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/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;
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 buffer_io_error(bh);
139 } while ((bh = bh->b_this_page) != head);
140 spin_unlock_irqrestore(&head->b_uptodate_lock, flags);
141 if (!under_io) {
142 fscrypt_free_bounce_page(bounce_page);
143 end_page_writeback(page);
144 }
145 }
146}
147
148static void ext4_release_io_end(ext4_io_end_t *io_end)
149{
150 struct bio *bio, *next_bio;
151
152 BUG_ON(!list_empty(&io_end->list));
153 BUG_ON(io_end->flag & EXT4_IO_END_UNWRITTEN);
154 WARN_ON(io_end->handle);
155
156 for (bio = io_end->bio; bio; bio = next_bio) {
157 next_bio = bio->bi_private;
158 ext4_finish_bio(bio);
159 bio_put(bio);
160 }
161 ext4_free_io_end_vec(io_end);
162 kmem_cache_free(io_end_cachep, io_end);
163}
164
165/*
166 * Check a range of space and convert unwritten extents to written. Note that
167 * we are protected from truncate touching same part of extent tree by the
168 * fact that truncate code waits for all DIO to finish (thus exclusion from
169 * direct IO is achieved) and also waits for PageWriteback bits. Thus we
170 * cannot get to ext4_ext_truncate() before all IOs overlapping that range are
171 * completed (happens from ext4_free_ioend()).
172 */
173static int ext4_end_io_end(ext4_io_end_t *io_end)
174{
175 struct inode *inode = io_end->inode;
176 handle_t *handle = io_end->handle;
177 int ret = 0;
178
179 ext4_debug("ext4_end_io_nolock: io_end 0x%p from inode %lu,list->next 0x%p,"
180 "list->prev 0x%p\n",
181 io_end, inode->i_ino, io_end->list.next, io_end->list.prev);
182
183 io_end->handle = NULL; /* Following call will use up the handle */
184 ret = ext4_convert_unwritten_io_end_vec(handle, io_end);
185 if (ret < 0 && !ext4_forced_shutdown(EXT4_SB(inode->i_sb))) {
186 ext4_msg(inode->i_sb, KERN_EMERG,
187 "failed to convert unwritten extents to written "
188 "extents -- potential data loss! "
189 "(inode %lu, error %d)", inode->i_ino, ret);
190 }
191 ext4_clear_io_unwritten_flag(io_end);
192 ext4_release_io_end(io_end);
193 return ret;
194}
195
196static void dump_completed_IO(struct inode *inode, struct list_head *head)
197{
198#ifdef EXT4FS_DEBUG
199 struct list_head *cur, *before, *after;
200 ext4_io_end_t *io_end, *io_end0, *io_end1;
201
202 if (list_empty(head))
203 return;
204
205 ext4_debug("Dump inode %lu completed io list\n", inode->i_ino);
206 list_for_each_entry(io_end, head, list) {
207 cur = &io_end->list;
208 before = cur->prev;
209 io_end0 = container_of(before, ext4_io_end_t, list);
210 after = cur->next;
211 io_end1 = container_of(after, ext4_io_end_t, list);
212
213 ext4_debug("io 0x%p from inode %lu,prev 0x%p,next 0x%p\n",
214 io_end, inode->i_ino, io_end0, io_end1);
215 }
216#endif
217}
218
219/* Add the io_end to per-inode completed end_io list. */
220static void ext4_add_complete_io(ext4_io_end_t *io_end)
221{
222 struct ext4_inode_info *ei = EXT4_I(io_end->inode);
223 struct ext4_sb_info *sbi = EXT4_SB(io_end->inode->i_sb);
224 struct workqueue_struct *wq;
225 unsigned long flags;
226
227 /* Only reserved conversions from writeback should enter here */
228 WARN_ON(!(io_end->flag & EXT4_IO_END_UNWRITTEN));
229 WARN_ON(!io_end->handle && sbi->s_journal);
230 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
231 wq = sbi->rsv_conversion_wq;
232 if (list_empty(&ei->i_rsv_conversion_list))
233 queue_work(wq, &ei->i_rsv_conversion_work);
234 list_add_tail(&io_end->list, &ei->i_rsv_conversion_list);
235 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
236}
237
238static int ext4_do_flush_completed_IO(struct inode *inode,
239 struct list_head *head)
240{
241 ext4_io_end_t *io_end;
242 struct list_head unwritten;
243 unsigned long flags;
244 struct ext4_inode_info *ei = EXT4_I(inode);
245 int err, ret = 0;
246
247 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
248 dump_completed_IO(inode, head);
249 list_replace_init(head, &unwritten);
250 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
251
252 while (!list_empty(&unwritten)) {
253 io_end = list_entry(unwritten.next, ext4_io_end_t, list);
254 BUG_ON(!(io_end->flag & EXT4_IO_END_UNWRITTEN));
255 list_del_init(&io_end->list);
256
257 err = ext4_end_io_end(io_end);
258 if (unlikely(!ret && err))
259 ret = err;
260 }
261 return ret;
262}
263
264/*
265 * work on completed IO, to convert unwritten extents to extents
266 */
267void ext4_end_io_rsv_work(struct work_struct *work)
268{
269 struct ext4_inode_info *ei = container_of(work, struct ext4_inode_info,
270 i_rsv_conversion_work);
271 ext4_do_flush_completed_IO(&ei->vfs_inode, &ei->i_rsv_conversion_list);
272}
273
274ext4_io_end_t *ext4_init_io_end(struct inode *inode, gfp_t flags)
275{
276 ext4_io_end_t *io_end = kmem_cache_zalloc(io_end_cachep, flags);
277
278 if (io_end) {
279 io_end->inode = inode;
280 INIT_LIST_HEAD(&io_end->list);
281 INIT_LIST_HEAD(&io_end->list_vec);
282 atomic_set(&io_end->count, 1);
283 }
284 return io_end;
285}
286
287void ext4_put_io_end_defer(ext4_io_end_t *io_end)
288{
289 if (atomic_dec_and_test(&io_end->count)) {
290 if (!(io_end->flag & EXT4_IO_END_UNWRITTEN) ||
291 list_empty(&io_end->list_vec)) {
292 ext4_release_io_end(io_end);
293 return;
294 }
295 ext4_add_complete_io(io_end);
296 }
297}
298
299int ext4_put_io_end(ext4_io_end_t *io_end)
300{
301 int err = 0;
302
303 if (atomic_dec_and_test(&io_end->count)) {
304 if (io_end->flag & EXT4_IO_END_UNWRITTEN) {
305 err = ext4_convert_unwritten_io_end_vec(io_end->handle,
306 io_end);
307 io_end->handle = NULL;
308 ext4_clear_io_unwritten_flag(io_end);
309 }
310 ext4_release_io_end(io_end);
311 }
312 return err;
313}
314
315ext4_io_end_t *ext4_get_io_end(ext4_io_end_t *io_end)
316{
317 atomic_inc(&io_end->count);
318 return io_end;
319}
320
321/* BIO completion function for page writeback */
322static void ext4_end_bio(struct bio *bio)
323{
324 ext4_io_end_t *io_end = bio->bi_private;
325 sector_t bi_sector = bio->bi_iter.bi_sector;
326 char b[BDEVNAME_SIZE];
327
328 if (WARN_ONCE(!io_end, "io_end is NULL: %s: sector %Lu len %u err %d\n",
329 bio_devname(bio, b),
330 (long long) bio->bi_iter.bi_sector,
331 (unsigned) bio_sectors(bio),
332 bio->bi_status)) {
333 ext4_finish_bio(bio);
334 bio_put(bio);
335 return;
336 }
337 bio->bi_end_io = NULL;
338
339 if (bio->bi_status) {
340 struct inode *inode = io_end->inode;
341
342 ext4_warning(inode->i_sb, "I/O error %d writing to inode %lu "
343 "starting block %llu)",
344 bio->bi_status, inode->i_ino,
345 (unsigned long long)
346 bi_sector >> (inode->i_blkbits - 9));
347 mapping_set_error(inode->i_mapping,
348 blk_status_to_errno(bio->bi_status));
349 }
350
351 if (io_end->flag & EXT4_IO_END_UNWRITTEN) {
352 /*
353 * Link bio into list hanging from io_end. We have to do it
354 * atomically as bio completions can be racing against each
355 * other.
356 */
357 bio->bi_private = xchg(&io_end->bio, bio);
358 ext4_put_io_end_defer(io_end);
359 } else {
360 /*
361 * Drop io_end reference early. Inode can get freed once
362 * we finish the bio.
363 */
364 ext4_put_io_end_defer(io_end);
365 ext4_finish_bio(bio);
366 bio_put(bio);
367 }
368}
369
370void ext4_io_submit(struct ext4_io_submit *io)
371{
372 struct bio *bio = io->io_bio;
373
374 if (bio) {
375 int io_op_flags = io->io_wbc->sync_mode == WB_SYNC_ALL ?
376 REQ_SYNC : 0;
377 io->io_bio->bi_write_hint = io->io_end->inode->i_write_hint;
378 bio_set_op_attrs(io->io_bio, REQ_OP_WRITE, io_op_flags);
379 submit_bio(io->io_bio);
380 }
381 io->io_bio = NULL;
382}
383
384void ext4_io_submit_init(struct ext4_io_submit *io,
385 struct writeback_control *wbc)
386{
387 io->io_wbc = wbc;
388 io->io_bio = NULL;
389 io->io_end = NULL;
390}
391
392static void io_submit_init_bio(struct ext4_io_submit *io,
393 struct buffer_head *bh)
394{
395 struct bio *bio;
396
397 /*
398 * bio_alloc will _always_ be able to allocate a bio if
399 * __GFP_DIRECT_RECLAIM is set, see comments for bio_alloc_bioset().
400 */
401 bio = bio_alloc(GFP_NOIO, BIO_MAX_VECS);
402 fscrypt_set_bio_crypt_ctx_bh(bio, bh, GFP_NOIO);
403 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
404 bio_set_dev(bio, bh->b_bdev);
405 bio->bi_end_io = ext4_end_bio;
406 bio->bi_private = ext4_get_io_end(io->io_end);
407 io->io_bio = bio;
408 io->io_next_block = bh->b_blocknr;
409 wbc_init_bio(io->io_wbc, bio);
410}
411
412static void io_submit_add_bh(struct ext4_io_submit *io,
413 struct inode *inode,
414 struct page *page,
415 struct buffer_head *bh)
416{
417 int ret;
418
419 if (io->io_bio && (bh->b_blocknr != io->io_next_block ||
420 !fscrypt_mergeable_bio_bh(io->io_bio, bh))) {
421submit_and_retry:
422 ext4_io_submit(io);
423 }
424 if (io->io_bio == NULL) {
425 io_submit_init_bio(io, bh);
426 io->io_bio->bi_write_hint = inode->i_write_hint;
427 }
428 ret = bio_add_page(io->io_bio, page, bh->b_size, bh_offset(bh));
429 if (ret != bh->b_size)
430 goto submit_and_retry;
431 wbc_account_cgroup_owner(io->io_wbc, page, bh->b_size);
432 io->io_next_block++;
433}
434
435int ext4_bio_write_page(struct ext4_io_submit *io,
436 struct page *page,
437 int len,
438 bool keep_towrite)
439{
440 struct page *bounce_page = NULL;
441 struct inode *inode = page->mapping->host;
442 unsigned block_start;
443 struct buffer_head *bh, *head;
444 int ret = 0;
445 int nr_submitted = 0;
446 int nr_to_submit = 0;
447 struct writeback_control *wbc = io->io_wbc;
448
449 BUG_ON(!PageLocked(page));
450 BUG_ON(PageWriteback(page));
451
452 if (keep_towrite)
453 set_page_writeback_keepwrite(page);
454 else
455 set_page_writeback(page);
456 ClearPageError(page);
457
458 /*
459 * Comments copied from block_write_full_page:
460 *
461 * The page straddles i_size. It must be zeroed out on each and every
462 * writepage invocation because it may be mmapped. "A file is mapped
463 * in multiples of the page size. For a file that is not a multiple of
464 * the page size, the remaining memory is zeroed when mapped, and
465 * writes to that region are not written out to the file."
466 */
467 if (len < PAGE_SIZE)
468 zero_user_segment(page, len, PAGE_SIZE);
469 /*
470 * In the first loop we prepare and mark buffers to submit. We have to
471 * mark all buffers in the page before submitting so that
472 * end_page_writeback() cannot be called from ext4_bio_end_io() when IO
473 * on the first buffer finishes and we are still working on submitting
474 * the second buffer.
475 */
476 bh = head = page_buffers(page);
477 do {
478 block_start = bh_offset(bh);
479 if (block_start >= len) {
480 clear_buffer_dirty(bh);
481 set_buffer_uptodate(bh);
482 continue;
483 }
484 if (!buffer_dirty(bh) || buffer_delay(bh) ||
485 !buffer_mapped(bh) || buffer_unwritten(bh)) {
486 /* A hole? We can safely clear the dirty bit */
487 if (!buffer_mapped(bh))
488 clear_buffer_dirty(bh);
489 if (io->io_bio)
490 ext4_io_submit(io);
491 continue;
492 }
493 if (buffer_new(bh))
494 clear_buffer_new(bh);
495 set_buffer_async_write(bh);
496 nr_to_submit++;
497 } while ((bh = bh->b_this_page) != head);
498
499 bh = head = page_buffers(page);
500
501 /*
502 * If any blocks are being written to an encrypted file, encrypt them
503 * into a bounce page. For simplicity, just encrypt until the last
504 * block which might be needed. This may cause some unneeded blocks
505 * (e.g. holes) to be unnecessarily encrypted, but this is rare and
506 * can't happen in the common case of blocksize == PAGE_SIZE.
507 */
508 if (fscrypt_inode_uses_fs_layer_crypto(inode) && nr_to_submit) {
509 gfp_t gfp_flags = GFP_NOFS;
510 unsigned int enc_bytes = round_up(len, i_blocksize(inode));
511
512 /*
513 * Since bounce page allocation uses a mempool, we can only use
514 * a waiting mask (i.e. request guaranteed allocation) on the
515 * first page of the bio. Otherwise it can deadlock.
516 */
517 if (io->io_bio)
518 gfp_flags = GFP_NOWAIT | __GFP_NOWARN;
519 retry_encrypt:
520 bounce_page = fscrypt_encrypt_pagecache_blocks(page, enc_bytes,
521 0, gfp_flags);
522 if (IS_ERR(bounce_page)) {
523 ret = PTR_ERR(bounce_page);
524 if (ret == -ENOMEM &&
525 (io->io_bio || wbc->sync_mode == WB_SYNC_ALL)) {
526 gfp_flags = GFP_NOFS;
527 if (io->io_bio)
528 ext4_io_submit(io);
529 else
530 gfp_flags |= __GFP_NOFAIL;
531 congestion_wait(BLK_RW_ASYNC, HZ/50);
532 goto retry_encrypt;
533 }
534
535 printk_ratelimited(KERN_ERR "%s: ret = %d\n", __func__, ret);
536 redirty_page_for_writepage(wbc, page);
537 do {
538 clear_buffer_async_write(bh);
539 bh = bh->b_this_page;
540 } while (bh != head);
541 goto unlock;
542 }
543 }
544
545 /* Now submit buffers to write */
546 do {
547 if (!buffer_async_write(bh))
548 continue;
549 io_submit_add_bh(io, inode,
550 bounce_page ? bounce_page : page, bh);
551 nr_submitted++;
552 clear_buffer_dirty(bh);
553 } while ((bh = bh->b_this_page) != head);
554
555unlock:
556 unlock_page(page);
557 /* Nothing submitted - we have to end page writeback */
558 if (!nr_submitted)
559 end_page_writeback(page);
560 return ret;
561}
1/*
2 * linux/fs/ext4/page-io.c
3 *
4 * This contains the new page_io functions for ext4
5 *
6 * Written by Theodore Ts'o, 2010.
7 */
8
9#include <linux/fs.h>
10#include <linux/time.h>
11#include <linux/highuid.h>
12#include <linux/pagemap.h>
13#include <linux/quotaops.h>
14#include <linux/string.h>
15#include <linux/buffer_head.h>
16#include <linux/writeback.h>
17#include <linux/pagevec.h>
18#include <linux/mpage.h>
19#include <linux/namei.h>
20#include <linux/uio.h>
21#include <linux/bio.h>
22#include <linux/workqueue.h>
23#include <linux/kernel.h>
24#include <linux/slab.h>
25#include <linux/mm.h>
26#include <linux/backing-dev.h>
27
28#include "ext4_jbd2.h"
29#include "xattr.h"
30#include "acl.h"
31
32static struct kmem_cache *io_end_cachep;
33
34int __init ext4_init_pageio(void)
35{
36 io_end_cachep = KMEM_CACHE(ext4_io_end, SLAB_RECLAIM_ACCOUNT);
37 if (io_end_cachep == NULL)
38 return -ENOMEM;
39 return 0;
40}
41
42void ext4_exit_pageio(void)
43{
44 kmem_cache_destroy(io_end_cachep);
45}
46
47/*
48 * Print an buffer I/O error compatible with the fs/buffer.c. This
49 * provides compatibility with dmesg scrapers that look for a specific
50 * buffer I/O error message. We really need a unified error reporting
51 * structure to userspace ala Digital Unix's uerf system, but it's
52 * probably not going to happen in my lifetime, due to LKML politics...
53 */
54static void buffer_io_error(struct buffer_head *bh)
55{
56 printk_ratelimited(KERN_ERR "Buffer I/O error on device %pg, logical block %llu\n",
57 bh->b_bdev,
58 (unsigned long long)bh->b_blocknr);
59}
60
61static void ext4_finish_bio(struct bio *bio)
62{
63 int i;
64 struct bio_vec *bvec;
65
66 bio_for_each_segment_all(bvec, bio, i) {
67 struct page *page = bvec->bv_page;
68#ifdef CONFIG_EXT4_FS_ENCRYPTION
69 struct page *data_page = NULL;
70 struct ext4_crypto_ctx *ctx = NULL;
71#endif
72 struct buffer_head *bh, *head;
73 unsigned bio_start = bvec->bv_offset;
74 unsigned bio_end = bio_start + bvec->bv_len;
75 unsigned under_io = 0;
76 unsigned long flags;
77
78 if (!page)
79 continue;
80
81#ifdef CONFIG_EXT4_FS_ENCRYPTION
82 if (!page->mapping) {
83 /* The bounce data pages are unmapped. */
84 data_page = page;
85 ctx = (struct ext4_crypto_ctx *)page_private(data_page);
86 page = ctx->w.control_page;
87 }
88#endif
89
90 if (bio->bi_error) {
91 SetPageError(page);
92 set_bit(AS_EIO, &page->mapping->flags);
93 }
94 bh = head = page_buffers(page);
95 /*
96 * We check all buffers in the page under BH_Uptodate_Lock
97 * to avoid races with other end io clearing async_write flags
98 */
99 local_irq_save(flags);
100 bit_spin_lock(BH_Uptodate_Lock, &head->b_state);
101 do {
102 if (bh_offset(bh) < bio_start ||
103 bh_offset(bh) + bh->b_size > bio_end) {
104 if (buffer_async_write(bh))
105 under_io++;
106 continue;
107 }
108 clear_buffer_async_write(bh);
109 if (bio->bi_error)
110 buffer_io_error(bh);
111 } while ((bh = bh->b_this_page) != head);
112 bit_spin_unlock(BH_Uptodate_Lock, &head->b_state);
113 local_irq_restore(flags);
114 if (!under_io) {
115#ifdef CONFIG_EXT4_FS_ENCRYPTION
116 if (ctx)
117 ext4_restore_control_page(data_page);
118#endif
119 end_page_writeback(page);
120 }
121 }
122}
123
124static void ext4_release_io_end(ext4_io_end_t *io_end)
125{
126 struct bio *bio, *next_bio;
127
128 BUG_ON(!list_empty(&io_end->list));
129 BUG_ON(io_end->flag & EXT4_IO_END_UNWRITTEN);
130 WARN_ON(io_end->handle);
131
132 for (bio = io_end->bio; bio; bio = next_bio) {
133 next_bio = bio->bi_private;
134 ext4_finish_bio(bio);
135 bio_put(bio);
136 }
137 kmem_cache_free(io_end_cachep, io_end);
138}
139
140/*
141 * Check a range of space and convert unwritten extents to written. Note that
142 * we are protected from truncate touching same part of extent tree by the
143 * fact that truncate code waits for all DIO to finish (thus exclusion from
144 * direct IO is achieved) and also waits for PageWriteback bits. Thus we
145 * cannot get to ext4_ext_truncate() before all IOs overlapping that range are
146 * completed (happens from ext4_free_ioend()).
147 */
148static int ext4_end_io(ext4_io_end_t *io)
149{
150 struct inode *inode = io->inode;
151 loff_t offset = io->offset;
152 ssize_t size = io->size;
153 handle_t *handle = io->handle;
154 int ret = 0;
155
156 ext4_debug("ext4_end_io_nolock: io 0x%p from inode %lu,list->next 0x%p,"
157 "list->prev 0x%p\n",
158 io, inode->i_ino, io->list.next, io->list.prev);
159
160 io->handle = NULL; /* Following call will use up the handle */
161 ret = ext4_convert_unwritten_extents(handle, inode, offset, size);
162 if (ret < 0) {
163 ext4_msg(inode->i_sb, KERN_EMERG,
164 "failed to convert unwritten extents to written "
165 "extents -- potential data loss! "
166 "(inode %lu, offset %llu, size %zd, error %d)",
167 inode->i_ino, offset, size, ret);
168 }
169 ext4_clear_io_unwritten_flag(io);
170 ext4_release_io_end(io);
171 return ret;
172}
173
174static void dump_completed_IO(struct inode *inode, struct list_head *head)
175{
176#ifdef EXT4FS_DEBUG
177 struct list_head *cur, *before, *after;
178 ext4_io_end_t *io, *io0, *io1;
179
180 if (list_empty(head))
181 return;
182
183 ext4_debug("Dump inode %lu completed io list\n", inode->i_ino);
184 list_for_each_entry(io, head, list) {
185 cur = &io->list;
186 before = cur->prev;
187 io0 = container_of(before, ext4_io_end_t, list);
188 after = cur->next;
189 io1 = container_of(after, ext4_io_end_t, list);
190
191 ext4_debug("io 0x%p from inode %lu,prev 0x%p,next 0x%p\n",
192 io, inode->i_ino, io0, io1);
193 }
194#endif
195}
196
197/* Add the io_end to per-inode completed end_io list. */
198static void ext4_add_complete_io(ext4_io_end_t *io_end)
199{
200 struct ext4_inode_info *ei = EXT4_I(io_end->inode);
201 struct ext4_sb_info *sbi = EXT4_SB(io_end->inode->i_sb);
202 struct workqueue_struct *wq;
203 unsigned long flags;
204
205 /* Only reserved conversions from writeback should enter here */
206 WARN_ON(!(io_end->flag & EXT4_IO_END_UNWRITTEN));
207 WARN_ON(!io_end->handle && sbi->s_journal);
208 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
209 wq = sbi->rsv_conversion_wq;
210 if (list_empty(&ei->i_rsv_conversion_list))
211 queue_work(wq, &ei->i_rsv_conversion_work);
212 list_add_tail(&io_end->list, &ei->i_rsv_conversion_list);
213 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
214}
215
216static int ext4_do_flush_completed_IO(struct inode *inode,
217 struct list_head *head)
218{
219 ext4_io_end_t *io;
220 struct list_head unwritten;
221 unsigned long flags;
222 struct ext4_inode_info *ei = EXT4_I(inode);
223 int err, ret = 0;
224
225 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
226 dump_completed_IO(inode, head);
227 list_replace_init(head, &unwritten);
228 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
229
230 while (!list_empty(&unwritten)) {
231 io = list_entry(unwritten.next, ext4_io_end_t, list);
232 BUG_ON(!(io->flag & EXT4_IO_END_UNWRITTEN));
233 list_del_init(&io->list);
234
235 err = ext4_end_io(io);
236 if (unlikely(!ret && err))
237 ret = err;
238 }
239 return ret;
240}
241
242/*
243 * work on completed IO, to convert unwritten extents to extents
244 */
245void ext4_end_io_rsv_work(struct work_struct *work)
246{
247 struct ext4_inode_info *ei = container_of(work, struct ext4_inode_info,
248 i_rsv_conversion_work);
249 ext4_do_flush_completed_IO(&ei->vfs_inode, &ei->i_rsv_conversion_list);
250}
251
252ext4_io_end_t *ext4_init_io_end(struct inode *inode, gfp_t flags)
253{
254 ext4_io_end_t *io = kmem_cache_zalloc(io_end_cachep, flags);
255 if (io) {
256 io->inode = inode;
257 INIT_LIST_HEAD(&io->list);
258 atomic_set(&io->count, 1);
259 }
260 return io;
261}
262
263void ext4_put_io_end_defer(ext4_io_end_t *io_end)
264{
265 if (atomic_dec_and_test(&io_end->count)) {
266 if (!(io_end->flag & EXT4_IO_END_UNWRITTEN) || !io_end->size) {
267 ext4_release_io_end(io_end);
268 return;
269 }
270 ext4_add_complete_io(io_end);
271 }
272}
273
274int ext4_put_io_end(ext4_io_end_t *io_end)
275{
276 int err = 0;
277
278 if (atomic_dec_and_test(&io_end->count)) {
279 if (io_end->flag & EXT4_IO_END_UNWRITTEN) {
280 err = ext4_convert_unwritten_extents(io_end->handle,
281 io_end->inode, io_end->offset,
282 io_end->size);
283 io_end->handle = NULL;
284 ext4_clear_io_unwritten_flag(io_end);
285 }
286 ext4_release_io_end(io_end);
287 }
288 return err;
289}
290
291ext4_io_end_t *ext4_get_io_end(ext4_io_end_t *io_end)
292{
293 atomic_inc(&io_end->count);
294 return io_end;
295}
296
297/* BIO completion function for page writeback */
298static void ext4_end_bio(struct bio *bio)
299{
300 ext4_io_end_t *io_end = bio->bi_private;
301 sector_t bi_sector = bio->bi_iter.bi_sector;
302
303 BUG_ON(!io_end);
304 bio->bi_end_io = NULL;
305
306 if (bio->bi_error) {
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_error, 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, bio->bi_error);
317 }
318
319 if (io_end->flag & EXT4_IO_END_UNWRITTEN) {
320 /*
321 * Link bio into list hanging from io_end. We have to do it
322 * atomically as bio completions can be racing against each
323 * other.
324 */
325 bio->bi_private = xchg(&io_end->bio, bio);
326 ext4_put_io_end_defer(io_end);
327 } else {
328 /*
329 * Drop io_end reference early. Inode can get freed once
330 * we finish the bio.
331 */
332 ext4_put_io_end_defer(io_end);
333 ext4_finish_bio(bio);
334 bio_put(bio);
335 }
336}
337
338void ext4_io_submit(struct ext4_io_submit *io)
339{
340 struct bio *bio = io->io_bio;
341
342 if (bio) {
343 int io_op = io->io_wbc->sync_mode == WB_SYNC_ALL ?
344 WRITE_SYNC : WRITE;
345 bio_get(io->io_bio);
346 submit_bio(io_op, io->io_bio);
347 bio_put(io->io_bio);
348 }
349 io->io_bio = NULL;
350}
351
352void ext4_io_submit_init(struct ext4_io_submit *io,
353 struct writeback_control *wbc)
354{
355 io->io_wbc = wbc;
356 io->io_bio = NULL;
357 io->io_end = NULL;
358}
359
360static int io_submit_init_bio(struct ext4_io_submit *io,
361 struct buffer_head *bh)
362{
363 struct bio *bio;
364
365 bio = bio_alloc(GFP_NOIO, BIO_MAX_PAGES);
366 if (!bio)
367 return -ENOMEM;
368 wbc_init_bio(io->io_wbc, bio);
369 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
370 bio->bi_bdev = 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 return 0;
376}
377
378static int io_submit_add_bh(struct ext4_io_submit *io,
379 struct inode *inode,
380 struct page *page,
381 struct buffer_head *bh)
382{
383 int ret;
384
385 if (io->io_bio && bh->b_blocknr != io->io_next_block) {
386submit_and_retry:
387 ext4_io_submit(io);
388 }
389 if (io->io_bio == NULL) {
390 ret = io_submit_init_bio(io, bh);
391 if (ret)
392 return ret;
393 }
394 ret = bio_add_page(io->io_bio, page, bh->b_size, bh_offset(bh));
395 if (ret != bh->b_size)
396 goto submit_and_retry;
397 wbc_account_io(io->io_wbc, page, bh->b_size);
398 io->io_next_block++;
399 return 0;
400}
401
402int ext4_bio_write_page(struct ext4_io_submit *io,
403 struct page *page,
404 int len,
405 struct writeback_control *wbc,
406 bool keep_towrite)
407{
408 struct page *data_page = NULL;
409 struct inode *inode = page->mapping->host;
410 unsigned block_start, blocksize;
411 struct buffer_head *bh, *head;
412 int ret = 0;
413 int nr_submitted = 0;
414 int nr_to_submit = 0;
415
416 blocksize = 1 << inode->i_blkbits;
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 unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
465 }
466 set_buffer_async_write(bh);
467 nr_to_submit++;
468 } while ((bh = bh->b_this_page) != head);
469
470 bh = head = page_buffers(page);
471
472 if (ext4_encrypted_inode(inode) && S_ISREG(inode->i_mode) &&
473 nr_to_submit) {
474 gfp_t gfp_flags = GFP_NOFS;
475
476 retry_encrypt:
477 data_page = ext4_encrypt(inode, page, gfp_flags);
478 if (IS_ERR(data_page)) {
479 ret = PTR_ERR(data_page);
480 if (ret == -ENOMEM && wbc->sync_mode == WB_SYNC_ALL) {
481 if (io->io_bio) {
482 ext4_io_submit(io);
483 congestion_wait(BLK_RW_ASYNC, HZ/50);
484 }
485 gfp_flags |= __GFP_NOFAIL;
486 goto retry_encrypt;
487 }
488 data_page = NULL;
489 goto out;
490 }
491 }
492
493 /* Now submit buffers to write */
494 do {
495 if (!buffer_async_write(bh))
496 continue;
497 ret = io_submit_add_bh(io, inode,
498 data_page ? data_page : page, bh);
499 if (ret) {
500 /*
501 * We only get here on ENOMEM. Not much else
502 * we can do but mark the page as dirty, and
503 * better luck next time.
504 */
505 break;
506 }
507 nr_submitted++;
508 clear_buffer_dirty(bh);
509 } while ((bh = bh->b_this_page) != head);
510
511 /* Error stopped previous loop? Clean up buffers... */
512 if (ret) {
513 out:
514 if (data_page)
515 ext4_restore_control_page(data_page);
516 printk_ratelimited(KERN_ERR "%s: ret = %d\n", __func__, ret);
517 redirty_page_for_writepage(wbc, page);
518 do {
519 clear_buffer_async_write(bh);
520 bh = bh->b_this_page;
521 } while (bh != head);
522 }
523 unlock_page(page);
524 /* Nothing submitted - we have to end page writeback */
525 if (!nr_submitted)
526 end_page_writeback(page);
527 return ret;
528}