4 * Copyright (C) 2002, Linus Torvalds.
8 * 04Jul2002 Andrew Morton
10 * 11Sep2002 janetinc@us.ibm.com
11 * added readv/writev support.
12 * 29Oct2002 Andrew Morton
13 * rewrote bio_add_page() support.
14 * 30Oct2002 pbadari@us.ibm.com
15 * added support for non-aligned IO.
16 * 06Nov2002 pbadari@us.ibm.com
17 * added asynchronous IO support.
18 * 21Jul2003 nathans@sgi.com
19 * added IO completion notifier.
22 #include <linux/kernel.h>
23 #include <linux/module.h>
24 #include <linux/types.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>
42 * How many user pages to map in one call to get_user_pages(). This determines
43 * the size of a structure in the slab cache
48 * Flags for dio_complete()
50 #define DIO_COMPLETE_ASYNC 0x01 /* This is async IO */
51 #define DIO_COMPLETE_INVALIDATE 0x02 /* Can invalidate pages */
54 * This code generally works in units of "dio_blocks". A dio_block is
55 * somewhere between the hard sector size and the filesystem block size. it
56 * is determined on a per-invocation basis. When talking to the filesystem
57 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
58 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
59 * to bio_block quantities by shifting left by blkfactor.
61 * If blkfactor is zero then the user's request was aligned to the filesystem's
65 /* dio_state only used in the submission path */
68 struct bio
*bio
; /* bio under assembly */
69 unsigned blkbits
; /* doesn't change */
70 unsigned blkfactor
; /* When we're using an alignment which
71 is finer than the filesystem's soft
72 blocksize, this specifies how much
73 finer. blkfactor=2 means 1/4-block
74 alignment. Does not change */
75 unsigned start_zero_done
; /* flag: sub-blocksize zeroing has
76 been performed at the start of a
78 int pages_in_io
; /* approximate total IO pages */
79 sector_t block_in_file
; /* Current offset into the underlying
80 file in dio_block units. */
81 unsigned blocks_available
; /* At block_in_file. changes */
82 int reap_counter
; /* rate limit reaping */
83 sector_t final_block_in_request
;/* doesn't change */
84 int boundary
; /* prev block is at a boundary */
85 get_block_t
*get_block
; /* block mapping function */
86 dio_submit_t
*submit_io
; /* IO submition function */
88 loff_t logical_offset_in_bio
; /* current first logical block in bio */
89 sector_t final_block_in_bio
; /* current final block in bio + 1 */
90 sector_t next_block_for_io
; /* next block to be put under IO,
91 in dio_blocks units */
94 * Deferred addition of a page to the dio. These variables are
95 * private to dio_send_cur_page(), submit_page_section() and
98 struct page
*cur_page
; /* The page */
99 unsigned cur_page_offset
; /* Offset into it, in bytes */
100 unsigned cur_page_len
; /* Nr of bytes at cur_page_offset */
101 sector_t cur_page_block
; /* Where it starts */
102 loff_t cur_page_fs_offset
; /* Offset in file */
104 struct iov_iter
*iter
;
106 * Page queue. These variables belong to dio_refill_pages() and
109 unsigned head
; /* next page to process */
110 unsigned tail
; /* last valid page + 1 */
114 /* dio_state communicated between submission path and end_io */
116 int flags
; /* doesn't change */
120 struct gendisk
*bio_disk
;
122 loff_t i_size
; /* i_size when submitted */
123 dio_iodone_t
*end_io
; /* IO completion function */
125 void *private; /* copy from map_bh.b_private */
127 /* BIO completion state */
128 spinlock_t bio_lock
; /* protects BIO fields below */
129 int page_errors
; /* errno from get_user_pages() */
130 int is_async
; /* is IO async ? */
131 bool defer_completion
; /* defer AIO completion to workqueue? */
132 bool should_dirty
; /* if pages should be dirtied */
133 int io_error
; /* IO error in completion path */
134 unsigned long refcount
; /* direct_io_worker() and bios */
135 struct bio
*bio_list
; /* singly linked via bi_private */
136 struct task_struct
*waiter
; /* waiting task (NULL if none) */
138 /* AIO related stuff */
139 struct kiocb
*iocb
; /* kiocb */
140 ssize_t result
; /* IO result */
143 * pages[] (and any fields placed after it) are not zeroed out at
144 * allocation time. Don't add new fields after pages[] unless you
145 * wish that they not be zeroed.
148 struct page
*pages
[DIO_PAGES
]; /* page buffer */
149 struct work_struct complete_work
;/* deferred AIO completion */
151 } ____cacheline_aligned_in_smp
;
153 static struct kmem_cache
*dio_cache __read_mostly
;
156 * How many pages are in the queue?
158 static inline unsigned dio_pages_present(struct dio_submit
*sdio
)
160 return sdio
->tail
- sdio
->head
;
164 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
166 static inline int dio_refill_pages(struct dio
*dio
, struct dio_submit
*sdio
)
170 ret
= iov_iter_get_pages(sdio
->iter
, dio
->pages
, LONG_MAX
, DIO_PAGES
,
173 if (ret
< 0 && sdio
->blocks_available
&& (dio
->op
== REQ_OP_WRITE
)) {
174 struct page
*page
= ZERO_PAGE(0);
176 * A memory fault, but the filesystem has some outstanding
177 * mapped blocks. We need to use those blocks up to avoid
178 * leaking stale data in the file.
180 if (dio
->page_errors
== 0)
181 dio
->page_errors
= ret
;
183 dio
->pages
[0] = page
;
187 sdio
->to
= PAGE_SIZE
;
192 iov_iter_advance(sdio
->iter
, ret
);
195 sdio
->tail
= (ret
+ PAGE_SIZE
- 1) / PAGE_SIZE
;
196 sdio
->to
= ((ret
- 1) & (PAGE_SIZE
- 1)) + 1;
203 * Get another userspace page. Returns an ERR_PTR on error. Pages are
204 * buffered inside the dio so that we can call get_user_pages() against a
205 * decent number of pages, less frequently. To provide nicer use of the
208 static inline struct page
*dio_get_page(struct dio
*dio
,
209 struct dio_submit
*sdio
)
211 if (dio_pages_present(sdio
) == 0) {
214 ret
= dio_refill_pages(dio
, sdio
);
217 BUG_ON(dio_pages_present(sdio
) == 0);
219 return dio
->pages
[sdio
->head
];
223 * dio_complete() - called when all DIO BIO I/O has been completed
224 * @offset: the byte offset in the file of the completed operation
226 * This drops i_dio_count, lets interested parties know that a DIO operation
227 * has completed, and calculates the resulting return code for the operation.
229 * It lets the filesystem know if it registered an interest earlier via
230 * get_block. Pass the private field of the map buffer_head so that
231 * filesystems can use it to hold additional state between get_block calls and
234 static ssize_t
dio_complete(struct dio
*dio
, ssize_t ret
, unsigned int flags
)
236 loff_t offset
= dio
->iocb
->ki_pos
;
237 ssize_t transferred
= 0;
241 * AIO submission can race with bio completion to get here while
242 * expecting to have the last io completed by bio completion.
243 * In that case -EIOCBQUEUED is in fact not an error we want
244 * to preserve through this call.
246 if (ret
== -EIOCBQUEUED
)
250 transferred
= dio
->result
;
252 /* Check for short read case */
253 if ((dio
->op
== REQ_OP_READ
) &&
254 ((offset
+ transferred
) > dio
->i_size
))
255 transferred
= dio
->i_size
- offset
;
256 /* ignore EFAULT if some IO has been done */
257 if (unlikely(ret
== -EFAULT
) && transferred
)
262 ret
= dio
->page_errors
;
270 err
= dio
->end_io(dio
->iocb
, offset
, ret
, dio
->private);
276 * Try again to invalidate clean pages which might have been cached by
277 * non-direct readahead, or faulted in by get_user_pages() if the source
278 * of the write was an mmap'ed region of the file we're writing. Either
279 * one is a pretty crazy thing to do, so we don't support it 100%. If
280 * this invalidation fails, tough, the write still worked...
282 * And this page cache invalidation has to be after dio->end_io(), as
283 * some filesystems convert unwritten extents to real allocations in
284 * end_io() when necessary, otherwise a racing buffer read would cache
285 * zeros from unwritten extents.
287 if (flags
& DIO_COMPLETE_INVALIDATE
&&
288 ret
> 0 && dio
->op
== REQ_OP_WRITE
&&
289 dio
->inode
->i_mapping
->nrpages
) {
290 err
= invalidate_inode_pages2_range(dio
->inode
->i_mapping
,
291 offset
>> PAGE_SHIFT
,
292 (offset
+ ret
- 1) >> PAGE_SHIFT
);
296 if (!(dio
->flags
& DIO_SKIP_DIO_COUNT
))
297 inode_dio_end(dio
->inode
);
299 if (flags
& DIO_COMPLETE_ASYNC
) {
301 * generic_write_sync expects ki_pos to have been updated
302 * already, but the submission path only does this for
305 dio
->iocb
->ki_pos
+= transferred
;
307 if (ret
> 0 && dio
->op
== REQ_OP_WRITE
)
308 ret
= generic_write_sync(dio
->iocb
, ret
);
309 dio
->iocb
->ki_complete(dio
->iocb
, ret
, 0);
312 kmem_cache_free(dio_cache
, dio
);
316 static void dio_aio_complete_work(struct work_struct
*work
)
318 struct dio
*dio
= container_of(work
, struct dio
, complete_work
);
320 dio_complete(dio
, 0, DIO_COMPLETE_ASYNC
| DIO_COMPLETE_INVALIDATE
);
323 static blk_status_t
dio_bio_complete(struct dio
*dio
, struct bio
*bio
);
326 * Asynchronous IO callback.
328 static void dio_bio_end_aio(struct bio
*bio
)
330 struct dio
*dio
= bio
->bi_private
;
331 unsigned long remaining
;
333 bool defer_completion
= false;
335 /* cleanup the bio */
336 dio_bio_complete(dio
, bio
);
338 spin_lock_irqsave(&dio
->bio_lock
, flags
);
339 remaining
= --dio
->refcount
;
340 if (remaining
== 1 && dio
->waiter
)
341 wake_up_process(dio
->waiter
);
342 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
344 if (remaining
== 0) {
346 * Defer completion when defer_completion is set or
347 * when the inode has pages mapped and this is AIO write.
348 * We need to invalidate those pages because there is a
349 * chance they contain stale data in the case buffered IO
350 * went in between AIO submission and completion into the
354 defer_completion
= dio
->defer_completion
||
355 (dio
->op
== REQ_OP_WRITE
&&
356 dio
->inode
->i_mapping
->nrpages
);
357 if (defer_completion
) {
358 INIT_WORK(&dio
->complete_work
, dio_aio_complete_work
);
359 queue_work(dio
->inode
->i_sb
->s_dio_done_wq
,
360 &dio
->complete_work
);
362 dio_complete(dio
, 0, DIO_COMPLETE_ASYNC
);
368 * The BIO completion handler simply queues the BIO up for the process-context
371 * During I/O bi_private points at the dio. After I/O, bi_private is used to
372 * implement a singly-linked list of completed BIOs, at dio->bio_list.
374 static void dio_bio_end_io(struct bio
*bio
)
376 struct dio
*dio
= bio
->bi_private
;
379 spin_lock_irqsave(&dio
->bio_lock
, flags
);
380 bio
->bi_private
= dio
->bio_list
;
382 if (--dio
->refcount
== 1 && dio
->waiter
)
383 wake_up_process(dio
->waiter
);
384 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
388 * dio_end_io - handle the end io action for the given bio
389 * @bio: The direct io bio thats being completed
391 * This is meant to be called by any filesystem that uses their own dio_submit_t
392 * so that the DIO specific endio actions are dealt with after the filesystem
393 * has done it's completion work.
395 void dio_end_io(struct bio
*bio
)
397 struct dio
*dio
= bio
->bi_private
;
400 dio_bio_end_aio(bio
);
404 EXPORT_SYMBOL_GPL(dio_end_io
);
407 dio_bio_alloc(struct dio
*dio
, struct dio_submit
*sdio
,
408 struct block_device
*bdev
,
409 sector_t first_sector
, int nr_vecs
)
414 * bio_alloc() is guaranteed to return a bio when called with
415 * __GFP_RECLAIM and we request a valid number of vectors.
417 bio
= bio_alloc(GFP_KERNEL
, nr_vecs
);
419 bio_set_dev(bio
, bdev
);
420 bio
->bi_iter
.bi_sector
= first_sector
;
421 bio_set_op_attrs(bio
, dio
->op
, dio
->op_flags
);
423 bio
->bi_end_io
= dio_bio_end_aio
;
425 bio
->bi_end_io
= dio_bio_end_io
;
427 bio
->bi_write_hint
= dio
->iocb
->ki_hint
;
430 sdio
->logical_offset_in_bio
= sdio
->cur_page_fs_offset
;
434 * In the AIO read case we speculatively dirty the pages before starting IO.
435 * During IO completion, any of these pages which happen to have been written
436 * back will be redirtied by bio_check_pages_dirty().
438 * bios hold a dio reference between submit_bio and ->end_io.
440 static inline void dio_bio_submit(struct dio
*dio
, struct dio_submit
*sdio
)
442 struct bio
*bio
= sdio
->bio
;
445 bio
->bi_private
= dio
;
447 spin_lock_irqsave(&dio
->bio_lock
, flags
);
449 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
451 if (dio
->is_async
&& dio
->op
== REQ_OP_READ
&& dio
->should_dirty
)
452 bio_set_pages_dirty(bio
);
454 dio
->bio_disk
= bio
->bi_disk
;
456 if (sdio
->submit_io
) {
457 sdio
->submit_io(bio
, dio
->inode
, sdio
->logical_offset_in_bio
);
458 dio
->bio_cookie
= BLK_QC_T_NONE
;
460 dio
->bio_cookie
= submit_bio(bio
);
464 sdio
->logical_offset_in_bio
= 0;
468 * Release any resources in case of a failure
470 static inline void dio_cleanup(struct dio
*dio
, struct dio_submit
*sdio
)
472 while (sdio
->head
< sdio
->tail
)
473 put_page(dio
->pages
[sdio
->head
++]);
477 * Wait for the next BIO to complete. Remove it and return it. NULL is
478 * returned once all BIOs have been completed. This must only be called once
479 * all bios have been issued so that dio->refcount can only decrease. This
480 * requires that that the caller hold a reference on the dio.
482 static struct bio
*dio_await_one(struct dio
*dio
)
485 struct bio
*bio
= NULL
;
487 spin_lock_irqsave(&dio
->bio_lock
, flags
);
490 * Wait as long as the list is empty and there are bios in flight. bio
491 * completion drops the count, maybe adds to the list, and wakes while
492 * holding the bio_lock so we don't need set_current_state()'s barrier
493 * and can call it after testing our condition.
495 while (dio
->refcount
> 1 && dio
->bio_list
== NULL
) {
496 __set_current_state(TASK_UNINTERRUPTIBLE
);
497 dio
->waiter
= current
;
498 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
499 if (!(dio
->iocb
->ki_flags
& IOCB_HIPRI
) ||
500 !blk_mq_poll(dio
->bio_disk
->queue
, dio
->bio_cookie
))
502 /* wake up sets us TASK_RUNNING */
503 spin_lock_irqsave(&dio
->bio_lock
, flags
);
508 dio
->bio_list
= bio
->bi_private
;
510 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
515 * Process one completed BIO. No locks are held.
517 static blk_status_t
dio_bio_complete(struct dio
*dio
, struct bio
*bio
)
519 struct bio_vec
*bvec
;
521 blk_status_t err
= bio
->bi_status
;
524 if (err
== BLK_STS_AGAIN
&& (bio
->bi_opf
& REQ_NOWAIT
))
525 dio
->io_error
= -EAGAIN
;
527 dio
->io_error
= -EIO
;
530 if (dio
->is_async
&& dio
->op
== REQ_OP_READ
&& dio
->should_dirty
) {
531 bio_check_pages_dirty(bio
); /* transfers ownership */
533 bio_for_each_segment_all(bvec
, bio
, i
) {
534 struct page
*page
= bvec
->bv_page
;
536 if (dio
->op
== REQ_OP_READ
&& !PageCompound(page
) &&
538 set_page_dirty_lock(page
);
547 * Wait on and process all in-flight BIOs. This must only be called once
548 * all bios have been issued so that the refcount can only decrease.
549 * This just waits for all bios to make it through dio_bio_complete. IO
550 * errors are propagated through dio->io_error and should be propagated via
553 static void dio_await_completion(struct dio
*dio
)
557 bio
= dio_await_one(dio
);
559 dio_bio_complete(dio
, bio
);
564 * A really large O_DIRECT read or write can generate a lot of BIOs. So
565 * to keep the memory consumption sane we periodically reap any completed BIOs
566 * during the BIO generation phase.
568 * This also helps to limit the peak amount of pinned userspace memory.
570 static inline int dio_bio_reap(struct dio
*dio
, struct dio_submit
*sdio
)
574 if (sdio
->reap_counter
++ >= 64) {
575 while (dio
->bio_list
) {
580 spin_lock_irqsave(&dio
->bio_lock
, flags
);
582 dio
->bio_list
= bio
->bi_private
;
583 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
584 ret2
= blk_status_to_errno(dio_bio_complete(dio
, bio
));
588 sdio
->reap_counter
= 0;
594 * Create workqueue for deferred direct IO completions. We allocate the
595 * workqueue when it's first needed. This avoids creating workqueue for
596 * filesystems that don't need it and also allows us to create the workqueue
597 * late enough so the we can include s_id in the name of the workqueue.
599 int sb_init_dio_done_wq(struct super_block
*sb
)
601 struct workqueue_struct
*old
;
602 struct workqueue_struct
*wq
= alloc_workqueue("dio/%s",
608 * This has to be atomic as more DIOs can race to create the workqueue
610 old
= cmpxchg(&sb
->s_dio_done_wq
, NULL
, wq
);
611 /* Someone created workqueue before us? Free ours... */
613 destroy_workqueue(wq
);
617 static int dio_set_defer_completion(struct dio
*dio
)
619 struct super_block
*sb
= dio
->inode
->i_sb
;
621 if (dio
->defer_completion
)
623 dio
->defer_completion
= true;
624 if (!sb
->s_dio_done_wq
)
625 return sb_init_dio_done_wq(sb
);
630 * Call into the fs to map some more disk blocks. We record the current number
631 * of available blocks at sdio->blocks_available. These are in units of the
632 * fs blocksize, i_blocksize(inode).
634 * The fs is allowed to map lots of blocks at once. If it wants to do that,
635 * it uses the passed inode-relative block number as the file offset, as usual.
637 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
638 * has remaining to do. The fs should not map more than this number of blocks.
640 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
641 * indicate how much contiguous disk space has been made available at
644 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
645 * This isn't very efficient...
647 * In the case of filesystem holes: the fs may return an arbitrarily-large
648 * hole by returning an appropriate value in b_size and by clearing
649 * buffer_mapped(). However the direct-io code will only process holes one
650 * block at a time - it will repeatedly call get_block() as it walks the hole.
652 static int get_more_blocks(struct dio
*dio
, struct dio_submit
*sdio
,
653 struct buffer_head
*map_bh
)
656 sector_t fs_startblk
; /* Into file, in filesystem-sized blocks */
657 sector_t fs_endblk
; /* Into file, in filesystem-sized blocks */
658 unsigned long fs_count
; /* Number of filesystem-sized blocks */
660 unsigned int i_blkbits
= sdio
->blkbits
+ sdio
->blkfactor
;
664 * If there was a memory error and we've overwritten all the
665 * mapped blocks then we can now return that memory error
667 ret
= dio
->page_errors
;
669 BUG_ON(sdio
->block_in_file
>= sdio
->final_block_in_request
);
670 fs_startblk
= sdio
->block_in_file
>> sdio
->blkfactor
;
671 fs_endblk
= (sdio
->final_block_in_request
- 1) >>
673 fs_count
= fs_endblk
- fs_startblk
+ 1;
676 map_bh
->b_size
= fs_count
<< i_blkbits
;
679 * For writes that could fill holes inside i_size on a
680 * DIO_SKIP_HOLES filesystem we forbid block creations: only
681 * overwrites are permitted. We will return early to the caller
682 * once we see an unmapped buffer head returned, and the caller
683 * will fall back to buffered I/O.
685 * Otherwise the decision is left to the get_blocks method,
686 * which may decide to handle it or also return an unmapped
689 create
= dio
->op
== REQ_OP_WRITE
;
690 if (dio
->flags
& DIO_SKIP_HOLES
) {
691 i_size
= i_size_read(dio
->inode
);
692 if (i_size
&& fs_startblk
<= (i_size
- 1) >> i_blkbits
)
696 ret
= (*sdio
->get_block
)(dio
->inode
, fs_startblk
,
699 /* Store for completion */
700 dio
->private = map_bh
->b_private
;
702 if (ret
== 0 && buffer_defer_completion(map_bh
))
703 ret
= dio_set_defer_completion(dio
);
709 * There is no bio. Make one now.
711 static inline int dio_new_bio(struct dio
*dio
, struct dio_submit
*sdio
,
712 sector_t start_sector
, struct buffer_head
*map_bh
)
717 ret
= dio_bio_reap(dio
, sdio
);
720 sector
= start_sector
<< (sdio
->blkbits
- 9);
721 nr_pages
= min(sdio
->pages_in_io
, BIO_MAX_PAGES
);
722 BUG_ON(nr_pages
<= 0);
723 dio_bio_alloc(dio
, sdio
, map_bh
->b_bdev
, sector
, nr_pages
);
730 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
731 * that was successful then update final_block_in_bio and take a ref against
732 * the just-added page.
734 * Return zero on success. Non-zero means the caller needs to start a new BIO.
736 static inline int dio_bio_add_page(struct dio_submit
*sdio
)
740 ret
= bio_add_page(sdio
->bio
, sdio
->cur_page
,
741 sdio
->cur_page_len
, sdio
->cur_page_offset
);
742 if (ret
== sdio
->cur_page_len
) {
744 * Decrement count only, if we are done with this page
746 if ((sdio
->cur_page_len
+ sdio
->cur_page_offset
) == PAGE_SIZE
)
748 get_page(sdio
->cur_page
);
749 sdio
->final_block_in_bio
= sdio
->cur_page_block
+
750 (sdio
->cur_page_len
>> sdio
->blkbits
);
759 * Put cur_page under IO. The section of cur_page which is described by
760 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
761 * starts on-disk at cur_page_block.
763 * We take a ref against the page here (on behalf of its presence in the bio).
765 * The caller of this function is responsible for removing cur_page from the
766 * dio, and for dropping the refcount which came from that presence.
768 static inline int dio_send_cur_page(struct dio
*dio
, struct dio_submit
*sdio
,
769 struct buffer_head
*map_bh
)
774 loff_t cur_offset
= sdio
->cur_page_fs_offset
;
775 loff_t bio_next_offset
= sdio
->logical_offset_in_bio
+
776 sdio
->bio
->bi_iter
.bi_size
;
779 * See whether this new request is contiguous with the old.
781 * Btrfs cannot handle having logically non-contiguous requests
782 * submitted. For example if you have
784 * Logical: [0-4095][HOLE][8192-12287]
785 * Physical: [0-4095] [4096-8191]
787 * We cannot submit those pages together as one BIO. So if our
788 * current logical offset in the file does not equal what would
789 * be the next logical offset in the bio, submit the bio we
792 if (sdio
->final_block_in_bio
!= sdio
->cur_page_block
||
793 cur_offset
!= bio_next_offset
)
794 dio_bio_submit(dio
, sdio
);
797 if (sdio
->bio
== NULL
) {
798 ret
= dio_new_bio(dio
, sdio
, sdio
->cur_page_block
, map_bh
);
803 if (dio_bio_add_page(sdio
) != 0) {
804 dio_bio_submit(dio
, sdio
);
805 ret
= dio_new_bio(dio
, sdio
, sdio
->cur_page_block
, map_bh
);
807 ret
= dio_bio_add_page(sdio
);
816 * An autonomous function to put a chunk of a page under deferred IO.
818 * The caller doesn't actually know (or care) whether this piece of page is in
819 * a BIO, or is under IO or whatever. We just take care of all possible
820 * situations here. The separation between the logic of do_direct_IO() and
821 * that of submit_page_section() is important for clarity. Please don't break.
823 * The chunk of page starts on-disk at blocknr.
825 * We perform deferred IO, by recording the last-submitted page inside our
826 * private part of the dio structure. If possible, we just expand the IO
827 * across that page here.
829 * If that doesn't work out then we put the old page into the bio and add this
830 * page to the dio instead.
833 submit_page_section(struct dio
*dio
, struct dio_submit
*sdio
, struct page
*page
,
834 unsigned offset
, unsigned len
, sector_t blocknr
,
835 struct buffer_head
*map_bh
)
839 if (dio
->op
== REQ_OP_WRITE
) {
841 * Read accounting is performed in submit_bio()
843 task_io_account_write(len
);
847 * Can we just grow the current page's presence in the dio?
849 if (sdio
->cur_page
== page
&&
850 sdio
->cur_page_offset
+ sdio
->cur_page_len
== offset
&&
851 sdio
->cur_page_block
+
852 (sdio
->cur_page_len
>> sdio
->blkbits
) == blocknr
) {
853 sdio
->cur_page_len
+= len
;
858 * If there's a deferred page already there then send it.
860 if (sdio
->cur_page
) {
861 ret
= dio_send_cur_page(dio
, sdio
, map_bh
);
862 put_page(sdio
->cur_page
);
863 sdio
->cur_page
= NULL
;
868 get_page(page
); /* It is in dio */
869 sdio
->cur_page
= page
;
870 sdio
->cur_page_offset
= offset
;
871 sdio
->cur_page_len
= len
;
872 sdio
->cur_page_block
= blocknr
;
873 sdio
->cur_page_fs_offset
= sdio
->block_in_file
<< sdio
->blkbits
;
876 * If sdio->boundary then we want to schedule the IO now to
877 * avoid metadata seeks.
879 if (sdio
->boundary
) {
880 ret
= dio_send_cur_page(dio
, sdio
, map_bh
);
882 dio_bio_submit(dio
, sdio
);
883 put_page(sdio
->cur_page
);
884 sdio
->cur_page
= NULL
;
890 * If we are not writing the entire block and get_block() allocated
891 * the block for us, we need to fill-in the unused portion of the
892 * block with zeros. This happens only if user-buffer, fileoffset or
893 * io length is not filesystem block-size multiple.
895 * `end' is zero if we're doing the start of the IO, 1 at the end of the
898 static inline void dio_zero_block(struct dio
*dio
, struct dio_submit
*sdio
,
899 int end
, struct buffer_head
*map_bh
)
901 unsigned dio_blocks_per_fs_block
;
902 unsigned this_chunk_blocks
; /* In dio_blocks */
903 unsigned this_chunk_bytes
;
906 sdio
->start_zero_done
= 1;
907 if (!sdio
->blkfactor
|| !buffer_new(map_bh
))
910 dio_blocks_per_fs_block
= 1 << sdio
->blkfactor
;
911 this_chunk_blocks
= sdio
->block_in_file
& (dio_blocks_per_fs_block
- 1);
913 if (!this_chunk_blocks
)
917 * We need to zero out part of an fs block. It is either at the
918 * beginning or the end of the fs block.
921 this_chunk_blocks
= dio_blocks_per_fs_block
- this_chunk_blocks
;
923 this_chunk_bytes
= this_chunk_blocks
<< sdio
->blkbits
;
926 if (submit_page_section(dio
, sdio
, page
, 0, this_chunk_bytes
,
927 sdio
->next_block_for_io
, map_bh
))
930 sdio
->next_block_for_io
+= this_chunk_blocks
;
934 * Walk the user pages, and the file, mapping blocks to disk and generating
935 * a sequence of (page,offset,len,block) mappings. These mappings are injected
936 * into submit_page_section(), which takes care of the next stage of submission
938 * Direct IO against a blockdev is different from a file. Because we can
939 * happily perform page-sized but 512-byte aligned IOs. It is important that
940 * blockdev IO be able to have fine alignment and large sizes.
942 * So what we do is to permit the ->get_block function to populate bh.b_size
943 * with the size of IO which is permitted at this offset and this i_blkbits.
945 * For best results, the blockdev should be set up with 512-byte i_blkbits and
946 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
947 * fine alignment but still allows this function to work in PAGE_SIZE units.
949 static int do_direct_IO(struct dio
*dio
, struct dio_submit
*sdio
,
950 struct buffer_head
*map_bh
)
952 const unsigned blkbits
= sdio
->blkbits
;
953 const unsigned i_blkbits
= blkbits
+ sdio
->blkfactor
;
956 while (sdio
->block_in_file
< sdio
->final_block_in_request
) {
960 page
= dio_get_page(dio
, sdio
);
965 from
= sdio
->head
? 0 : sdio
->from
;
966 to
= (sdio
->head
== sdio
->tail
- 1) ? sdio
->to
: PAGE_SIZE
;
970 unsigned this_chunk_bytes
; /* # of bytes mapped */
971 unsigned this_chunk_blocks
; /* # of blocks */
974 if (sdio
->blocks_available
== 0) {
976 * Need to go and map some more disk
978 unsigned long blkmask
;
979 unsigned long dio_remainder
;
981 ret
= get_more_blocks(dio
, sdio
, map_bh
);
986 if (!buffer_mapped(map_bh
))
989 sdio
->blocks_available
=
990 map_bh
->b_size
>> blkbits
;
991 sdio
->next_block_for_io
=
992 map_bh
->b_blocknr
<< sdio
->blkfactor
;
993 if (buffer_new(map_bh
)) {
997 map_bh
->b_size
>> i_blkbits
);
1000 if (!sdio
->blkfactor
)
1003 blkmask
= (1 << sdio
->blkfactor
) - 1;
1004 dio_remainder
= (sdio
->block_in_file
& blkmask
);
1007 * If we are at the start of IO and that IO
1008 * starts partway into a fs-block,
1009 * dio_remainder will be non-zero. If the IO
1010 * is a read then we can simply advance the IO
1011 * cursor to the first block which is to be
1012 * read. But if the IO is a write and the
1013 * block was newly allocated we cannot do that;
1014 * the start of the fs block must be zeroed out
1017 if (!buffer_new(map_bh
))
1018 sdio
->next_block_for_io
+= dio_remainder
;
1019 sdio
->blocks_available
-= dio_remainder
;
1023 if (!buffer_mapped(map_bh
)) {
1024 loff_t i_size_aligned
;
1026 /* AKPM: eargh, -ENOTBLK is a hack */
1027 if (dio
->op
== REQ_OP_WRITE
) {
1033 * Be sure to account for a partial block as the
1034 * last block in the file
1036 i_size_aligned
= ALIGN(i_size_read(dio
->inode
),
1038 if (sdio
->block_in_file
>=
1039 i_size_aligned
>> blkbits
) {
1044 zero_user(page
, from
, 1 << blkbits
);
1045 sdio
->block_in_file
++;
1046 from
+= 1 << blkbits
;
1047 dio
->result
+= 1 << blkbits
;
1052 * If we're performing IO which has an alignment which
1053 * is finer than the underlying fs, go check to see if
1054 * we must zero out the start of this block.
1056 if (unlikely(sdio
->blkfactor
&& !sdio
->start_zero_done
))
1057 dio_zero_block(dio
, sdio
, 0, map_bh
);
1060 * Work out, in this_chunk_blocks, how much disk we
1061 * can add to this page
1063 this_chunk_blocks
= sdio
->blocks_available
;
1064 u
= (to
- from
) >> blkbits
;
1065 if (this_chunk_blocks
> u
)
1066 this_chunk_blocks
= u
;
1067 u
= sdio
->final_block_in_request
- sdio
->block_in_file
;
1068 if (this_chunk_blocks
> u
)
1069 this_chunk_blocks
= u
;
1070 this_chunk_bytes
= this_chunk_blocks
<< blkbits
;
1071 BUG_ON(this_chunk_bytes
== 0);
1073 if (this_chunk_blocks
== sdio
->blocks_available
)
1074 sdio
->boundary
= buffer_boundary(map_bh
);
1075 ret
= submit_page_section(dio
, sdio
, page
,
1078 sdio
->next_block_for_io
,
1084 sdio
->next_block_for_io
+= this_chunk_blocks
;
1086 sdio
->block_in_file
+= this_chunk_blocks
;
1087 from
+= this_chunk_bytes
;
1088 dio
->result
+= this_chunk_bytes
;
1089 sdio
->blocks_available
-= this_chunk_blocks
;
1091 BUG_ON(sdio
->block_in_file
> sdio
->final_block_in_request
);
1092 if (sdio
->block_in_file
== sdio
->final_block_in_request
)
1096 /* Drop the ref which was taken in get_user_pages() */
1103 static inline int drop_refcount(struct dio
*dio
)
1106 unsigned long flags
;
1109 * Sync will always be dropping the final ref and completing the
1110 * operation. AIO can if it was a broken operation described above or
1111 * in fact if all the bios race to complete before we get here. In
1112 * that case dio_complete() translates the EIOCBQUEUED into the proper
1113 * return code that the caller will hand to ->complete().
1115 * This is managed by the bio_lock instead of being an atomic_t so that
1116 * completion paths can drop their ref and use the remaining count to
1117 * decide to wake the submission path atomically.
1119 spin_lock_irqsave(&dio
->bio_lock
, flags
);
1120 ret2
= --dio
->refcount
;
1121 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
1126 * This is a library function for use by filesystem drivers.
1128 * The locking rules are governed by the flags parameter:
1129 * - if the flags value contains DIO_LOCKING we use a fancy locking
1130 * scheme for dumb filesystems.
1131 * For writes this function is called under i_mutex and returns with
1132 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1133 * taken and dropped again before returning.
1134 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1135 * internal locking but rather rely on the filesystem to synchronize
1136 * direct I/O reads/writes versus each other and truncate.
1138 * To help with locking against truncate we incremented the i_dio_count
1139 * counter before starting direct I/O, and decrement it once we are done.
1140 * Truncate can wait for it to reach zero to provide exclusion. It is
1141 * expected that filesystem provide exclusion between new direct I/O
1142 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1143 * but other filesystems need to take care of this on their own.
1145 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1146 * is always inlined. Otherwise gcc is unable to split the structure into
1147 * individual fields and will generate much worse code. This is important
1148 * for the whole file.
1150 static inline ssize_t
1151 do_blockdev_direct_IO(struct kiocb
*iocb
, struct inode
*inode
,
1152 struct block_device
*bdev
, struct iov_iter
*iter
,
1153 get_block_t get_block
, dio_iodone_t end_io
,
1154 dio_submit_t submit_io
, int flags
)
1156 unsigned i_blkbits
= ACCESS_ONCE(inode
->i_blkbits
);
1157 unsigned blkbits
= i_blkbits
;
1158 unsigned blocksize_mask
= (1 << blkbits
) - 1;
1159 ssize_t retval
= -EINVAL
;
1160 size_t count
= iov_iter_count(iter
);
1161 loff_t offset
= iocb
->ki_pos
;
1162 loff_t end
= offset
+ count
;
1164 struct dio_submit sdio
= { 0, };
1165 struct buffer_head map_bh
= { 0, };
1166 struct blk_plug plug
;
1167 unsigned long align
= offset
| iov_iter_alignment(iter
);
1170 * Avoid references to bdev if not absolutely needed to give
1171 * the early prefetch in the caller enough time.
1174 if (align
& blocksize_mask
) {
1176 blkbits
= blksize_bits(bdev_logical_block_size(bdev
));
1177 blocksize_mask
= (1 << blkbits
) - 1;
1178 if (align
& blocksize_mask
)
1182 /* watch out for a 0 len io from a tricksy fs */
1183 if (iov_iter_rw(iter
) == READ
&& !iov_iter_count(iter
))
1186 dio
= kmem_cache_alloc(dio_cache
, GFP_KERNEL
);
1191 * Believe it or not, zeroing out the page array caused a .5%
1192 * performance regression in a database benchmark. So, we take
1193 * care to only zero out what's needed.
1195 memset(dio
, 0, offsetof(struct dio
, pages
));
1198 if (dio
->flags
& DIO_LOCKING
) {
1199 if (iov_iter_rw(iter
) == READ
) {
1200 struct address_space
*mapping
=
1201 iocb
->ki_filp
->f_mapping
;
1203 /* will be released by direct_io_worker */
1206 retval
= filemap_write_and_wait_range(mapping
, offset
,
1209 inode_unlock(inode
);
1210 kmem_cache_free(dio_cache
, dio
);
1216 /* Once we sampled i_size check for reads beyond EOF */
1217 dio
->i_size
= i_size_read(inode
);
1218 if (iov_iter_rw(iter
) == READ
&& offset
>= dio
->i_size
) {
1219 if (dio
->flags
& DIO_LOCKING
)
1220 inode_unlock(inode
);
1221 kmem_cache_free(dio_cache
, dio
);
1227 * For file extending writes updating i_size before data writeouts
1228 * complete can expose uninitialized blocks in dumb filesystems.
1229 * In that case we need to wait for I/O completion even if asked
1230 * for an asynchronous write.
1232 if (is_sync_kiocb(iocb
))
1233 dio
->is_async
= false;
1234 else if (!(dio
->flags
& DIO_ASYNC_EXTEND
) &&
1235 iov_iter_rw(iter
) == WRITE
&& end
> i_size_read(inode
))
1236 dio
->is_async
= false;
1238 dio
->is_async
= true;
1241 if (iov_iter_rw(iter
) == WRITE
) {
1242 dio
->op
= REQ_OP_WRITE
;
1243 dio
->op_flags
= REQ_SYNC
| REQ_IDLE
;
1244 if (iocb
->ki_flags
& IOCB_NOWAIT
)
1245 dio
->op_flags
|= REQ_NOWAIT
;
1247 dio
->op
= REQ_OP_READ
;
1251 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1252 * so that we can call ->fsync.
1254 if (dio
->is_async
&& iov_iter_rw(iter
) == WRITE
) {
1256 if (iocb
->ki_flags
& IOCB_DSYNC
)
1257 retval
= dio_set_defer_completion(dio
);
1258 else if (!dio
->inode
->i_sb
->s_dio_done_wq
) {
1260 * In case of AIO write racing with buffered read we
1261 * need to defer completion. We can't decide this now,
1262 * however the workqueue needs to be initialized here.
1264 retval
= sb_init_dio_done_wq(dio
->inode
->i_sb
);
1268 * We grab i_mutex only for reads so we don't have
1269 * to release it here
1271 kmem_cache_free(dio_cache
, dio
);
1277 * Will be decremented at I/O completion time.
1279 if (!(dio
->flags
& DIO_SKIP_DIO_COUNT
))
1280 inode_dio_begin(inode
);
1283 sdio
.blkbits
= blkbits
;
1284 sdio
.blkfactor
= i_blkbits
- blkbits
;
1285 sdio
.block_in_file
= offset
>> blkbits
;
1287 sdio
.get_block
= get_block
;
1288 dio
->end_io
= end_io
;
1289 sdio
.submit_io
= submit_io
;
1290 sdio
.final_block_in_bio
= -1;
1291 sdio
.next_block_for_io
= -1;
1295 spin_lock_init(&dio
->bio_lock
);
1298 dio
->should_dirty
= (iter
->type
== ITER_IOVEC
);
1300 sdio
.final_block_in_request
=
1301 (offset
+ iov_iter_count(iter
)) >> blkbits
;
1304 * In case of non-aligned buffers, we may need 2 more
1305 * pages since we need to zero out first and last block.
1307 if (unlikely(sdio
.blkfactor
))
1308 sdio
.pages_in_io
= 2;
1310 sdio
.pages_in_io
+= iov_iter_npages(iter
, INT_MAX
);
1312 blk_start_plug(&plug
);
1314 retval
= do_direct_IO(dio
, &sdio
, &map_bh
);
1316 dio_cleanup(dio
, &sdio
);
1318 if (retval
== -ENOTBLK
) {
1320 * The remaining part of the request will be
1321 * be handled by buffered I/O when we return
1326 * There may be some unwritten disk at the end of a part-written
1327 * fs-block-sized block. Go zero that now.
1329 dio_zero_block(dio
, &sdio
, 1, &map_bh
);
1331 if (sdio
.cur_page
) {
1334 ret2
= dio_send_cur_page(dio
, &sdio
, &map_bh
);
1337 put_page(sdio
.cur_page
);
1338 sdio
.cur_page
= NULL
;
1341 dio_bio_submit(dio
, &sdio
);
1343 blk_finish_plug(&plug
);
1346 * It is possible that, we return short IO due to end of file.
1347 * In that case, we need to release all the pages we got hold on.
1349 dio_cleanup(dio
, &sdio
);
1352 * All block lookups have been performed. For READ requests
1353 * we can let i_mutex go now that its achieved its purpose
1354 * of protecting us from looking up uninitialized blocks.
1356 if (iov_iter_rw(iter
) == READ
&& (dio
->flags
& DIO_LOCKING
))
1357 inode_unlock(dio
->inode
);
1360 * The only time we want to leave bios in flight is when a successful
1361 * partial aio read or full aio write have been setup. In that case
1362 * bio completion will call aio_complete. The only time it's safe to
1363 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1364 * This had *better* be the only place that raises -EIOCBQUEUED.
1366 BUG_ON(retval
== -EIOCBQUEUED
);
1367 if (dio
->is_async
&& retval
== 0 && dio
->result
&&
1368 (iov_iter_rw(iter
) == READ
|| dio
->result
== count
))
1369 retval
= -EIOCBQUEUED
;
1371 dio_await_completion(dio
);
1373 if (drop_refcount(dio
) == 0) {
1374 retval
= dio_complete(dio
, retval
, DIO_COMPLETE_INVALIDATE
);
1376 BUG_ON(retval
!= -EIOCBQUEUED
);
1382 ssize_t
__blockdev_direct_IO(struct kiocb
*iocb
, struct inode
*inode
,
1383 struct block_device
*bdev
, struct iov_iter
*iter
,
1384 get_block_t get_block
,
1385 dio_iodone_t end_io
, dio_submit_t submit_io
,
1389 * The block device state is needed in the end to finally
1390 * submit everything. Since it's likely to be cache cold
1391 * prefetch it here as first thing to hide some of the
1394 * Attempt to prefetch the pieces we likely need later.
1396 prefetch(&bdev
->bd_disk
->part_tbl
);
1397 prefetch(bdev
->bd_queue
);
1398 prefetch((char *)bdev
->bd_queue
+ SMP_CACHE_BYTES
);
1400 return do_blockdev_direct_IO(iocb
, inode
, bdev
, iter
, get_block
,
1401 end_io
, submit_io
, flags
);
1404 EXPORT_SYMBOL(__blockdev_direct_IO
);
1406 static __init
int dio_init(void)
1408 dio_cache
= KMEM_CACHE(dio
, SLAB_PANIC
);
1411 module_init(dio_init
)