1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23#include <linux/kernel.h>
24#include <linux/module.h>
25#include <linux/types.h>
26#include <linux/fs.h>
27#include <linux/mm.h>
28#include <linux/slab.h>
29#include <linux/highmem.h>
30#include <linux/pagemap.h>
31#include <linux/task_io_accounting_ops.h>
32#include <linux/bio.h>
33#include <linux/wait.h>
34#include <linux/err.h>
35#include <linux/blkdev.h>
36#include <linux/buffer_head.h>
37#include <linux/rwsem.h>
38#include <linux/uio.h>
39#include <linux/atomic.h>
40#include <linux/prefetch.h>
41
42#include "internal.h"
43
44
45
46
47
48#define DIO_PAGES 64
49
50
51
52
53#define DIO_COMPLETE_ASYNC 0x01
54#define DIO_COMPLETE_INVALIDATE 0x02
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70struct dio_submit {
71 struct bio *bio;
72 unsigned blkbits;
73 unsigned blkfactor;
74
75
76
77
78 unsigned start_zero_done;
79
80
81 int pages_in_io;
82 sector_t block_in_file;
83
84 unsigned blocks_available;
85 int reap_counter;
86 sector_t final_block_in_request;
87 int boundary;
88 get_block_t *get_block;
89 dio_submit_t *submit_io;
90
91 loff_t logical_offset_in_bio;
92 sector_t final_block_in_bio;
93 sector_t next_block_for_io;
94
95
96
97
98
99
100
101 struct page *cur_page;
102 unsigned cur_page_offset;
103 unsigned cur_page_len;
104 sector_t cur_page_block;
105 loff_t cur_page_fs_offset;
106
107 struct iov_iter *iter;
108
109
110
111
112 unsigned head;
113 unsigned tail;
114 size_t from, to;
115};
116
117
118struct dio {
119 int flags;
120 int op;
121 int op_flags;
122 blk_qc_t bio_cookie;
123 struct gendisk *bio_disk;
124 struct inode *inode;
125 loff_t i_size;
126 dio_iodone_t *end_io;
127
128 void *private;
129
130
131 spinlock_t bio_lock;
132 int page_errors;
133 int is_async;
134 bool defer_completion;
135 bool should_dirty;
136 int io_error;
137 unsigned long refcount;
138 struct bio *bio_list;
139 struct task_struct *waiter;
140
141
142 struct kiocb *iocb;
143 ssize_t result;
144
145
146
147
148
149
150 union {
151 struct page *pages[DIO_PAGES];
152 struct work_struct complete_work;
153 };
154} ____cacheline_aligned_in_smp;
155
156static struct kmem_cache *dio_cache __read_mostly;
157
158
159
160
161static inline unsigned dio_pages_present(struct dio_submit *sdio)
162{
163 return sdio->tail - sdio->head;
164}
165
166
167
168
169static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
170{
171 ssize_t ret;
172
173 ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES,
174 &sdio->from);
175
176 if (ret < 0 && sdio->blocks_available && (dio->op == REQ_OP_WRITE)) {
177 struct page *page = ZERO_PAGE(0);
178
179
180
181
182
183 if (dio->page_errors == 0)
184 dio->page_errors = ret;
185 get_page(page);
186 dio->pages[0] = page;
187 sdio->head = 0;
188 sdio->tail = 1;
189 sdio->from = 0;
190 sdio->to = PAGE_SIZE;
191 return 0;
192 }
193
194 if (ret >= 0) {
195 iov_iter_advance(sdio->iter, ret);
196 ret += sdio->from;
197 sdio->head = 0;
198 sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
199 sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1;
200 return 0;
201 }
202 return ret;
203}
204
205
206
207
208
209
210
211static inline struct page *dio_get_page(struct dio *dio,
212 struct dio_submit *sdio)
213{
214 if (dio_pages_present(sdio) == 0) {
215 int ret;
216
217 ret = dio_refill_pages(dio, sdio);
218 if (ret)
219 return ERR_PTR(ret);
220 BUG_ON(dio_pages_present(sdio) == 0);
221 }
222 return dio->pages[sdio->head];
223}
224
225
226
227
228
229
230
231
232
233
234
235
236static ssize_t dio_complete(struct dio *dio, ssize_t ret, unsigned int flags)
237{
238 loff_t offset = dio->iocb->ki_pos;
239 ssize_t transferred = 0;
240 int err;
241
242
243
244
245
246
247
248 if (ret == -EIOCBQUEUED)
249 ret = 0;
250
251 if (dio->result) {
252 transferred = dio->result;
253
254
255 if ((dio->op == REQ_OP_READ) &&
256 ((offset + transferred) > dio->i_size))
257 transferred = dio->i_size - offset;
258
259 if (unlikely(ret == -EFAULT) && transferred)
260 ret = 0;
261 }
262
263 if (ret == 0)
264 ret = dio->page_errors;
265 if (ret == 0)
266 ret = dio->io_error;
267 if (ret == 0)
268 ret = transferred;
269
270 if (dio->end_io) {
271
272 err = dio->end_io(dio->iocb, offset, ret, dio->private);
273 if (err)
274 ret = err;
275 }
276
277
278
279
280
281
282
283
284
285
286
287
288
289 if (flags & DIO_COMPLETE_INVALIDATE &&
290 ret > 0 && dio->op == REQ_OP_WRITE &&
291 dio->inode->i_mapping->nrpages) {
292 err = invalidate_inode_pages2_range(dio->inode->i_mapping,
293 offset >> PAGE_SHIFT,
294 (offset + ret - 1) >> PAGE_SHIFT);
295 if (err)
296 dio_warn_stale_pagecache(dio->iocb->ki_filp);
297 }
298
299 inode_dio_end(dio->inode);
300
301 if (flags & DIO_COMPLETE_ASYNC) {
302
303
304
305
306
307 dio->iocb->ki_pos += transferred;
308
309 if (ret > 0 && dio->op == REQ_OP_WRITE)
310 ret = generic_write_sync(dio->iocb, ret);
311 dio->iocb->ki_complete(dio->iocb, ret, 0);
312 }
313
314 kmem_cache_free(dio_cache, dio);
315 return ret;
316}
317
318static void dio_aio_complete_work(struct work_struct *work)
319{
320 struct dio *dio = container_of(work, struct dio, complete_work);
321
322 dio_complete(dio, 0, DIO_COMPLETE_ASYNC | DIO_COMPLETE_INVALIDATE);
323}
324
325static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio);
326
327
328
329
330static void dio_bio_end_aio(struct bio *bio)
331{
332 struct dio *dio = bio->bi_private;
333 unsigned long remaining;
334 unsigned long flags;
335 bool defer_completion = false;
336
337
338 dio_bio_complete(dio, bio);
339
340 spin_lock_irqsave(&dio->bio_lock, flags);
341 remaining = --dio->refcount;
342 if (remaining == 1 && dio->waiter)
343 wake_up_process(dio->waiter);
344 spin_unlock_irqrestore(&dio->bio_lock, flags);
345
346 if (remaining == 0) {
347
348
349
350
351
352
353
354
355 if (dio->result)
356 defer_completion = dio->defer_completion ||
357 (dio->op == REQ_OP_WRITE &&
358 dio->inode->i_mapping->nrpages);
359 if (defer_completion) {
360 INIT_WORK(&dio->complete_work, dio_aio_complete_work);
361 queue_work(dio->inode->i_sb->s_dio_done_wq,
362 &dio->complete_work);
363 } else {
364 dio_complete(dio, 0, DIO_COMPLETE_ASYNC);
365 }
366 }
367}
368
369
370
371
372
373
374
375
376static void dio_bio_end_io(struct bio *bio)
377{
378 struct dio *dio = bio->bi_private;
379 unsigned long flags;
380
381 spin_lock_irqsave(&dio->bio_lock, flags);
382 bio->bi_private = dio->bio_list;
383 dio->bio_list = bio;
384 if (--dio->refcount == 1 && dio->waiter)
385 wake_up_process(dio->waiter);
386 spin_unlock_irqrestore(&dio->bio_lock, flags);
387}
388
389static inline void
390dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
391 struct block_device *bdev,
392 sector_t first_sector, int nr_vecs)
393{
394 struct bio *bio;
395
396
397
398
399
400 bio = bio_alloc(GFP_KERNEL, nr_vecs);
401
402 bio_set_dev(bio, bdev);
403 bio->bi_iter.bi_sector = first_sector;
404 bio_set_op_attrs(bio, dio->op, dio->op_flags);
405 if (dio->is_async)
406 bio->bi_end_io = dio_bio_end_aio;
407 else
408 bio->bi_end_io = dio_bio_end_io;
409
410 bio->bi_write_hint = dio->iocb->ki_hint;
411
412 sdio->bio = bio;
413 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
414}
415
416
417
418
419
420
421
422
423static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
424{
425 struct bio *bio = sdio->bio;
426 unsigned long flags;
427
428 bio->bi_private = dio;
429
430 bio_clear_flag(bio, BIO_WORKINGSET);
431
432 spin_lock_irqsave(&dio->bio_lock, flags);
433 dio->refcount++;
434 spin_unlock_irqrestore(&dio->bio_lock, flags);
435
436 if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty)
437 bio_set_pages_dirty(bio);
438
439 dio->bio_disk = bio->bi_bdev->bd_disk;
440
441 if (sdio->submit_io) {
442 sdio->submit_io(bio, dio->inode, sdio->logical_offset_in_bio);
443 dio->bio_cookie = BLK_QC_T_NONE;
444 } else
445 dio->bio_cookie = submit_bio(bio);
446
447 sdio->bio = NULL;
448 sdio->boundary = 0;
449 sdio->logical_offset_in_bio = 0;
450}
451
452
453
454
455static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
456{
457 while (sdio->head < sdio->tail)
458 put_page(dio->pages[sdio->head++]);
459}
460
461
462
463
464
465
466
467static struct bio *dio_await_one(struct dio *dio)
468{
469 unsigned long flags;
470 struct bio *bio = NULL;
471
472 spin_lock_irqsave(&dio->bio_lock, flags);
473
474
475
476
477
478
479
480 while (dio->refcount > 1 && dio->bio_list == NULL) {
481 __set_current_state(TASK_UNINTERRUPTIBLE);
482 dio->waiter = current;
483 spin_unlock_irqrestore(&dio->bio_lock, flags);
484 if (!(dio->iocb->ki_flags & IOCB_HIPRI) ||
485 !blk_poll(dio->bio_disk->queue, dio->bio_cookie, true))
486 blk_io_schedule();
487
488 spin_lock_irqsave(&dio->bio_lock, flags);
489 dio->waiter = NULL;
490 }
491 if (dio->bio_list) {
492 bio = dio->bio_list;
493 dio->bio_list = bio->bi_private;
494 }
495 spin_unlock_irqrestore(&dio->bio_lock, flags);
496 return bio;
497}
498
499
500
501
502static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio)
503{
504 blk_status_t err = bio->bi_status;
505 bool should_dirty = dio->op == REQ_OP_READ && dio->should_dirty;
506
507 if (err) {
508 if (err == BLK_STS_AGAIN && (bio->bi_opf & REQ_NOWAIT))
509 dio->io_error = -EAGAIN;
510 else
511 dio->io_error = -EIO;
512 }
513
514 if (dio->is_async && should_dirty) {
515 bio_check_pages_dirty(bio);
516 } else {
517 bio_release_pages(bio, should_dirty);
518 bio_put(bio);
519 }
520 return err;
521}
522
523
524
525
526
527
528
529
530static void dio_await_completion(struct dio *dio)
531{
532 struct bio *bio;
533 do {
534 bio = dio_await_one(dio);
535 if (bio)
536 dio_bio_complete(dio, bio);
537 } while (bio);
538}
539
540
541
542
543
544
545
546
547static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
548{
549 int ret = 0;
550
551 if (sdio->reap_counter++ >= 64) {
552 while (dio->bio_list) {
553 unsigned long flags;
554 struct bio *bio;
555 int ret2;
556
557 spin_lock_irqsave(&dio->bio_lock, flags);
558 bio = dio->bio_list;
559 dio->bio_list = bio->bi_private;
560 spin_unlock_irqrestore(&dio->bio_lock, flags);
561 ret2 = blk_status_to_errno(dio_bio_complete(dio, bio));
562 if (ret == 0)
563 ret = ret2;
564 }
565 sdio->reap_counter = 0;
566 }
567 return ret;
568}
569
570
571
572
573
574
575
576int sb_init_dio_done_wq(struct super_block *sb)
577{
578 struct workqueue_struct *old;
579 struct workqueue_struct *wq = alloc_workqueue("dio/%s",
580 WQ_MEM_RECLAIM, 0,
581 sb->s_id);
582 if (!wq)
583 return -ENOMEM;
584
585
586
587 old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
588
589 if (old)
590 destroy_workqueue(wq);
591 return 0;
592}
593
594static int dio_set_defer_completion(struct dio *dio)
595{
596 struct super_block *sb = dio->inode->i_sb;
597
598 if (dio->defer_completion)
599 return 0;
600 dio->defer_completion = true;
601 if (!sb->s_dio_done_wq)
602 return sb_init_dio_done_wq(sb);
603 return 0;
604}
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
630 struct buffer_head *map_bh)
631{
632 int ret;
633 sector_t fs_startblk;
634 sector_t fs_endblk;
635 unsigned long fs_count;
636 int create;
637 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
638 loff_t i_size;
639
640
641
642
643
644 ret = dio->page_errors;
645 if (ret == 0) {
646 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
647 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
648 fs_endblk = (sdio->final_block_in_request - 1) >>
649 sdio->blkfactor;
650 fs_count = fs_endblk - fs_startblk + 1;
651
652 map_bh->b_state = 0;
653 map_bh->b_size = fs_count << i_blkbits;
654
655
656
657
658
659
660
661
662
663
664
665
666 create = dio->op == REQ_OP_WRITE;
667 if (dio->flags & DIO_SKIP_HOLES) {
668 i_size = i_size_read(dio->inode);
669 if (i_size && fs_startblk <= (i_size - 1) >> i_blkbits)
670 create = 0;
671 }
672
673 ret = (*sdio->get_block)(dio->inode, fs_startblk,
674 map_bh, create);
675
676
677 dio->private = map_bh->b_private;
678
679 if (ret == 0 && buffer_defer_completion(map_bh))
680 ret = dio_set_defer_completion(dio);
681 }
682 return ret;
683}
684
685
686
687
688static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
689 sector_t start_sector, struct buffer_head *map_bh)
690{
691 sector_t sector;
692 int ret, nr_pages;
693
694 ret = dio_bio_reap(dio, sdio);
695 if (ret)
696 goto out;
697 sector = start_sector << (sdio->blkbits - 9);
698 nr_pages = bio_max_segs(sdio->pages_in_io);
699 BUG_ON(nr_pages <= 0);
700 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
701 sdio->boundary = 0;
702out:
703 return ret;
704}
705
706
707
708
709
710
711
712
713static inline int dio_bio_add_page(struct dio_submit *sdio)
714{
715 int ret;
716
717 ret = bio_add_page(sdio->bio, sdio->cur_page,
718 sdio->cur_page_len, sdio->cur_page_offset);
719 if (ret == sdio->cur_page_len) {
720
721
722
723 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
724 sdio->pages_in_io--;
725 get_page(sdio->cur_page);
726 sdio->final_block_in_bio = sdio->cur_page_block +
727 (sdio->cur_page_len >> sdio->blkbits);
728 ret = 0;
729 } else {
730 ret = 1;
731 }
732 return ret;
733}
734
735
736
737
738
739
740
741
742
743
744
745static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
746 struct buffer_head *map_bh)
747{
748 int ret = 0;
749
750 if (sdio->bio) {
751 loff_t cur_offset = sdio->cur_page_fs_offset;
752 loff_t bio_next_offset = sdio->logical_offset_in_bio +
753 sdio->bio->bi_iter.bi_size;
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769 if (sdio->final_block_in_bio != sdio->cur_page_block ||
770 cur_offset != bio_next_offset)
771 dio_bio_submit(dio, sdio);
772 }
773
774 if (sdio->bio == NULL) {
775 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
776 if (ret)
777 goto out;
778 }
779
780 if (dio_bio_add_page(sdio) != 0) {
781 dio_bio_submit(dio, sdio);
782 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
783 if (ret == 0) {
784 ret = dio_bio_add_page(sdio);
785 BUG_ON(ret != 0);
786 }
787 }
788out:
789 return ret;
790}
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809static inline int
810submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
811 unsigned offset, unsigned len, sector_t blocknr,
812 struct buffer_head *map_bh)
813{
814 int ret = 0;
815 int boundary = sdio->boundary;
816
817 if (dio->op == REQ_OP_WRITE) {
818
819
820
821 task_io_account_write(len);
822 }
823
824
825
826
827 if (sdio->cur_page == page &&
828 sdio->cur_page_offset + sdio->cur_page_len == offset &&
829 sdio->cur_page_block +
830 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
831 sdio->cur_page_len += len;
832 goto out;
833 }
834
835
836
837
838 if (sdio->cur_page) {
839 ret = dio_send_cur_page(dio, sdio, map_bh);
840 put_page(sdio->cur_page);
841 sdio->cur_page = NULL;
842 if (ret)
843 return ret;
844 }
845
846 get_page(page);
847 sdio->cur_page = page;
848 sdio->cur_page_offset = offset;
849 sdio->cur_page_len = len;
850 sdio->cur_page_block = blocknr;
851 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
852out:
853
854
855
856
857 if (boundary) {
858 ret = dio_send_cur_page(dio, sdio, map_bh);
859 if (sdio->bio)
860 dio_bio_submit(dio, sdio);
861 put_page(sdio->cur_page);
862 sdio->cur_page = NULL;
863 }
864 return ret;
865}
866
867
868
869
870
871
872
873
874
875
876static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
877 int end, struct buffer_head *map_bh)
878{
879 unsigned dio_blocks_per_fs_block;
880 unsigned this_chunk_blocks;
881 unsigned this_chunk_bytes;
882 struct page *page;
883
884 sdio->start_zero_done = 1;
885 if (!sdio->blkfactor || !buffer_new(map_bh))
886 return;
887
888 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
889 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
890
891 if (!this_chunk_blocks)
892 return;
893
894
895
896
897
898 if (end)
899 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
900
901 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
902
903 page = ZERO_PAGE(0);
904 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
905 sdio->next_block_for_io, map_bh))
906 return;
907
908 sdio->next_block_for_io += this_chunk_blocks;
909}
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
928 struct buffer_head *map_bh)
929{
930 const unsigned blkbits = sdio->blkbits;
931 const unsigned i_blkbits = blkbits + sdio->blkfactor;
932 int ret = 0;
933
934 while (sdio->block_in_file < sdio->final_block_in_request) {
935 struct page *page;
936 size_t from, to;
937
938 page = dio_get_page(dio, sdio);
939 if (IS_ERR(page)) {
940 ret = PTR_ERR(page);
941 goto out;
942 }
943 from = sdio->head ? 0 : sdio->from;
944 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
945 sdio->head++;
946
947 while (from < to) {
948 unsigned this_chunk_bytes;
949 unsigned this_chunk_blocks;
950 unsigned u;
951
952 if (sdio->blocks_available == 0) {
953
954
955
956 unsigned long blkmask;
957 unsigned long dio_remainder;
958
959 ret = get_more_blocks(dio, sdio, map_bh);
960 if (ret) {
961 put_page(page);
962 goto out;
963 }
964 if (!buffer_mapped(map_bh))
965 goto do_holes;
966
967 sdio->blocks_available =
968 map_bh->b_size >> blkbits;
969 sdio->next_block_for_io =
970 map_bh->b_blocknr << sdio->blkfactor;
971 if (buffer_new(map_bh)) {
972 clean_bdev_aliases(
973 map_bh->b_bdev,
974 map_bh->b_blocknr,
975 map_bh->b_size >> i_blkbits);
976 }
977
978 if (!sdio->blkfactor)
979 goto do_holes;
980
981 blkmask = (1 << sdio->blkfactor) - 1;
982 dio_remainder = (sdio->block_in_file & blkmask);
983
984
985
986
987
988
989
990
991
992
993
994
995 if (!buffer_new(map_bh))
996 sdio->next_block_for_io += dio_remainder;
997 sdio->blocks_available -= dio_remainder;
998 }
999do_holes:
1000
1001 if (!buffer_mapped(map_bh)) {
1002 loff_t i_size_aligned;
1003
1004
1005 if (dio->op == REQ_OP_WRITE) {
1006 put_page(page);
1007 return -ENOTBLK;
1008 }
1009
1010
1011
1012
1013
1014 i_size_aligned = ALIGN(i_size_read(dio->inode),
1015 1 << blkbits);
1016 if (sdio->block_in_file >=
1017 i_size_aligned >> blkbits) {
1018
1019 put_page(page);
1020 goto out;
1021 }
1022 zero_user(page, from, 1 << blkbits);
1023 sdio->block_in_file++;
1024 from += 1 << blkbits;
1025 dio->result += 1 << blkbits;
1026 goto next_block;
1027 }
1028
1029
1030
1031
1032
1033
1034 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
1035 dio_zero_block(dio, sdio, 0, map_bh);
1036
1037
1038
1039
1040
1041 this_chunk_blocks = sdio->blocks_available;
1042 u = (to - from) >> blkbits;
1043 if (this_chunk_blocks > u)
1044 this_chunk_blocks = u;
1045 u = sdio->final_block_in_request - sdio->block_in_file;
1046 if (this_chunk_blocks > u)
1047 this_chunk_blocks = u;
1048 this_chunk_bytes = this_chunk_blocks << blkbits;
1049 BUG_ON(this_chunk_bytes == 0);
1050
1051 if (this_chunk_blocks == sdio->blocks_available)
1052 sdio->boundary = buffer_boundary(map_bh);
1053 ret = submit_page_section(dio, sdio, page,
1054 from,
1055 this_chunk_bytes,
1056 sdio->next_block_for_io,
1057 map_bh);
1058 if (ret) {
1059 put_page(page);
1060 goto out;
1061 }
1062 sdio->next_block_for_io += this_chunk_blocks;
1063
1064 sdio->block_in_file += this_chunk_blocks;
1065 from += this_chunk_bytes;
1066 dio->result += this_chunk_bytes;
1067 sdio->blocks_available -= this_chunk_blocks;
1068next_block:
1069 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1070 if (sdio->block_in_file == sdio->final_block_in_request)
1071 break;
1072 }
1073
1074
1075 put_page(page);
1076 }
1077out:
1078 return ret;
1079}
1080
1081static inline int drop_refcount(struct dio *dio)
1082{
1083 int ret2;
1084 unsigned long flags;
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097 spin_lock_irqsave(&dio->bio_lock, flags);
1098 ret2 = --dio->refcount;
1099 spin_unlock_irqrestore(&dio->bio_lock, flags);
1100 return ret2;
1101}
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128static inline ssize_t
1129do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1130 struct block_device *bdev, struct iov_iter *iter,
1131 get_block_t get_block, dio_iodone_t end_io,
1132 dio_submit_t submit_io, int flags)
1133{
1134 unsigned i_blkbits = READ_ONCE(inode->i_blkbits);
1135 unsigned blkbits = i_blkbits;
1136 unsigned blocksize_mask = (1 << blkbits) - 1;
1137 ssize_t retval = -EINVAL;
1138 const size_t count = iov_iter_count(iter);
1139 loff_t offset = iocb->ki_pos;
1140 const loff_t end = offset + count;
1141 struct dio *dio;
1142 struct dio_submit sdio = { 0, };
1143 struct buffer_head map_bh = { 0, };
1144 struct blk_plug plug;
1145 unsigned long align = offset | iov_iter_alignment(iter);
1146
1147
1148
1149
1150
1151
1152
1153 if (iov_iter_rw(iter) == READ && !count)
1154 return 0;
1155
1156 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
1157 if (!dio)
1158 return -ENOMEM;
1159
1160
1161
1162
1163
1164 memset(dio, 0, offsetof(struct dio, pages));
1165
1166 dio->flags = flags;
1167 if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ) {
1168
1169 inode_lock(inode);
1170 }
1171
1172
1173 dio->i_size = i_size_read(inode);
1174 if (iov_iter_rw(iter) == READ && offset >= dio->i_size) {
1175 retval = 0;
1176 goto fail_dio;
1177 }
1178
1179 if (align & blocksize_mask) {
1180 if (bdev)
1181 blkbits = blksize_bits(bdev_logical_block_size(bdev));
1182 blocksize_mask = (1 << blkbits) - 1;
1183 if (align & blocksize_mask)
1184 goto fail_dio;
1185 }
1186
1187 if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ) {
1188 struct address_space *mapping = iocb->ki_filp->f_mapping;
1189
1190 retval = filemap_write_and_wait_range(mapping, offset, end - 1);
1191 if (retval)
1192 goto fail_dio;
1193 }
1194
1195
1196
1197
1198
1199
1200
1201 if (is_sync_kiocb(iocb))
1202 dio->is_async = false;
1203 else if (iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
1204 dio->is_async = false;
1205 else
1206 dio->is_async = true;
1207
1208 dio->inode = inode;
1209 if (iov_iter_rw(iter) == WRITE) {
1210 dio->op = REQ_OP_WRITE;
1211 dio->op_flags = REQ_SYNC | REQ_IDLE;
1212 if (iocb->ki_flags & IOCB_NOWAIT)
1213 dio->op_flags |= REQ_NOWAIT;
1214 } else {
1215 dio->op = REQ_OP_READ;
1216 }
1217 if (iocb->ki_flags & IOCB_HIPRI)
1218 dio->op_flags |= REQ_HIPRI;
1219
1220
1221
1222
1223
1224 if (dio->is_async && iov_iter_rw(iter) == WRITE) {
1225 retval = 0;
1226 if (iocb->ki_flags & IOCB_DSYNC)
1227 retval = dio_set_defer_completion(dio);
1228 else if (!dio->inode->i_sb->s_dio_done_wq) {
1229
1230
1231
1232
1233
1234 retval = sb_init_dio_done_wq(dio->inode->i_sb);
1235 }
1236 if (retval)
1237 goto fail_dio;
1238 }
1239
1240
1241
1242
1243 inode_dio_begin(inode);
1244
1245 retval = 0;
1246 sdio.blkbits = blkbits;
1247 sdio.blkfactor = i_blkbits - blkbits;
1248 sdio.block_in_file = offset >> blkbits;
1249
1250 sdio.get_block = get_block;
1251 dio->end_io = end_io;
1252 sdio.submit_io = submit_io;
1253 sdio.final_block_in_bio = -1;
1254 sdio.next_block_for_io = -1;
1255
1256 dio->iocb = iocb;
1257
1258 spin_lock_init(&dio->bio_lock);
1259 dio->refcount = 1;
1260
1261 dio->should_dirty = iter_is_iovec(iter) && iov_iter_rw(iter) == READ;
1262 sdio.iter = iter;
1263 sdio.final_block_in_request = end >> blkbits;
1264
1265
1266
1267
1268
1269 if (unlikely(sdio.blkfactor))
1270 sdio.pages_in_io = 2;
1271
1272 sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
1273
1274 blk_start_plug(&plug);
1275
1276 retval = do_direct_IO(dio, &sdio, &map_bh);
1277 if (retval)
1278 dio_cleanup(dio, &sdio);
1279
1280 if (retval == -ENOTBLK) {
1281
1282
1283
1284
1285 retval = 0;
1286 }
1287
1288
1289
1290
1291 dio_zero_block(dio, &sdio, 1, &map_bh);
1292
1293 if (sdio.cur_page) {
1294 ssize_t ret2;
1295
1296 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1297 if (retval == 0)
1298 retval = ret2;
1299 put_page(sdio.cur_page);
1300 sdio.cur_page = NULL;
1301 }
1302 if (sdio.bio)
1303 dio_bio_submit(dio, &sdio);
1304
1305 blk_finish_plug(&plug);
1306
1307
1308
1309
1310
1311 dio_cleanup(dio, &sdio);
1312
1313
1314
1315
1316
1317
1318 if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING))
1319 inode_unlock(dio->inode);
1320
1321
1322
1323
1324
1325
1326
1327
1328 BUG_ON(retval == -EIOCBQUEUED);
1329 if (dio->is_async && retval == 0 && dio->result &&
1330 (iov_iter_rw(iter) == READ || dio->result == count))
1331 retval = -EIOCBQUEUED;
1332 else
1333 dio_await_completion(dio);
1334
1335 if (drop_refcount(dio) == 0) {
1336 retval = dio_complete(dio, retval, DIO_COMPLETE_INVALIDATE);
1337 } else
1338 BUG_ON(retval != -EIOCBQUEUED);
1339
1340 return retval;
1341
1342fail_dio:
1343 if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ)
1344 inode_unlock(inode);
1345
1346 kmem_cache_free(dio_cache, dio);
1347 return retval;
1348}
1349
1350ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1351 struct block_device *bdev, struct iov_iter *iter,
1352 get_block_t get_block,
1353 dio_iodone_t end_io, dio_submit_t submit_io,
1354 int flags)
1355{
1356
1357
1358
1359
1360
1361
1362
1363
1364 prefetch(&bdev->bd_disk->part_tbl);
1365 prefetch(bdev->bd_disk->queue);
1366 prefetch((char *)bdev->bd_disk->queue + SMP_CACHE_BYTES);
1367
1368 return do_blockdev_direct_IO(iocb, inode, bdev, iter, get_block,
1369 end_io, submit_io, flags);
1370}
1371
1372EXPORT_SYMBOL(__blockdev_direct_IO);
1373
1374static __init int dio_init(void)
1375{
1376 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1377 return 0;
1378}
1379module_init(dio_init)
1380