direct-io.c (fc30eea1542dd787c6aa46e970014e97e390c5b2) direct-io.c (fcb14cb1bdacec5b4374fe161e83fb8208164a85)
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * fs/direct-io.c
4 *
5 * Copyright (C) 2002, Linus Torvalds.
6 *
7 * O_DIRECT
8 *

--- 103 unchanged lines hidden (view full) ---

112 unsigned head; /* next page to process */
113 unsigned tail; /* last valid page + 1 */
114 size_t from, to;
115};
116
117/* dio_state communicated between submission path and end_io */
118struct dio {
119 int flags; /* doesn't change */
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * fs/direct-io.c
4 *
5 * Copyright (C) 2002, Linus Torvalds.
6 *
7 * O_DIRECT
8 *

--- 103 unchanged lines hidden (view full) ---

112 unsigned head; /* next page to process */
113 unsigned tail; /* last valid page + 1 */
114 size_t from, to;
115};
116
117/* dio_state communicated between submission path and end_io */
118struct dio {
119 int flags; /* doesn't change */
120 int op;
121 int op_flags;
120 blk_opf_t opf; /* request operation type and flags */
122 struct gendisk *bio_disk;
123 struct inode *inode;
124 loff_t i_size; /* i_size when submitted */
125 dio_iodone_t *end_io; /* IO completion function */
126
127 void *private; /* copy from map_bh.b_private */
128
129 /* BIO completion state */

--- 32 unchanged lines hidden (view full) ---

162 return sdio->tail - sdio->head;
163}
164
165/*
166 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
167 */
168static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
169{
121 struct gendisk *bio_disk;
122 struct inode *inode;
123 loff_t i_size; /* i_size when submitted */
124 dio_iodone_t *end_io; /* IO completion function */
125
126 void *private; /* copy from map_bh.b_private */
127
128 /* BIO completion state */

--- 32 unchanged lines hidden (view full) ---

161 return sdio->tail - sdio->head;
162}
163
164/*
165 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
166 */
167static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
168{
169 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
170 ssize_t ret;
171
172 ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES,
173 &sdio->from);
174
170 ssize_t ret;
171
172 ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES,
173 &sdio->from);
174
175 if (ret < 0 && sdio->blocks_available && (dio->op == REQ_OP_WRITE)) {
175 if (ret < 0 && sdio->blocks_available && dio_op == REQ_OP_WRITE) {
176 struct page *page = ZERO_PAGE(0);
177 /*
178 * A memory fault, but the filesystem has some outstanding
179 * mapped blocks. We need to use those blocks up to avoid
180 * leaking stale data in the file.
181 */
182 if (dio->page_errors == 0)
183 dio->page_errors = ret;

--- 45 unchanged lines hidden (view full) ---

229 *
230 * It lets the filesystem know if it registered an interest earlier via
231 * get_block. Pass the private field of the map buffer_head so that
232 * filesystems can use it to hold additional state between get_block calls and
233 * dio_complete.
234 */
235static ssize_t dio_complete(struct dio *dio, ssize_t ret, unsigned int flags)
236{
176 struct page *page = ZERO_PAGE(0);
177 /*
178 * A memory fault, but the filesystem has some outstanding
179 * mapped blocks. We need to use those blocks up to avoid
180 * leaking stale data in the file.
181 */
182 if (dio->page_errors == 0)
183 dio->page_errors = ret;

--- 45 unchanged lines hidden (view full) ---

229 *
230 * It lets the filesystem know if it registered an interest earlier via
231 * get_block. Pass the private field of the map buffer_head so that
232 * filesystems can use it to hold additional state between get_block calls and
233 * dio_complete.
234 */
235static ssize_t dio_complete(struct dio *dio, ssize_t ret, unsigned int flags)
236{
237 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
237 loff_t offset = dio->iocb->ki_pos;
238 ssize_t transferred = 0;
239 int err;
240
241 /*
242 * AIO submission can race with bio completion to get here while
243 * expecting to have the last io completed by bio completion.
244 * In that case -EIOCBQUEUED is in fact not an error we want
245 * to preserve through this call.
246 */
247 if (ret == -EIOCBQUEUED)
248 ret = 0;
249
250 if (dio->result) {
251 transferred = dio->result;
252
253 /* Check for short read case */
238 loff_t offset = dio->iocb->ki_pos;
239 ssize_t transferred = 0;
240 int err;
241
242 /*
243 * AIO submission can race with bio completion to get here while
244 * expecting to have the last io completed by bio completion.
245 * In that case -EIOCBQUEUED is in fact not an error we want
246 * to preserve through this call.
247 */
248 if (ret == -EIOCBQUEUED)
249 ret = 0;
250
251 if (dio->result) {
252 transferred = dio->result;
253
254 /* Check for short read case */
254 if ((dio->op == REQ_OP_READ) &&
255 if (dio_op == REQ_OP_READ &&
255 ((offset + transferred) > dio->i_size))
256 transferred = dio->i_size - offset;
257 /* ignore EFAULT if some IO has been done */
258 if (unlikely(ret == -EFAULT) && transferred)
259 ret = 0;
260 }
261
262 if (ret == 0)

--- 18 unchanged lines hidden (view full) ---

281 * this invalidation fails, tough, the write still worked...
282 *
283 * And this page cache invalidation has to be after dio->end_io(), as
284 * some filesystems convert unwritten extents to real allocations in
285 * end_io() when necessary, otherwise a racing buffer read would cache
286 * zeros from unwritten extents.
287 */
288 if (flags & DIO_COMPLETE_INVALIDATE &&
256 ((offset + transferred) > dio->i_size))
257 transferred = dio->i_size - offset;
258 /* ignore EFAULT if some IO has been done */
259 if (unlikely(ret == -EFAULT) && transferred)
260 ret = 0;
261 }
262
263 if (ret == 0)

--- 18 unchanged lines hidden (view full) ---

282 * this invalidation fails, tough, the write still worked...
283 *
284 * And this page cache invalidation has to be after dio->end_io(), as
285 * some filesystems convert unwritten extents to real allocations in
286 * end_io() when necessary, otherwise a racing buffer read would cache
287 * zeros from unwritten extents.
288 */
289 if (flags & DIO_COMPLETE_INVALIDATE &&
289 ret > 0 && dio->op == REQ_OP_WRITE &&
290 ret > 0 && dio_op == REQ_OP_WRITE &&
290 dio->inode->i_mapping->nrpages) {
291 err = invalidate_inode_pages2_range(dio->inode->i_mapping,
292 offset >> PAGE_SHIFT,
293 (offset + ret - 1) >> PAGE_SHIFT);
294 if (err)
295 dio_warn_stale_pagecache(dio->iocb->ki_filp);
296 }
297
298 inode_dio_end(dio->inode);
299
300 if (flags & DIO_COMPLETE_ASYNC) {
301 /*
302 * generic_write_sync expects ki_pos to have been updated
303 * already, but the submission path only does this for
304 * synchronous I/O.
305 */
306 dio->iocb->ki_pos += transferred;
307
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 * generic_write_sync expects ki_pos to have been updated
304 * already, but the submission path only does this for
305 * synchronous I/O.
306 */
307 dio->iocb->ki_pos += transferred;
308
308 if (ret > 0 && dio->op == REQ_OP_WRITE)
309 if (ret > 0 && dio_op == REQ_OP_WRITE)
309 ret = generic_write_sync(dio->iocb, ret);
310 dio->iocb->ki_complete(dio->iocb, ret);
311 }
312
313 kmem_cache_free(dio_cache, dio);
314 return ret;
315}
316

--- 7 unchanged lines hidden (view full) ---

324static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio);
325
326/*
327 * Asynchronous IO callback.
328 */
329static void dio_bio_end_aio(struct bio *bio)
330{
331 struct dio *dio = bio->bi_private;
310 ret = generic_write_sync(dio->iocb, ret);
311 dio->iocb->ki_complete(dio->iocb, ret);
312 }
313
314 kmem_cache_free(dio_cache, dio);
315 return ret;
316}
317

--- 7 unchanged lines hidden (view full) ---

325static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio);
326
327/*
328 * Asynchronous IO callback.
329 */
330static void dio_bio_end_aio(struct bio *bio)
331{
332 struct dio *dio = bio->bi_private;
333 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
332 unsigned long remaining;
333 unsigned long flags;
334 bool defer_completion = false;
335
336 /* cleanup the bio */
337 dio_bio_complete(dio, bio);
338
339 spin_lock_irqsave(&dio->bio_lock, flags);

--- 8 unchanged lines hidden (view full) ---

348 * when the inode has pages mapped and this is AIO write.
349 * We need to invalidate those pages because there is a
350 * chance they contain stale data in the case buffered IO
351 * went in between AIO submission and completion into the
352 * same region.
353 */
354 if (dio->result)
355 defer_completion = dio->defer_completion ||
334 unsigned long remaining;
335 unsigned long flags;
336 bool defer_completion = false;
337
338 /* cleanup the bio */
339 dio_bio_complete(dio, bio);
340
341 spin_lock_irqsave(&dio->bio_lock, flags);

--- 8 unchanged lines hidden (view full) ---

350 * when the inode has pages mapped and this is AIO write.
351 * We need to invalidate those pages because there is a
352 * chance they contain stale data in the case buffered IO
353 * went in between AIO submission and completion into the
354 * same region.
355 */
356 if (dio->result)
357 defer_completion = dio->defer_completion ||
356 (dio->op == REQ_OP_WRITE &&
358 (dio_op == REQ_OP_WRITE &&
357 dio->inode->i_mapping->nrpages);
358 if (defer_completion) {
359 INIT_WORK(&dio->complete_work, dio_aio_complete_work);
360 queue_work(dio->inode->i_sb->s_dio_done_wq,
361 &dio->complete_work);
362 } else {
363 dio_complete(dio, 0, DIO_COMPLETE_ASYNC);
364 }

--- 26 unchanged lines hidden (view full) ---

391 sector_t first_sector, int nr_vecs)
392{
393 struct bio *bio;
394
395 /*
396 * bio_alloc() is guaranteed to return a bio when allowed to sleep and
397 * we request a valid number of vectors.
398 */
359 dio->inode->i_mapping->nrpages);
360 if (defer_completion) {
361 INIT_WORK(&dio->complete_work, dio_aio_complete_work);
362 queue_work(dio->inode->i_sb->s_dio_done_wq,
363 &dio->complete_work);
364 } else {
365 dio_complete(dio, 0, DIO_COMPLETE_ASYNC);
366 }

--- 26 unchanged lines hidden (view full) ---

393 sector_t first_sector, int nr_vecs)
394{
395 struct bio *bio;
396
397 /*
398 * bio_alloc() is guaranteed to return a bio when allowed to sleep and
399 * we request a valid number of vectors.
400 */
399 bio = bio_alloc(bdev, nr_vecs, dio->op | dio->op_flags, GFP_KERNEL);
401 bio = bio_alloc(bdev, nr_vecs, dio->opf, GFP_KERNEL);
400 bio->bi_iter.bi_sector = first_sector;
401 if (dio->is_async)
402 bio->bi_end_io = dio_bio_end_aio;
403 else
404 bio->bi_end_io = dio_bio_end_io;
405 sdio->bio = bio;
406 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
407}
408
409/*
410 * In the AIO read case we speculatively dirty the pages before starting IO.
411 * During IO completion, any of these pages which happen to have been written
412 * back will be redirtied by bio_check_pages_dirty().
413 *
414 * bios hold a dio reference between submit_bio and ->end_io.
415 */
416static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
417{
402 bio->bi_iter.bi_sector = first_sector;
403 if (dio->is_async)
404 bio->bi_end_io = dio_bio_end_aio;
405 else
406 bio->bi_end_io = dio_bio_end_io;
407 sdio->bio = bio;
408 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
409}
410
411/*
412 * In the AIO read case we speculatively dirty the pages before starting IO.
413 * During IO completion, any of these pages which happen to have been written
414 * back will be redirtied by bio_check_pages_dirty().
415 *
416 * bios hold a dio reference between submit_bio and ->end_io.
417 */
418static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
419{
420 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
418 struct bio *bio = sdio->bio;
419 unsigned long flags;
420
421 bio->bi_private = dio;
422 /* don't account direct I/O as memory stall */
423 bio_clear_flag(bio, BIO_WORKINGSET);
424
425 spin_lock_irqsave(&dio->bio_lock, flags);
426 dio->refcount++;
427 spin_unlock_irqrestore(&dio->bio_lock, flags);
428
421 struct bio *bio = sdio->bio;
422 unsigned long flags;
423
424 bio->bi_private = dio;
425 /* don't account direct I/O as memory stall */
426 bio_clear_flag(bio, BIO_WORKINGSET);
427
428 spin_lock_irqsave(&dio->bio_lock, flags);
429 dio->refcount++;
430 spin_unlock_irqrestore(&dio->bio_lock, flags);
431
429 if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty)
432 if (dio->is_async && dio_op == REQ_OP_READ && dio->should_dirty)
430 bio_set_pages_dirty(bio);
431
432 dio->bio_disk = bio->bi_bdev->bd_disk;
433
434 if (sdio->submit_io)
435 sdio->submit_io(bio, dio->inode, sdio->logical_offset_in_bio);
436 else
437 submit_bio(bio);

--- 49 unchanged lines hidden (view full) ---

487}
488
489/*
490 * Process one completed BIO. No locks are held.
491 */
492static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio)
493{
494 blk_status_t err = bio->bi_status;
433 bio_set_pages_dirty(bio);
434
435 dio->bio_disk = bio->bi_bdev->bd_disk;
436
437 if (sdio->submit_io)
438 sdio->submit_io(bio, dio->inode, sdio->logical_offset_in_bio);
439 else
440 submit_bio(bio);

--- 49 unchanged lines hidden (view full) ---

490}
491
492/*
493 * Process one completed BIO. No locks are held.
494 */
495static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio)
496{
497 blk_status_t err = bio->bi_status;
495 bool should_dirty = dio->op == REQ_OP_READ && dio->should_dirty;
498 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
499 bool should_dirty = dio_op == REQ_OP_READ && dio->should_dirty;
496
497 if (err) {
498 if (err == BLK_STS_AGAIN && (bio->bi_opf & REQ_NOWAIT))
499 dio->io_error = -EAGAIN;
500 else
501 dio->io_error = -EIO;
502 }
503

--- 110 unchanged lines hidden (view full) ---

614 * In the case of filesystem holes: the fs may return an arbitrarily-large
615 * hole by returning an appropriate value in b_size and by clearing
616 * buffer_mapped(). However the direct-io code will only process holes one
617 * block at a time - it will repeatedly call get_block() as it walks the hole.
618 */
619static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
620 struct buffer_head *map_bh)
621{
500
501 if (err) {
502 if (err == BLK_STS_AGAIN && (bio->bi_opf & REQ_NOWAIT))
503 dio->io_error = -EAGAIN;
504 else
505 dio->io_error = -EIO;
506 }
507

--- 110 unchanged lines hidden (view full) ---

618 * In the case of filesystem holes: the fs may return an arbitrarily-large
619 * hole by returning an appropriate value in b_size and by clearing
620 * buffer_mapped(). However the direct-io code will only process holes one
621 * block at a time - it will repeatedly call get_block() as it walks the hole.
622 */
623static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
624 struct buffer_head *map_bh)
625{
626 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
622 int ret;
623 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
624 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
625 unsigned long fs_count; /* Number of filesystem-sized blocks */
626 int create;
627 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
628 loff_t i_size;
629

--- 18 unchanged lines hidden (view full) ---

648 * overwrites are permitted. We will return early to the caller
649 * once we see an unmapped buffer head returned, and the caller
650 * will fall back to buffered I/O.
651 *
652 * Otherwise the decision is left to the get_blocks method,
653 * which may decide to handle it or also return an unmapped
654 * buffer head.
655 */
627 int ret;
628 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
629 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
630 unsigned long fs_count; /* Number of filesystem-sized blocks */
631 int create;
632 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
633 loff_t i_size;
634

--- 18 unchanged lines hidden (view full) ---

653 * overwrites are permitted. We will return early to the caller
654 * once we see an unmapped buffer head returned, and the caller
655 * will fall back to buffered I/O.
656 *
657 * Otherwise the decision is left to the get_blocks method,
658 * which may decide to handle it or also return an unmapped
659 * buffer head.
660 */
656 create = dio->op == REQ_OP_WRITE;
661 create = dio_op == REQ_OP_WRITE;
657 if (dio->flags & DIO_SKIP_HOLES) {
658 i_size = i_size_read(dio->inode);
659 if (i_size && fs_startblk <= (i_size - 1) >> i_blkbits)
660 create = 0;
661 }
662
663 ret = (*sdio->get_block)(dio->inode, fs_startblk,
664 map_bh, create);

--- 131 unchanged lines hidden (view full) ---

796 * If that doesn't work out then we put the old page into the bio and add this
797 * page to the dio instead.
798 */
799static inline int
800submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
801 unsigned offset, unsigned len, sector_t blocknr,
802 struct buffer_head *map_bh)
803{
662 if (dio->flags & DIO_SKIP_HOLES) {
663 i_size = i_size_read(dio->inode);
664 if (i_size && fs_startblk <= (i_size - 1) >> i_blkbits)
665 create = 0;
666 }
667
668 ret = (*sdio->get_block)(dio->inode, fs_startblk,
669 map_bh, create);

--- 131 unchanged lines hidden (view full) ---

801 * If that doesn't work out then we put the old page into the bio and add this
802 * page to the dio instead.
803 */
804static inline int
805submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
806 unsigned offset, unsigned len, sector_t blocknr,
807 struct buffer_head *map_bh)
808{
809 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
804 int ret = 0;
805 int boundary = sdio->boundary; /* dio_send_cur_page may clear it */
806
810 int ret = 0;
811 int boundary = sdio->boundary; /* dio_send_cur_page may clear it */
812
807 if (dio->op == REQ_OP_WRITE) {
813 if (dio_op == REQ_OP_WRITE) {
808 /*
809 * Read accounting is performed in submit_bio()
810 */
811 task_io_account_write(len);
812 }
813
814 /*
815 * Can we just grow the current page's presence in the dio?

--- 96 unchanged lines hidden (view full) ---

912 *
913 * For best results, the blockdev should be set up with 512-byte i_blkbits and
914 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
915 * fine alignment but still allows this function to work in PAGE_SIZE units.
916 */
917static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
918 struct buffer_head *map_bh)
919{
814 /*
815 * Read accounting is performed in submit_bio()
816 */
817 task_io_account_write(len);
818 }
819
820 /*
821 * Can we just grow the current page's presence in the dio?

--- 96 unchanged lines hidden (view full) ---

918 *
919 * For best results, the blockdev should be set up with 512-byte i_blkbits and
920 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
921 * fine alignment but still allows this function to work in PAGE_SIZE units.
922 */
923static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
924 struct buffer_head *map_bh)
925{
926 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
920 const unsigned blkbits = sdio->blkbits;
921 const unsigned i_blkbits = blkbits + sdio->blkfactor;
922 int ret = 0;
923
924 while (sdio->block_in_file < sdio->final_block_in_request) {
925 struct page *page;
926 size_t from, to;
927

--- 59 unchanged lines hidden (view full) ---

987 sdio->blocks_available -= dio_remainder;
988 }
989do_holes:
990 /* Handle holes */
991 if (!buffer_mapped(map_bh)) {
992 loff_t i_size_aligned;
993
994 /* AKPM: eargh, -ENOTBLK is a hack */
927 const unsigned blkbits = sdio->blkbits;
928 const unsigned i_blkbits = blkbits + sdio->blkfactor;
929 int ret = 0;
930
931 while (sdio->block_in_file < sdio->final_block_in_request) {
932 struct page *page;
933 size_t from, to;
934

--- 59 unchanged lines hidden (view full) ---

994 sdio->blocks_available -= dio_remainder;
995 }
996do_holes:
997 /* Handle holes */
998 if (!buffer_mapped(map_bh)) {
999 loff_t i_size_aligned;
1000
1001 /* AKPM: eargh, -ENOTBLK is a hack */
995 if (dio->op == REQ_OP_WRITE) {
1002 if (dio_op == REQ_OP_WRITE) {
996 put_page(page);
997 return -ENOTBLK;
998 }
999
1000 /*
1001 * Be sure to account for a partial block as the
1002 * last block in the file
1003 */

--- 187 unchanged lines hidden (view full) ---

1191 dio->is_async = false;
1192 else if (iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
1193 dio->is_async = false;
1194 else
1195 dio->is_async = true;
1196
1197 dio->inode = inode;
1198 if (iov_iter_rw(iter) == WRITE) {
1003 put_page(page);
1004 return -ENOTBLK;
1005 }
1006
1007 /*
1008 * Be sure to account for a partial block as the
1009 * last block in the file
1010 */

--- 187 unchanged lines hidden (view full) ---

1198 dio->is_async = false;
1199 else if (iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
1200 dio->is_async = false;
1201 else
1202 dio->is_async = true;
1203
1204 dio->inode = inode;
1205 if (iov_iter_rw(iter) == WRITE) {
1199 dio->op = REQ_OP_WRITE;
1200 dio->op_flags = REQ_SYNC | REQ_IDLE;
1206 dio->opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
1201 if (iocb->ki_flags & IOCB_NOWAIT)
1207 if (iocb->ki_flags & IOCB_NOWAIT)
1202 dio->op_flags |= REQ_NOWAIT;
1208 dio->opf |= REQ_NOWAIT;
1203 } else {
1209 } else {
1204 dio->op = REQ_OP_READ;
1210 dio->opf = REQ_OP_READ;
1205 }
1206
1207 /*
1208 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1209 * so that we can call ->fsync.
1210 */
1211 if (dio->is_async && iov_iter_rw(iter) == WRITE) {
1212 retval = 0;
1211 }
1212
1213 /*
1214 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1215 * so that we can call ->fsync.
1216 */
1217 if (dio->is_async && iov_iter_rw(iter) == WRITE) {
1218 retval = 0;
1213 if (iocb->ki_flags & IOCB_DSYNC)
1219 if (iocb_is_dsync(iocb))
1214 retval = dio_set_defer_completion(dio);
1215 else if (!dio->inode->i_sb->s_dio_done_wq) {
1216 /*
1217 * In case of AIO write racing with buffered read we
1218 * need to defer completion. We can't decide this now,
1219 * however the workqueue needs to be initialized here.
1220 */
1221 retval = sb_init_dio_done_wq(dio->inode->i_sb);

--- 18 unchanged lines hidden (view full) ---

1240 sdio.final_block_in_bio = -1;
1241 sdio.next_block_for_io = -1;
1242
1243 dio->iocb = iocb;
1244
1245 spin_lock_init(&dio->bio_lock);
1246 dio->refcount = 1;
1247
1220 retval = dio_set_defer_completion(dio);
1221 else if (!dio->inode->i_sb->s_dio_done_wq) {
1222 /*
1223 * In case of AIO write racing with buffered read we
1224 * need to defer completion. We can't decide this now,
1225 * however the workqueue needs to be initialized here.
1226 */
1227 retval = sb_init_dio_done_wq(dio->inode->i_sb);

--- 18 unchanged lines hidden (view full) ---

1246 sdio.final_block_in_bio = -1;
1247 sdio.next_block_for_io = -1;
1248
1249 dio->iocb = iocb;
1250
1251 spin_lock_init(&dio->bio_lock);
1252 dio->refcount = 1;
1253
1248 dio->should_dirty = iter_is_iovec(iter) && iov_iter_rw(iter) == READ;
1254 dio->should_dirty = user_backed_iter(iter) && iov_iter_rw(iter) == READ;
1249 sdio.iter = iter;
1250 sdio.final_block_in_request = end >> blkbits;
1251
1252 /*
1253 * In case of non-aligned buffers, we may need 2 more
1254 * pages since we need to zero out first and last block.
1255 */
1256 if (unlikely(sdio.blkfactor))

--- 87 unchanged lines hidden --- 		1255 sdio.iter = iter;
1256 sdio.final_block_in_request = end >> blkbits;
1257
1258 /*
1259 * In case of non-aligned buffers, we may need 2 more
1260 * pages since we need to zero out first and last block.
1261 */
1262 if (unlikely(sdio.blkfactor))

--- 87 unchanged lines hidden ---