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