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