_SYNC_ALL)
377 io->io_bio->bi_opf |= REQ_SYNC;
378 submit_bio(io->io_bio);
379 }
380 io->io_bio = NULL;
381}
382
383void ext4_io_submit_init(struct ext4_io_submit *io,
384 struct writeback_control *wbc)
385{
386 io->io_wbc = wbc;
387 io->io_bio = NULL;
388 io->io_end = NULL;
389}
390
391static void io_submit_init_bio(struct ext4_io_submit *io,
392 struct buffer_head *bh)
393{
394 struct bio *bio;
395
396 /*
397 * bio_alloc will _always_ be able to allocate a bio if
398 * __GFP_DIRECT_RECLAIM is set, see comments for bio_alloc_bioset().
399 */
400 bio = bio_alloc(bh->b_bdev, BIO_MAX_VECS, REQ_OP_WRITE, GFP_NOIO);
401 fscrypt_set_bio_crypt_ctx_bh(bio, bh, GFP_NOIO);
402 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
403 bio->bi_end_io = ext4_end_bio;
404 bio->bi_private = ext4_get_io_end(io->io_end);
405 io->io_bio = bio;
406 io->io_next_block = bh->b_blocknr;
407 wbc_init_bio(io->io_wbc, bio);
408}
409
410static void io_submit_add_bh(struct ext4_io_submit *io,
411 struct inode *inode,
412 struct page *page,
413 struct buffer_head *bh)
414{
415 int ret;
416
417 if (io->io_bio && (bh->b_blocknr != io->io_next_block ||
418 !fscrypt_mergeable_bio_bh(io->io_bio, bh))) {
419submit_and_retry:
420 ext4_io_submit(io);
421 }
422 if (io->io_bio == NULL)
423 io_submit_init_bio(io, bh);
424 ret = bio_add_page(io->io_bio, page, bh->b_size, bh_offset(bh));
425 if (ret != bh->b_size)
426 goto submit_and_retry;
427 wbc_account_cgroup_owner(io->io_wbc, page, bh->b_size);
428 io->io_next_block++;
429}
430
431int ext4_bio_write_page(struct ext4_io_submit *io,
432 struct page *page,
433 int len)
434{
435 struct page *bounce_page = NULL;
436 struct inode *inode = page->mapping->host;
437 unsigned block_start;
438 struct buffer_head *bh, *head;
439 int ret = 0;
440 int nr_to_submit = 0;
441 struct writeback_control *wbc = io->io_wbc;
442 bool keep_towrite = false;
443
444 BUG_ON(!PageLocked(page));
445 BUG_ON(PageWriteback(page));
446
447 ClearPageError(page);
448
449 /*
450 * Comments copied from block_write_full_page:
451 *
452 * The page straddles i_size. It must be zeroed out on each and every
453 * writepage invocation because it may be mmapped. "A file is mapped
454 * in multiples of the page size. For a file that is not a multiple of
455 * the page size, the remaining memory is zeroed when mapped, and
456 * writes to that region are not written out to the file."
457 */
458 if (len < PAGE_SIZE)
459 zero_user_segment(page, len, PAGE_SIZE);
460 /*
461 * In the first loop we prepare and mark buffers to submit. We have to
462 * mark all buffers in the page before submitting so that
463 * end_page_writeback() cannot be called from ext4_end_bio() when IO
464 * on the first buffer finishes and we are still working on submitting
465 * the second buffer.
466 */
467 bh = head = page_buffers(page);
468 do {
469 block_start = bh_offset(bh);
470 if (block_start >= len) {
471 clear_buffer_dirty(bh);
472 set_buffer_uptodate(bh);
473 continue;
474 }
475 if (!buffer_dirty(bh) || buffer_delay(bh) ||
476 !buffer_mapped(bh) || buffer_unwritten(bh)) {
477 /* A hole? We can safely clear the dirty bit */
478 if (!buffer_mapped(bh))
479 clear_buffer_dirty(bh);
480 /*
481 * Keeping dirty some buffer we cannot write? Make sure
482 * to redirty the page and keep TOWRITE tag so that
483 * racing WB_SYNC_ALL writeback does not skip the page.
484 * This happens e.g. when doing writeout for
485 * transaction commit.
486 */
487 if (buffer_dirty(bh)) {
488 if (!PageDirty(page))
489 redirty_page_for_writepage(wbc, page);
490 keep_towrite = true;
491 }
492 continue;
493 }
494 if (buffer_new(bh))
495 clear_buffer_new(bh);
496 set_buffer_async_write(bh);
497 clear_buffer_dirty(bh);
498 nr_to_submit++;
499 } while ((bh = bh->b_this_page) != head);
500
501 /* Nothing to submit? Just unlock the page... */
502 if (!nr_to_submit)
503 goto unlock;
504
505 bh = head = page_buffers(page);
506
507 /*
508 * If any blocks are being written to an encrypted file, encrypt them
509 * into a bounce page. For simplicity, just encrypt until the last
510 * block which might be needed. This may cause some unneeded blocks
511 * (e.g. holes) to be unnecessarily encrypted, but this is rare and
512 * can't happen in the common case of blocksize == PAGE_SIZE.
513 */
514 if (fscrypt_inode_uses_fs_layer_crypto(inode) && nr_to_submit) {
515 gfp_t gfp_flags = GFP_NOFS;
516 unsigned int enc_bytes = round_up(len, i_blocksize(inode));
517
518 /*
519 * Since bounce page allocation uses a mempool, we can only use
520 * a waiting mask (i.e. request guaranteed allocation) on the
521 * first page of the bio. Otherwise it can deadlock.
522 */
523 if (io->io_bio)
524 gfp_flags = GFP_NOWAIT | __GFP_NOWARN;
525 retry_encrypt:
526 bounce_page = fscrypt_encrypt_pagecache_blocks(page, enc_bytes,
527 0, gfp_flags);
528 if (IS_ERR(bounce_page)) {
529 ret = PTR_ERR(bounce_page);
530 if (ret == -ENOMEM &&
531 (io->io_bio || wbc->sync_mode == WB_SYNC_ALL)) {
532 gfp_t new_gfp_flags = GFP_NOFS;
533 if (io->io_bio)
534 ext4_io_submit(io);
535 else
536 new_gfp_flags |= __GFP_NOFAIL;
537 memalloc_retry_wait(gfp_flags);
538 gfp_flags = new_gfp_flags;
539 goto retry_encrypt;
540 }
541
542 printk_ratelimited(KERN_ERR "%s: ret = %d\n", __func__, ret);
543 redirty_page_for_writepage(wbc, page);
544 do {
545 if (buffer_async_write(bh)) {
546 clear_buffer_async_write(bh);
547 set_buffer_dirty(bh);
548 }
549 bh = bh->b_this_page;
550 } while (bh != head);
551 goto unlock;
552 }
553 }
554
555 if (keep_towrite)
556 set_page_writeback_keepwrite(page);
557 else
558 set_page_writeback(page);
559
560 /* Now submit buffers to write */
561 do {
562 if (!buffer_async_write(bh))
563 continue;
564 io_submit_add_bh(io, inode,
565 bounce_page ? bounce_page : page, bh);
566 } while ((bh = bh->b_this_page) != head);
567unlock:
568 unlock_page(page);
569 return ret;
570}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * linux/fs/ext4/page-io.c
4 *
5 * This contains the new page_io functions for ext4
6 *
7 * Written by Theodore Ts'o, 2010.
8 */
9
10#include <linux/fs.h>
11#include <linux/time.h>
12#include <linux/highuid.h>
13#include <linux/pagemap.h>
14#include <linux/quotaops.h>
15#include <linux/string.h>
16#include <linux/buffer_head.h>
17#include <linux/writeback.h>
18#include <linux/pagevec.h>
19#include <linux/mpage.h>
20#include <linux/namei.h>
21#include <linux/uio.h>
22#include <linux/bio.h>
23#include <linux/workqueue.h>
24#include <linux/kernel.h>
25#include <linux/slab.h>
26#include <linux/mm.h>
27#include <linux/backing-dev.h>
28
29#include "ext4_jbd2.h"
30#include "xattr.h"
31#include "acl.h"
32
33static struct kmem_cache *io_end_cachep;
34
35int __init ext4_init_pageio(void)
36{
37 io_end_cachep = KMEM_CACHE(ext4_io_end, SLAB_RECLAIM_ACCOUNT);
38 if (io_end_cachep == NULL)
39 return -ENOMEM;
40 return 0;
41}
42
43void ext4_exit_pageio(void)
44{
45 kmem_cache_destroy(io_end_cachep);
46}
47
48/*
49 * Print an buffer I/O error compatible with the fs/buffer.c. This
50 * provides compatibility with dmesg scrapers that look for a specific
51 * buffer I/O error message. We really need a unified error reporting
52 * structure to userspace ala Digital Unix's uerf system, but it's
53 * probably not going to happen in my lifetime, due to LKML politics...
54 */
55static void buffer_io_error(struct buffer_head *bh)
56{
57 printk_ratelimited(KERN_ERR "Buffer I/O error on device %pg, logical block %llu\n",
58 bh->b_bdev,
59 (unsigned long long)bh->b_blocknr);
60}
61
62static void ext4_finish_bio(struct bio *bio)
63{
64 struct bio_vec *bvec;
65 struct bvec_iter_all iter_all;
66
67 bio_for_each_segment_all(bvec, bio, iter_all) {
68 struct page *page = bvec->bv_page;
69 struct page *bounce_page = NULL;
70 struct buffer_head *bh, *head;
71 unsigned bio_start = bvec->bv_offset;
72 unsigned bio_end = bio_start + bvec->bv_len;
73 unsigned under_io = 0;
74 unsigned long flags;
75
76 if (!page)
77 continue;
78
79 if (fscrypt_is_bounce_page(page)) {
80 bounce_page = page;
81 page = fscrypt_pagecache_page(bounce_page);
82 }
83
84 if (bio->bi_status) {
85 SetPageError(page);
86 mapping_set_error(page->mapping, -EIO);
87 }
88 bh = head = page_buffers(page);
89 /*
90 * We check all buffers in the page under BH_Uptodate_Lock
91 * to avoid races with other end io clearing async_write flags
92 */
93 local_irq_save(flags);
94 bit_spin_lock(BH_Uptodate_Lock, &head->b_state);
95 do {
96 if (bh_offset(bh) < bio_start ||
97 bh_offset(bh) + bh->b_size > bio_end) {
98 if (buffer_async_write(bh))
99 under_io++;
100 continue;
101 }
102 clear_buffer_async_write(bh);
103 if (bio->bi_status)
104 buffer_io_error(bh);
105 } while ((bh = bh->b_this_page) != head);
106 bit_spin_unlock(BH_Uptodate_Lock, &head->b_state);
107 local_irq_restore(flags);
108 if (!under_io) {
109 fscrypt_free_bounce_page(bounce_page);
110 end_page_writeback(page);
111 }
112 }
113}
114
115static void ext4_release_io_end(ext4_io_end_t *io_end)
116{
117 struct bio *bio, *next_bio;
118
119 BUG_ON(!list_empty(&io_end->list));
120 BUG_ON(io_end->flag & EXT4_IO_END_UNWRITTEN);
121 WARN_ON(io_end->handle);
122
123 for (bio = io_end->bio; bio; bio = next_bio) {
124 next_bio = bio->bi_private;
125 ext4_finish_bio(bio);
126 bio_put(bio);
127 }
128 kmem_cache_free(io_end_cachep, io_end);
129}
130
131/*
132 * Check a range of space and convert unwritten extents to written. Note that
133 * we are protected from truncate touching same part of extent tree by the
134 * fact that truncate code waits for all DIO to finish (thus exclusion from
135 * direct IO is achieved) and also waits for PageWriteback bits. Thus we
136 * cannot get to ext4_ext_truncate() before all IOs overlapping that range are
137 * completed (happens from ext4_free_ioend()).
138 */
139static int ext4_end_io(ext4_io_end_t *io)
140{
141 struct inode *inode = io->inode;
142 loff_t offset = io->offset;
143 ssize_t size = io->size;
144 handle_t *handle = io->handle;
145 int ret = 0;
146
147 ext4_debug("ext4_end_io_nolock: io 0x%p from inode %lu,list->next 0x%p,"
148 "list->prev 0x%p\n",
149 io, inode->i_ino, io->list.next, io->list.prev);
150
151 io->handle = NULL; /* Following call will use up the handle */
152 ret = ext4_convert_unwritten_extents(handle, inode, offset, size);
153 if (ret < 0 && !ext4_forced_shutdown(EXT4_SB(inode->i_sb))) {
154 ext4_msg(inode->i_sb, KERN_EMERG,
155 "failed to convert unwritten extents to written "
156 "extents -- potential data loss! "
157 "(inode %lu, offset %llu, size %zd, error %d)",
158 inode->i_ino, offset, size, ret);
159 }
160 ext4_clear_io_unwritten_flag(io);
161 ext4_release_io_end(io);
162 return ret;
163}
164
165static void dump_completed_IO(struct inode *inode, struct list_head *head)
166{
167#ifdef EXT4FS_DEBUG
168 struct list_head *cur, *before, *after;
169 ext4_io_end_t *io, *io0, *io1;
170
171 if (list_empty(head))
172 return;
173
174 ext4_debug("Dump inode %lu completed io list\n", inode->i_ino);
175 list_for_each_entry(io, head, list) {
176 cur = &io->list;
177 before = cur->prev;
178 io0 = container_of(before, ext4_io_end_t, list);
179 after = cur->next;
180 io1 = container_of(after, ext4_io_end_t, list);
181
182 ext4_debug("io 0x%p from inode %lu,prev 0x%p,next 0x%p\n",
183 io, inode->i_ino, io0, io1);
184 }
185#endif
186}
187
188/* Add the io_end to per-inode completed end_io list. */
189static void ext4_add_complete_io(ext4_io_end_t *io_end)
190{
191 struct ext4_inode_info *ei = EXT4_I(io_end->inode);
192 struct ext4_sb_info *sbi = EXT4_SB(io_end->inode->i_sb);
193 struct workqueue_struct *wq;
194 unsigned long flags;
195
196 /* Only reserved conversions from writeback should enter here */
197 WARN_ON(!(io_end->flag & EXT4_IO_END_UNWRITTEN));
198 WARN_ON(!io_end->handle && sbi->s_journal);
199 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
200 wq = sbi->rsv_conversion_wq;
201 if (list_empty(&ei->i_rsv_conversion_list))
202 queue_work(wq, &ei->i_rsv_conversion_work);
203 list_add_tail(&io_end->list, &ei->i_rsv_conversion_list);
204 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
205}
206
207static int ext4_do_flush_completed_IO(struct inode *inode,
208 struct list_head *head)
209{
210 ext4_io_end_t *io;
211 struct list_head unwritten;
212 unsigned long flags;
213 struct ext4_inode_info *ei = EXT4_I(inode);
214 int err, ret = 0;
215
216 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
217 dump_completed_IO(inode, head);
218 list_replace_init(head, &unwritten);
219 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
220
221 while (!list_empty(&unwritten)) {
222 io = list_entry(unwritten.next, ext4_io_end_t, list);
223 BUG_ON(!(io->flag & EXT4_IO_END_UNWRITTEN));
224 list_del_init(&io->list);
225
226 err = ext4_end_io(io);
227 if (unlikely(!ret && err))
228 ret = err;
229 }
230 return ret;
231}
232
233/*
234 * work on completed IO, to convert unwritten extents to extents
235 */
236void ext4_end_io_rsv_work(struct work_struct *work)
237{
238 struct ext4_inode_info *ei = container_of(work, struct ext4_inode_info,
239 i_rsv_conversion_work);
240 ext4_do_flush_completed_IO(&ei->vfs_inode, &ei->i_rsv_conversion_list);
241}
242
243ext4_io_end_t *ext4_init_io_end(struct inode *inode, gfp_t flags)
244{
245 ext4_io_end_t *io = kmem_cache_zalloc(io_end_cachep, flags);
246 if (io) {
247 io->inode = inode;
248 INIT_LIST_HEAD(&io->list);
249 atomic_set(&io->count, 1);
250 }
251 return io;
252}
253
254void ext4_put_io_end_defer(ext4_io_end_t *io_end)
255{
256 if (atomic_dec_and_test(&io_end->count)) {
257 if (!(io_end->flag & EXT4_IO_END_UNWRITTEN) || !io_end->size) {
258 ext4_release_io_end(io_end);
259 return;
260 }
261 ext4_add_complete_io(io_end);
262 }
263}
264
265int ext4_put_io_end(ext4_io_end_t *io_end)
266{
267 int err = 0;
268
269 if (atomic_dec_and_test(&io_end->count)) {
270 if (io_end->flag & EXT4_IO_END_UNWRITTEN) {
271 err = ext4_convert_unwritten_extents(io_end->handle,
272 io_end->inode, io_end->offset,
273 io_end->size);
274 io_end->handle = NULL;
275 ext4_clear_io_unwritten_flag(io_end);
276 }
277 ext4_release_io_end(io_end);
278 }
279 return err;
280}
281
282ext4_io_end_t *ext4_get_io_end(ext4_io_end_t *io_end)
283{
284 atomic_inc(&io_end->count);
285 return io_end;
286}
287
288/* BIO completion function for page writeback */
289static void ext4_end_bio(struct bio *bio)
290{
291 ext4_io_end_t *io_end = bio->bi_private;
292 sector_t bi_sector = bio->bi_iter.bi_sector;
293 char b[BDEVNAME_SIZE];
294
295 if (WARN_ONCE(!io_end, "io_end is NULL: %s: sector %Lu len %u err %d\n",
296 bio_devname(bio, b),
297 (long long) bio->bi_iter.bi_sector,
298 (unsigned) bio_sectors(bio),
299 bio->bi_status)) {
300 ext4_finish_bio(bio);
301 bio_put(bio);
302 return;
303 }
304 bio->bi_end_io = NULL;
305
306 if (bio->bi_status) {
307 struct inode *inode = io_end->inode;
308
309 ext4_warning(inode->i_sb, "I/O error %d writing to inode %lu "
310 "(offset %llu size %ld starting block %llu)",
311 bio->bi_status, inode->i_ino,
312 (unsigned long long) io_end->offset,
313 (long) io_end->size,
314 (unsigned long long)
315 bi_sector >> (inode->i_blkbits - 9));
316 mapping_set_error(inode->i_mapping,
317 blk_status_to_errno(bio->bi_status));
318 }
319
320 if (io_end->flag & EXT4_IO_END_UNWRITTEN) {
321 /*
322 * Link bio into list hanging from io_end. We have to do it
323 * atomically as bio completions can be racing against each
324 * other.
325 */
326 bio->bi_private = xchg(&io_end->bio, bio);
327 ext4_put_io_end_defer(io_end);
328 } else {
329 /*
330 * Drop io_end reference early. Inode can get freed once
331 * we finish the bio.
332 */
333 ext4_put_io_end_defer(io_end);
334 ext4_finish_bio(bio);
335 bio_put(bio);
336 }
337}
338
339void ext4_io_submit(struct ext4_io_submit *io)
340{
341 struct bio *bio = io->io_bio;
342
343 if (bio) {
344 int io_op_flags = io->io_wbc->sync_mode == WB_SYNC_ALL ?
345 REQ_SYNC : 0;
346 io->io_bio->bi_write_hint = io->io_end->inode->i_write_hint;
347 bio_set_op_attrs(io->io_bio, REQ_OP_WRITE, io_op_flags);
348 submit_bio(io->io_bio);
349 }
350 io->io_bio = NULL;
351}
352
353void ext4_io_submit_init(struct ext4_io_submit *io,
354 struct writeback_control *wbc)
355{
356 io->io_wbc = wbc;
357 io->io_bio = NULL;
358 io->io_end = NULL;
359}
360
361static int io_submit_init_bio(struct ext4_io_submit *io,
362 struct buffer_head *bh)
363{
364 struct bio *bio;
365
366 bio = bio_alloc(GFP_NOIO, BIO_MAX_PAGES);
367 if (!bio)
368 return -ENOMEM;
369 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
370 bio_set_dev(bio, bh->b_bdev);
371 bio->bi_end_io = ext4_end_bio;
372 bio->bi_private = ext4_get_io_end(io->io_end);
373 io->io_bio = bio;
374 io->io_next_block = bh->b_blocknr;
375 wbc_init_bio(io->io_wbc, bio);
376 return 0;
377}
378
379static int io_submit_add_bh(struct ext4_io_submit *io,
380 struct inode *inode,
381 struct page *page,
382 struct buffer_head *bh)
383{
384 int ret;
385
386 if (io->io_bio && bh->b_blocknr != io->io_next_block) {
387submit_and_retry:
388 ext4_io_submit(io);
389 }
390 if (io->io_bio == NULL) {
391 ret = io_submit_init_bio(io, bh);
392 if (ret)
393 return ret;
394 io->io_bio->bi_write_hint = inode->i_write_hint;
395 }
396 ret = bio_add_page(io->io_bio, page, bh->b_size, bh_offset(bh));
397 if (ret != bh->b_size)
398 goto submit_and_retry;
399 wbc_account_cgroup_owner(io->io_wbc, page, bh->b_size);
400 io->io_next_block++;
401 return 0;
402}
403
404int ext4_bio_write_page(struct ext4_io_submit *io,
405 struct page *page,
406 int len,
407 struct writeback_control *wbc,
408 bool keep_towrite)
409{
410 struct page *bounce_page = NULL;
411 struct inode *inode = page->mapping->host;
412 unsigned block_start;
413 struct buffer_head *bh, *head;
414 int ret = 0;
415 int nr_submitted = 0;
416 int nr_to_submit = 0;
417
418 BUG_ON(!PageLocked(page));
419 BUG_ON(PageWriteback(page));
420
421 if (keep_towrite)
422 set_page_writeback_keepwrite(page);
423 else
424 set_page_writeback(page);
425 ClearPageError(page);
426
427 /*
428 * Comments copied from block_write_full_page:
429 *
430 * The page straddles i_size. It must be zeroed out on each and every
431 * writepage invocation because it may be mmapped. "A file is mapped
432 * in multiples of the page size. For a file that is not a multiple of
433 * the page size, the remaining memory is zeroed when mapped, and
434 * writes to that region are not written out to the file."
435 */
436 if (len < PAGE_SIZE)
437 zero_user_segment(page, len, PAGE_SIZE);
438 /*
439 * In the first loop we prepare and mark buffers to submit. We have to
440 * mark all buffers in the page before submitting so that
441 * end_page_writeback() cannot be called from ext4_bio_end_io() when IO
442 * on the first buffer finishes and we are still working on submitting
443 * the second buffer.
444 */
445 bh = head = page_buffers(page);
446 do {
447 block_start = bh_offset(bh);
448 if (block_start >= len) {
449 clear_buffer_dirty(bh);
450 set_buffer_uptodate(bh);
451 continue;
452 }
453 if (!buffer_dirty(bh) || buffer_delay(bh) ||
454 !buffer_mapped(bh) || buffer_unwritten(bh)) {
455 /* A hole? We can safely clear the dirty bit */
456 if (!buffer_mapped(bh))
457 clear_buffer_dirty(bh);
458 if (io->io_bio)
459 ext4_io_submit(io);
460 continue;
461 }
462 if (buffer_new(bh))
463 clear_buffer_new(bh);
464 set_buffer_async_write(bh);
465 nr_to_submit++;
466 } while ((bh = bh->b_this_page) != head);
467
468 bh = head = page_buffers(page);
469
470 /*
471 * If any blocks are being written to an encrypted file, encrypt them
472 * into a bounce page. For simplicity, just encrypt until the last
473 * block which might be needed. This may cause some unneeded blocks
474 * (e.g. holes) to be unnecessarily encrypted, but this is rare and
475 * can't happen in the common case of blocksize == PAGE_SIZE.
476 */
477 if (IS_ENCRYPTED(inode) && S_ISREG(inode->i_mode) && nr_to_submit) {
478 gfp_t gfp_flags = GFP_NOFS;
479 unsigned int enc_bytes = round_up(len, i_blocksize(inode));
480
481 retry_encrypt:
482 bounce_page = fscrypt_encrypt_pagecache_blocks(page, enc_bytes,
483 0, gfp_flags);
484 if (IS_ERR(bounce_page)) {
485 ret = PTR_ERR(bounce_page);
486 if (ret == -ENOMEM && wbc->sync_mode == WB_SYNC_ALL) {
487 if (io->io_bio) {
488 ext4_io_submit(io);
489 congestion_wait(BLK_RW_ASYNC, HZ/50);
490 }
491 gfp_flags |= __GFP_NOFAIL;
492 goto retry_encrypt;
493 }
494 bounce_page = NULL;
495 goto out;
496 }
497 }
498
499 /* Now submit buffers to write */
500 do {
501 if (!buffer_async_write(bh))
502 continue;
503 ret = io_submit_add_bh(io, inode, bounce_page ?: page, bh);
504 if (ret) {
505 /*
506 * We only get here on ENOMEM. Not much else
507 * we can do but mark the page as dirty, and
508 * better luck next time.
509 */
510 break;
511 }
512 nr_submitted++;
513 clear_buffer_dirty(bh);
514 } while ((bh = bh->b_this_page) != head);
515
516 /* Error stopped previous loop? Clean up buffers... */
517 if (ret) {
518 out:
519 fscrypt_free_bounce_page(bounce_page);
520 printk_ratelimited(KERN_ERR "%s: ret = %d\n", __func__, ret);
521 redirty_page_for_writepage(wbc, page);
522 do {
523 clear_buffer_async_write(bh);
524 bh = bh->b_this_page;
525 } while (bh != head);
526 }
527 unlock_page(page);
528 /* Nothing submitted - we have to end page writeback */
529 if (!nr_submitted)
530 end_page_writeback(page);
531 return ret;
532}