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 * This code generally works in units of "dio_blocks". A dio_block is
49 * somewhere between the hard sector size and the filesystem block size. it
50 * is determined on a per-invocation basis. When talking to the filesystem
51 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
52 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
53 * to bio_block quantities by shifting left by blkfactor.
55 * If blkfactor is zero then the user's request was aligned to the filesystem's
59 /* dio_state only used in the submission path */
62 struct bio
*bio
; /* bio under assembly */
63 unsigned blkbits
; /* doesn't change */
64 unsigned blkfactor
; /* When we're using an alignment which
65 is finer than the filesystem's soft
66 blocksize, this specifies how much
67 finer. blkfactor=2 means 1/4-block
68 alignment. Does not change */
69 unsigned start_zero_done
; /* flag: sub-blocksize zeroing has
70 been performed at the start of a
72 int pages_in_io
; /* approximate total IO pages */
73 sector_t block_in_file
; /* Current offset into the underlying
74 file in dio_block units. */
75 unsigned blocks_available
; /* At block_in_file. changes */
76 int reap_counter
; /* rate limit reaping */
77 sector_t final_block_in_request
;/* doesn't change */
78 int boundary
; /* prev block is at a boundary */
79 get_block_t
*get_block
; /* block mapping function */
80 dio_submit_t
*submit_io
; /* IO submition function */
82 loff_t logical_offset_in_bio
; /* current first logical block in bio */
83 sector_t final_block_in_bio
; /* current final block in bio + 1 */
84 sector_t next_block_for_io
; /* next block to be put under IO,
85 in dio_blocks units */
88 * Deferred addition of a page to the dio. These variables are
89 * private to dio_send_cur_page(), submit_page_section() and
92 struct page
*cur_page
; /* The page */
93 unsigned cur_page_offset
; /* Offset into it, in bytes */
94 unsigned cur_page_len
; /* Nr of bytes at cur_page_offset */
95 sector_t cur_page_block
; /* Where it starts */
96 loff_t cur_page_fs_offset
; /* Offset in file */
98 struct iov_iter
*iter
;
100 * Page queue. These variables belong to dio_refill_pages() and
103 unsigned head
; /* next page to process */
104 unsigned tail
; /* last valid page + 1 */
108 /* dio_state communicated between submission path and end_io */
110 int flags
; /* doesn't change */
114 struct gendisk
*bio_disk
;
116 loff_t i_size
; /* i_size when submitted */
117 dio_iodone_t
*end_io
; /* IO completion function */
119 void *private; /* copy from map_bh.b_private */
121 /* BIO completion state */
122 spinlock_t bio_lock
; /* protects BIO fields below */
123 int page_errors
; /* errno from get_user_pages() */
124 int is_async
; /* is IO async ? */
125 bool defer_completion
; /* defer AIO completion to workqueue? */
126 bool should_dirty
; /* if pages should be dirtied */
127 int io_error
; /* IO error in completion path */
128 unsigned long refcount
; /* direct_io_worker() and bios */
129 struct bio
*bio_list
; /* singly linked via bi_private */
130 struct task_struct
*waiter
; /* waiting task (NULL if none) */
132 /* AIO related stuff */
133 struct kiocb
*iocb
; /* kiocb */
134 ssize_t result
; /* IO result */
137 * pages[] (and any fields placed after it) are not zeroed out at
138 * allocation time. Don't add new fields after pages[] unless you
139 * wish that they not be zeroed.
142 struct page
*pages
[DIO_PAGES
]; /* page buffer */
143 struct work_struct complete_work
;/* deferred AIO completion */
145 } ____cacheline_aligned_in_smp
;
147 static struct kmem_cache
*dio_cache __read_mostly
;
150 * How many pages are in the queue?
152 static inline unsigned dio_pages_present(struct dio_submit
*sdio
)
154 return sdio
->tail
- sdio
->head
;
158 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
160 static inline int dio_refill_pages(struct dio
*dio
, struct dio_submit
*sdio
)
164 ret
= iov_iter_get_pages(sdio
->iter
, dio
->pages
, LONG_MAX
, DIO_PAGES
,
167 if (ret
< 0 && sdio
->blocks_available
&& (dio
->op
== REQ_OP_WRITE
)) {
168 struct page
*page
= ZERO_PAGE(0);
170 * A memory fault, but the filesystem has some outstanding
171 * mapped blocks. We need to use those blocks up to avoid
172 * leaking stale data in the file.
174 if (dio
->page_errors
== 0)
175 dio
->page_errors
= ret
;
177 dio
->pages
[0] = page
;
181 sdio
->to
= PAGE_SIZE
;
186 iov_iter_advance(sdio
->iter
, ret
);
189 sdio
->tail
= (ret
+ PAGE_SIZE
- 1) / PAGE_SIZE
;
190 sdio
->to
= ((ret
- 1) & (PAGE_SIZE
- 1)) + 1;
197 * Get another userspace page. Returns an ERR_PTR on error. Pages are
198 * buffered inside the dio so that we can call get_user_pages() against a
199 * decent number of pages, less frequently. To provide nicer use of the
202 static inline struct page
*dio_get_page(struct dio
*dio
,
203 struct dio_submit
*sdio
)
205 if (dio_pages_present(sdio
) == 0) {
208 ret
= dio_refill_pages(dio
, sdio
);
211 BUG_ON(dio_pages_present(sdio
) == 0);
213 return dio
->pages
[sdio
->head
];
217 * dio_complete() - called when all DIO BIO I/O has been completed
218 * @offset: the byte offset in the file of the completed operation
220 * This drops i_dio_count, lets interested parties know that a DIO operation
221 * has completed, and calculates the resulting return code for the operation.
223 * It lets the filesystem know if it registered an interest earlier via
224 * get_block. Pass the private field of the map buffer_head so that
225 * filesystems can use it to hold additional state between get_block calls and
228 static ssize_t
dio_complete(struct dio
*dio
, ssize_t ret
, bool is_async
)
230 loff_t offset
= dio
->iocb
->ki_pos
;
231 ssize_t transferred
= 0;
235 * AIO submission can race with bio completion to get here while
236 * expecting to have the last io completed by bio completion.
237 * In that case -EIOCBQUEUED is in fact not an error we want
238 * to preserve through this call.
240 if (ret
== -EIOCBQUEUED
)
244 transferred
= dio
->result
;
246 /* Check for short read case */
247 if ((dio
->op
== REQ_OP_READ
) &&
248 ((offset
+ transferred
) > dio
->i_size
))
249 transferred
= dio
->i_size
- offset
;
250 /* ignore EFAULT if some IO has been done */
251 if (unlikely(ret
== -EFAULT
) && transferred
)
256 ret
= dio
->page_errors
;
263 * Try again to invalidate clean pages which might have been cached by
264 * non-direct readahead, or faulted in by get_user_pages() if the source
265 * of the write was an mmap'ed region of the file we're writing. Either
266 * one is a pretty crazy thing to do, so we don't support it 100%. If
267 * this invalidation fails, tough, the write still worked...
269 if (ret
> 0 && dio
->op
== REQ_OP_WRITE
&&
270 dio
->inode
->i_mapping
->nrpages
) {
271 err
= invalidate_inode_pages2_range(dio
->inode
->i_mapping
,
272 offset
>> PAGE_SHIFT
,
273 (offset
+ ret
- 1) >> PAGE_SHIFT
);
280 err
= dio
->end_io(dio
->iocb
, offset
, ret
, dio
->private);
285 if (!(dio
->flags
& DIO_SKIP_DIO_COUNT
))
286 inode_dio_end(dio
->inode
);
290 * generic_write_sync expects ki_pos to have been updated
291 * already, but the submission path only does this for
294 dio
->iocb
->ki_pos
+= transferred
;
296 if (dio
->op
== REQ_OP_WRITE
)
297 ret
= generic_write_sync(dio
->iocb
, transferred
);
298 dio
->iocb
->ki_complete(dio
->iocb
, ret
, 0);
301 kmem_cache_free(dio_cache
, dio
);
305 static void dio_aio_complete_work(struct work_struct
*work
)
307 struct dio
*dio
= container_of(work
, struct dio
, complete_work
);
309 dio_complete(dio
, 0, true);
312 static blk_status_t
dio_bio_complete(struct dio
*dio
, struct bio
*bio
);
315 * Asynchronous IO callback.
317 static void dio_bio_end_aio(struct bio
*bio
)
319 struct dio
*dio
= bio
->bi_private
;
320 unsigned long remaining
;
322 bool defer_completion
= false;
324 /* cleanup the bio */
325 dio_bio_complete(dio
, bio
);
327 spin_lock_irqsave(&dio
->bio_lock
, flags
);
328 remaining
= --dio
->refcount
;
329 if (remaining
== 1 && dio
->waiter
)
330 wake_up_process(dio
->waiter
);
331 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
333 if (remaining
== 0) {
335 * Defer completion when defer_completion is set or
336 * when the inode has pages mapped and this is AIO write.
337 * We need to invalidate those pages because there is a
338 * chance they contain stale data in the case buffered IO
339 * went in between AIO submission and completion into the
343 defer_completion
= dio
->defer_completion
||
344 (dio
->op
== REQ_OP_WRITE
&&
345 dio
->inode
->i_mapping
->nrpages
);
346 if (defer_completion
) {
347 INIT_WORK(&dio
->complete_work
, dio_aio_complete_work
);
348 queue_work(dio
->inode
->i_sb
->s_dio_done_wq
,
349 &dio
->complete_work
);
351 dio_complete(dio
, 0, true);
357 * The BIO completion handler simply queues the BIO up for the process-context
360 * During I/O bi_private points at the dio. After I/O, bi_private is used to
361 * implement a singly-linked list of completed BIOs, at dio->bio_list.
363 static void dio_bio_end_io(struct bio
*bio
)
365 struct dio
*dio
= bio
->bi_private
;
368 spin_lock_irqsave(&dio
->bio_lock
, flags
);
369 bio
->bi_private
= dio
->bio_list
;
371 if (--dio
->refcount
== 1 && dio
->waiter
)
372 wake_up_process(dio
->waiter
);
373 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
377 * dio_end_io - handle the end io action for the given bio
378 * @bio: The direct io bio thats being completed
380 * This is meant to be called by any filesystem that uses their own dio_submit_t
381 * so that the DIO specific endio actions are dealt with after the filesystem
382 * has done it's completion work.
384 void dio_end_io(struct bio
*bio
)
386 struct dio
*dio
= bio
->bi_private
;
389 dio_bio_end_aio(bio
);
393 EXPORT_SYMBOL_GPL(dio_end_io
);
396 dio_bio_alloc(struct dio
*dio
, struct dio_submit
*sdio
,
397 struct block_device
*bdev
,
398 sector_t first_sector
, int nr_vecs
)
403 * bio_alloc() is guaranteed to return a bio when called with
404 * __GFP_RECLAIM and we request a valid number of vectors.
406 bio
= bio_alloc(GFP_KERNEL
, nr_vecs
);
408 bio_set_dev(bio
, bdev
);
409 bio
->bi_iter
.bi_sector
= first_sector
;
410 bio_set_op_attrs(bio
, dio
->op
, dio
->op_flags
);
412 bio
->bi_end_io
= dio_bio_end_aio
;
414 bio
->bi_end_io
= dio_bio_end_io
;
416 bio
->bi_write_hint
= dio
->iocb
->ki_hint
;
419 sdio
->logical_offset_in_bio
= sdio
->cur_page_fs_offset
;
423 * In the AIO read case we speculatively dirty the pages before starting IO.
424 * During IO completion, any of these pages which happen to have been written
425 * back will be redirtied by bio_check_pages_dirty().
427 * bios hold a dio reference between submit_bio and ->end_io.
429 static inline void dio_bio_submit(struct dio
*dio
, struct dio_submit
*sdio
)
431 struct bio
*bio
= sdio
->bio
;
434 bio
->bi_private
= dio
;
436 spin_lock_irqsave(&dio
->bio_lock
, flags
);
438 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
440 if (dio
->is_async
&& dio
->op
== REQ_OP_READ
&& dio
->should_dirty
)
441 bio_set_pages_dirty(bio
);
443 dio
->bio_disk
= bio
->bi_disk
;
445 if (sdio
->submit_io
) {
446 sdio
->submit_io(bio
, dio
->inode
, sdio
->logical_offset_in_bio
);
447 dio
->bio_cookie
= BLK_QC_T_NONE
;
449 dio
->bio_cookie
= submit_bio(bio
);
453 sdio
->logical_offset_in_bio
= 0;
457 * Release any resources in case of a failure
459 static inline void dio_cleanup(struct dio
*dio
, struct dio_submit
*sdio
)
461 while (sdio
->head
< sdio
->tail
)
462 put_page(dio
->pages
[sdio
->head
++]);
466 * Wait for the next BIO to complete. Remove it and return it. NULL is
467 * returned once all BIOs have been completed. This must only be called once
468 * all bios have been issued so that dio->refcount can only decrease. This
469 * requires that that the caller hold a reference on the dio.
471 static struct bio
*dio_await_one(struct dio
*dio
)
474 struct bio
*bio
= NULL
;
476 spin_lock_irqsave(&dio
->bio_lock
, flags
);
479 * Wait as long as the list is empty and there are bios in flight. bio
480 * completion drops the count, maybe adds to the list, and wakes while
481 * holding the bio_lock so we don't need set_current_state()'s barrier
482 * and can call it after testing our condition.
484 while (dio
->refcount
> 1 && dio
->bio_list
== NULL
) {
485 __set_current_state(TASK_UNINTERRUPTIBLE
);
486 dio
->waiter
= current
;
487 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
488 if (!(dio
->iocb
->ki_flags
& IOCB_HIPRI
) ||
489 !blk_mq_poll(dio
->bio_disk
->queue
, dio
->bio_cookie
))
491 /* wake up sets us TASK_RUNNING */
492 spin_lock_irqsave(&dio
->bio_lock
, flags
);
497 dio
->bio_list
= bio
->bi_private
;
499 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
504 * Process one completed BIO. No locks are held.
506 static blk_status_t
dio_bio_complete(struct dio
*dio
, struct bio
*bio
)
508 struct bio_vec
*bvec
;
510 blk_status_t err
= bio
->bi_status
;
513 if (err
== BLK_STS_AGAIN
&& (bio
->bi_opf
& REQ_NOWAIT
))
514 dio
->io_error
= -EAGAIN
;
516 dio
->io_error
= -EIO
;
519 if (dio
->is_async
&& dio
->op
== REQ_OP_READ
&& dio
->should_dirty
) {
520 bio_check_pages_dirty(bio
); /* transfers ownership */
522 bio_for_each_segment_all(bvec
, bio
, i
) {
523 struct page
*page
= bvec
->bv_page
;
525 if (dio
->op
== REQ_OP_READ
&& !PageCompound(page
) &&
527 set_page_dirty_lock(page
);
536 * Wait on and process all in-flight BIOs. This must only be called once
537 * all bios have been issued so that the refcount can only decrease.
538 * This just waits for all bios to make it through dio_bio_complete. IO
539 * errors are propagated through dio->io_error and should be propagated via
542 static void dio_await_completion(struct dio
*dio
)
546 bio
= dio_await_one(dio
);
548 dio_bio_complete(dio
, bio
);
553 * A really large O_DIRECT read or write can generate a lot of BIOs. So
554 * to keep the memory consumption sane we periodically reap any completed BIOs
555 * during the BIO generation phase.
557 * This also helps to limit the peak amount of pinned userspace memory.
559 static inline int dio_bio_reap(struct dio
*dio
, struct dio_submit
*sdio
)
563 if (sdio
->reap_counter
++ >= 64) {
564 while (dio
->bio_list
) {
569 spin_lock_irqsave(&dio
->bio_lock
, flags
);
571 dio
->bio_list
= bio
->bi_private
;
572 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
573 ret2
= blk_status_to_errno(dio_bio_complete(dio
, bio
));
577 sdio
->reap_counter
= 0;
583 * Create workqueue for deferred direct IO completions. We allocate the
584 * workqueue when it's first needed. This avoids creating workqueue for
585 * filesystems that don't need it and also allows us to create the workqueue
586 * late enough so the we can include s_id in the name of the workqueue.
588 int sb_init_dio_done_wq(struct super_block
*sb
)
590 struct workqueue_struct
*old
;
591 struct workqueue_struct
*wq
= alloc_workqueue("dio/%s",
597 * This has to be atomic as more DIOs can race to create the workqueue
599 old
= cmpxchg(&sb
->s_dio_done_wq
, NULL
, wq
);
600 /* Someone created workqueue before us? Free ours... */
602 destroy_workqueue(wq
);
606 static int dio_set_defer_completion(struct dio
*dio
)
608 struct super_block
*sb
= dio
->inode
->i_sb
;
610 if (dio
->defer_completion
)
612 dio
->defer_completion
= true;
613 if (!sb
->s_dio_done_wq
)
614 return sb_init_dio_done_wq(sb
);
619 * Call into the fs to map some more disk blocks. We record the current number
620 * of available blocks at sdio->blocks_available. These are in units of the
621 * fs blocksize, i_blocksize(inode).
623 * The fs is allowed to map lots of blocks at once. If it wants to do that,
624 * it uses the passed inode-relative block number as the file offset, as usual.
626 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
627 * has remaining to do. The fs should not map more than this number of blocks.
629 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
630 * indicate how much contiguous disk space has been made available at
633 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
634 * This isn't very efficient...
636 * In the case of filesystem holes: the fs may return an arbitrarily-large
637 * hole by returning an appropriate value in b_size and by clearing
638 * buffer_mapped(). However the direct-io code will only process holes one
639 * block at a time - it will repeatedly call get_block() as it walks the hole.
641 static int get_more_blocks(struct dio
*dio
, struct dio_submit
*sdio
,
642 struct buffer_head
*map_bh
)
645 sector_t fs_startblk
; /* Into file, in filesystem-sized blocks */
646 sector_t fs_endblk
; /* Into file, in filesystem-sized blocks */
647 unsigned long fs_count
; /* Number of filesystem-sized blocks */
649 unsigned int i_blkbits
= sdio
->blkbits
+ sdio
->blkfactor
;
652 * If there was a memory error and we've overwritten all the
653 * mapped blocks then we can now return that memory error
655 ret
= dio
->page_errors
;
657 BUG_ON(sdio
->block_in_file
>= sdio
->final_block_in_request
);
658 fs_startblk
= sdio
->block_in_file
>> sdio
->blkfactor
;
659 fs_endblk
= (sdio
->final_block_in_request
- 1) >>
661 fs_count
= fs_endblk
- fs_startblk
+ 1;
664 map_bh
->b_size
= fs_count
<< i_blkbits
;
667 * For writes that could fill holes inside i_size on a
668 * DIO_SKIP_HOLES filesystem we forbid block creations: only
669 * overwrites are permitted. We will return early to the caller
670 * once we see an unmapped buffer head returned, and the caller
671 * will fall back to buffered I/O.
673 * Otherwise the decision is left to the get_blocks method,
674 * which may decide to handle it or also return an unmapped
677 create
= dio
->op
== REQ_OP_WRITE
;
678 if (dio
->flags
& DIO_SKIP_HOLES
) {
679 if (fs_startblk
<= ((i_size_read(dio
->inode
) - 1) >>
684 ret
= (*sdio
->get_block
)(dio
->inode
, fs_startblk
,
687 /* Store for completion */
688 dio
->private = map_bh
->b_private
;
690 if (ret
== 0 && buffer_defer_completion(map_bh
))
691 ret
= dio_set_defer_completion(dio
);
697 * There is no bio. Make one now.
699 static inline int dio_new_bio(struct dio
*dio
, struct dio_submit
*sdio
,
700 sector_t start_sector
, struct buffer_head
*map_bh
)
705 ret
= dio_bio_reap(dio
, sdio
);
708 sector
= start_sector
<< (sdio
->blkbits
- 9);
709 nr_pages
= min(sdio
->pages_in_io
, BIO_MAX_PAGES
);
710 BUG_ON(nr_pages
<= 0);
711 dio_bio_alloc(dio
, sdio
, map_bh
->b_bdev
, sector
, nr_pages
);
718 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
719 * that was successful then update final_block_in_bio and take a ref against
720 * the just-added page.
722 * Return zero on success. Non-zero means the caller needs to start a new BIO.
724 static inline int dio_bio_add_page(struct dio_submit
*sdio
)
728 ret
= bio_add_page(sdio
->bio
, sdio
->cur_page
,
729 sdio
->cur_page_len
, sdio
->cur_page_offset
);
730 if (ret
== sdio
->cur_page_len
) {
732 * Decrement count only, if we are done with this page
734 if ((sdio
->cur_page_len
+ sdio
->cur_page_offset
) == PAGE_SIZE
)
736 get_page(sdio
->cur_page
);
737 sdio
->final_block_in_bio
= sdio
->cur_page_block
+
738 (sdio
->cur_page_len
>> sdio
->blkbits
);
747 * Put cur_page under IO. The section of cur_page which is described by
748 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
749 * starts on-disk at cur_page_block.
751 * We take a ref against the page here (on behalf of its presence in the bio).
753 * The caller of this function is responsible for removing cur_page from the
754 * dio, and for dropping the refcount which came from that presence.
756 static inline int dio_send_cur_page(struct dio
*dio
, struct dio_submit
*sdio
,
757 struct buffer_head
*map_bh
)
762 loff_t cur_offset
= sdio
->cur_page_fs_offset
;
763 loff_t bio_next_offset
= sdio
->logical_offset_in_bio
+
764 sdio
->bio
->bi_iter
.bi_size
;
767 * See whether this new request is contiguous with the old.
769 * Btrfs cannot handle having logically non-contiguous requests
770 * submitted. For example if you have
772 * Logical: [0-4095][HOLE][8192-12287]
773 * Physical: [0-4095] [4096-8191]
775 * We cannot submit those pages together as one BIO. So if our
776 * current logical offset in the file does not equal what would
777 * be the next logical offset in the bio, submit the bio we
780 if (sdio
->final_block_in_bio
!= sdio
->cur_page_block
||
781 cur_offset
!= bio_next_offset
)
782 dio_bio_submit(dio
, sdio
);
785 if (sdio
->bio
== NULL
) {
786 ret
= dio_new_bio(dio
, sdio
, sdio
->cur_page_block
, map_bh
);
791 if (dio_bio_add_page(sdio
) != 0) {
792 dio_bio_submit(dio
, sdio
);
793 ret
= dio_new_bio(dio
, sdio
, sdio
->cur_page_block
, map_bh
);
795 ret
= dio_bio_add_page(sdio
);
804 * An autonomous function to put a chunk of a page under deferred IO.
806 * The caller doesn't actually know (or care) whether this piece of page is in
807 * a BIO, or is under IO or whatever. We just take care of all possible
808 * situations here. The separation between the logic of do_direct_IO() and
809 * that of submit_page_section() is important for clarity. Please don't break.
811 * The chunk of page starts on-disk at blocknr.
813 * We perform deferred IO, by recording the last-submitted page inside our
814 * private part of the dio structure. If possible, we just expand the IO
815 * across that page here.
817 * If that doesn't work out then we put the old page into the bio and add this
818 * page to the dio instead.
821 submit_page_section(struct dio
*dio
, struct dio_submit
*sdio
, struct page
*page
,
822 unsigned offset
, unsigned len
, sector_t blocknr
,
823 struct buffer_head
*map_bh
)
827 if (dio
->op
== REQ_OP_WRITE
) {
829 * Read accounting is performed in submit_bio()
831 task_io_account_write(len
);
835 * Can we just grow the current page's presence in the dio?
837 if (sdio
->cur_page
== page
&&
838 sdio
->cur_page_offset
+ sdio
->cur_page_len
== offset
&&
839 sdio
->cur_page_block
+
840 (sdio
->cur_page_len
>> sdio
->blkbits
) == blocknr
) {
841 sdio
->cur_page_len
+= len
;
846 * If there's a deferred page already there then send it.
848 if (sdio
->cur_page
) {
849 ret
= dio_send_cur_page(dio
, sdio
, map_bh
);
850 put_page(sdio
->cur_page
);
851 sdio
->cur_page
= NULL
;
856 get_page(page
); /* It is in dio */
857 sdio
->cur_page
= page
;
858 sdio
->cur_page_offset
= offset
;
859 sdio
->cur_page_len
= len
;
860 sdio
->cur_page_block
= blocknr
;
861 sdio
->cur_page_fs_offset
= sdio
->block_in_file
<< sdio
->blkbits
;
864 * If sdio->boundary then we want to schedule the IO now to
865 * avoid metadata seeks.
867 if (sdio
->boundary
) {
868 ret
= dio_send_cur_page(dio
, sdio
, map_bh
);
869 dio_bio_submit(dio
, sdio
);
870 put_page(sdio
->cur_page
);
871 sdio
->cur_page
= NULL
;
877 * If we are not writing the entire block and get_block() allocated
878 * the block for us, we need to fill-in the unused portion of the
879 * block with zeros. This happens only if user-buffer, fileoffset or
880 * io length is not filesystem block-size multiple.
882 * `end' is zero if we're doing the start of the IO, 1 at the end of the
885 static inline void dio_zero_block(struct dio
*dio
, struct dio_submit
*sdio
,
886 int end
, struct buffer_head
*map_bh
)
888 unsigned dio_blocks_per_fs_block
;
889 unsigned this_chunk_blocks
; /* In dio_blocks */
890 unsigned this_chunk_bytes
;
893 sdio
->start_zero_done
= 1;
894 if (!sdio
->blkfactor
|| !buffer_new(map_bh
))
897 dio_blocks_per_fs_block
= 1 << sdio
->blkfactor
;
898 this_chunk_blocks
= sdio
->block_in_file
& (dio_blocks_per_fs_block
- 1);
900 if (!this_chunk_blocks
)
904 * We need to zero out part of an fs block. It is either at the
905 * beginning or the end of the fs block.
908 this_chunk_blocks
= dio_blocks_per_fs_block
- this_chunk_blocks
;
910 this_chunk_bytes
= this_chunk_blocks
<< sdio
->blkbits
;
913 if (submit_page_section(dio
, sdio
, page
, 0, this_chunk_bytes
,
914 sdio
->next_block_for_io
, map_bh
))
917 sdio
->next_block_for_io
+= this_chunk_blocks
;
921 * Walk the user pages, and the file, mapping blocks to disk and generating
922 * a sequence of (page,offset,len,block) mappings. These mappings are injected
923 * into submit_page_section(), which takes care of the next stage of submission
925 * Direct IO against a blockdev is different from a file. Because we can
926 * happily perform page-sized but 512-byte aligned IOs. It is important that
927 * blockdev IO be able to have fine alignment and large sizes.
929 * So what we do is to permit the ->get_block function to populate bh.b_size
930 * with the size of IO which is permitted at this offset and this i_blkbits.
932 * For best results, the blockdev should be set up with 512-byte i_blkbits and
933 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
934 * fine alignment but still allows this function to work in PAGE_SIZE units.
936 static int do_direct_IO(struct dio
*dio
, struct dio_submit
*sdio
,
937 struct buffer_head
*map_bh
)
939 const unsigned blkbits
= sdio
->blkbits
;
940 const unsigned i_blkbits
= blkbits
+ sdio
->blkfactor
;
943 while (sdio
->block_in_file
< sdio
->final_block_in_request
) {
947 page
= dio_get_page(dio
, sdio
);
952 from
= sdio
->head
? 0 : sdio
->from
;
953 to
= (sdio
->head
== sdio
->tail
- 1) ? sdio
->to
: PAGE_SIZE
;
957 unsigned this_chunk_bytes
; /* # of bytes mapped */
958 unsigned this_chunk_blocks
; /* # of blocks */
961 if (sdio
->blocks_available
== 0) {
963 * Need to go and map some more disk
965 unsigned long blkmask
;
966 unsigned long dio_remainder
;
968 ret
= get_more_blocks(dio
, sdio
, map_bh
);
973 if (!buffer_mapped(map_bh
))
976 sdio
->blocks_available
=
977 map_bh
->b_size
>> blkbits
;
978 sdio
->next_block_for_io
=
979 map_bh
->b_blocknr
<< sdio
->blkfactor
;
980 if (buffer_new(map_bh
)) {
984 map_bh
->b_size
>> i_blkbits
);
987 if (!sdio
->blkfactor
)
990 blkmask
= (1 << sdio
->blkfactor
) - 1;
991 dio_remainder
= (sdio
->block_in_file
& blkmask
);
994 * If we are at the start of IO and that IO
995 * starts partway into a fs-block,
996 * dio_remainder will be non-zero. If the IO
997 * is a read then we can simply advance the IO
998 * cursor to the first block which is to be
999 * read. But if the IO is a write and the
1000 * block was newly allocated we cannot do that;
1001 * the start of the fs block must be zeroed out
1004 if (!buffer_new(map_bh
))
1005 sdio
->next_block_for_io
+= dio_remainder
;
1006 sdio
->blocks_available
-= dio_remainder
;
1010 if (!buffer_mapped(map_bh
)) {
1011 loff_t i_size_aligned
;
1013 /* AKPM: eargh, -ENOTBLK is a hack */
1014 if (dio
->op
== REQ_OP_WRITE
) {
1020 * Be sure to account for a partial block as the
1021 * last block in the file
1023 i_size_aligned
= ALIGN(i_size_read(dio
->inode
),
1025 if (sdio
->block_in_file
>=
1026 i_size_aligned
>> blkbits
) {
1031 zero_user(page
, from
, 1 << blkbits
);
1032 sdio
->block_in_file
++;
1033 from
+= 1 << blkbits
;
1034 dio
->result
+= 1 << blkbits
;
1039 * If we're performing IO which has an alignment which
1040 * is finer than the underlying fs, go check to see if
1041 * we must zero out the start of this block.
1043 if (unlikely(sdio
->blkfactor
&& !sdio
->start_zero_done
))
1044 dio_zero_block(dio
, sdio
, 0, map_bh
);
1047 * Work out, in this_chunk_blocks, how much disk we
1048 * can add to this page
1050 this_chunk_blocks
= sdio
->blocks_available
;
1051 u
= (to
- from
) >> blkbits
;
1052 if (this_chunk_blocks
> u
)
1053 this_chunk_blocks
= u
;
1054 u
= sdio
->final_block_in_request
- sdio
->block_in_file
;
1055 if (this_chunk_blocks
> u
)
1056 this_chunk_blocks
= u
;
1057 this_chunk_bytes
= this_chunk_blocks
<< blkbits
;
1058 BUG_ON(this_chunk_bytes
== 0);
1060 if (this_chunk_blocks
== sdio
->blocks_available
)
1061 sdio
->boundary
= buffer_boundary(map_bh
);
1062 ret
= submit_page_section(dio
, sdio
, page
,
1065 sdio
->next_block_for_io
,
1071 sdio
->next_block_for_io
+= this_chunk_blocks
;
1073 sdio
->block_in_file
+= this_chunk_blocks
;
1074 from
+= this_chunk_bytes
;
1075 dio
->result
+= this_chunk_bytes
;
1076 sdio
->blocks_available
-= this_chunk_blocks
;
1078 BUG_ON(sdio
->block_in_file
> sdio
->final_block_in_request
);
1079 if (sdio
->block_in_file
== sdio
->final_block_in_request
)
1083 /* Drop the ref which was taken in get_user_pages() */
1090 static inline int drop_refcount(struct dio
*dio
)
1093 unsigned long flags
;
1096 * Sync will always be dropping the final ref and completing the
1097 * operation. AIO can if it was a broken operation described above or
1098 * in fact if all the bios race to complete before we get here. In
1099 * that case dio_complete() translates the EIOCBQUEUED into the proper
1100 * return code that the caller will hand to ->complete().
1102 * This is managed by the bio_lock instead of being an atomic_t so that
1103 * completion paths can drop their ref and use the remaining count to
1104 * decide to wake the submission path atomically.
1106 spin_lock_irqsave(&dio
->bio_lock
, flags
);
1107 ret2
= --dio
->refcount
;
1108 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
1113 * This is a library function for use by filesystem drivers.
1115 * The locking rules are governed by the flags parameter:
1116 * - if the flags value contains DIO_LOCKING we use a fancy locking
1117 * scheme for dumb filesystems.
1118 * For writes this function is called under i_mutex and returns with
1119 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1120 * taken and dropped again before returning.
1121 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1122 * internal locking but rather rely on the filesystem to synchronize
1123 * direct I/O reads/writes versus each other and truncate.
1125 * To help with locking against truncate we incremented the i_dio_count
1126 * counter before starting direct I/O, and decrement it once we are done.
1127 * Truncate can wait for it to reach zero to provide exclusion. It is
1128 * expected that filesystem provide exclusion between new direct I/O
1129 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1130 * but other filesystems need to take care of this on their own.
1132 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1133 * is always inlined. Otherwise gcc is unable to split the structure into
1134 * individual fields and will generate much worse code. This is important
1135 * for the whole file.
1137 static inline ssize_t
1138 do_blockdev_direct_IO(struct kiocb
*iocb
, struct inode
*inode
,
1139 struct block_device
*bdev
, struct iov_iter
*iter
,
1140 get_block_t get_block
, dio_iodone_t end_io
,
1141 dio_submit_t submit_io
, int flags
)
1143 unsigned i_blkbits
= ACCESS_ONCE(inode
->i_blkbits
);
1144 unsigned blkbits
= i_blkbits
;
1145 unsigned blocksize_mask
= (1 << blkbits
) - 1;
1146 ssize_t retval
= -EINVAL
;
1147 size_t count
= iov_iter_count(iter
);
1148 loff_t offset
= iocb
->ki_pos
;
1149 loff_t end
= offset
+ count
;
1151 struct dio_submit sdio
= { 0, };
1152 struct buffer_head map_bh
= { 0, };
1153 struct blk_plug plug
;
1154 unsigned long align
= offset
| iov_iter_alignment(iter
);
1157 * Avoid references to bdev if not absolutely needed to give
1158 * the early prefetch in the caller enough time.
1161 if (align
& blocksize_mask
) {
1163 blkbits
= blksize_bits(bdev_logical_block_size(bdev
));
1164 blocksize_mask
= (1 << blkbits
) - 1;
1165 if (align
& blocksize_mask
)
1169 /* watch out for a 0 len io from a tricksy fs */
1170 if (iov_iter_rw(iter
) == READ
&& !iov_iter_count(iter
))
1173 dio
= kmem_cache_alloc(dio_cache
, GFP_KERNEL
);
1178 * Believe it or not, zeroing out the page array caused a .5%
1179 * performance regression in a database benchmark. So, we take
1180 * care to only zero out what's needed.
1182 memset(dio
, 0, offsetof(struct dio
, pages
));
1185 if (dio
->flags
& DIO_LOCKING
) {
1186 if (iov_iter_rw(iter
) == READ
) {
1187 struct address_space
*mapping
=
1188 iocb
->ki_filp
->f_mapping
;
1190 /* will be released by direct_io_worker */
1193 retval
= filemap_write_and_wait_range(mapping
, offset
,
1196 inode_unlock(inode
);
1197 kmem_cache_free(dio_cache
, dio
);
1203 /* Once we sampled i_size check for reads beyond EOF */
1204 dio
->i_size
= i_size_read(inode
);
1205 if (iov_iter_rw(iter
) == READ
&& offset
>= dio
->i_size
) {
1206 if (dio
->flags
& DIO_LOCKING
)
1207 inode_unlock(inode
);
1208 kmem_cache_free(dio_cache
, dio
);
1214 * For file extending writes updating i_size before data writeouts
1215 * complete can expose uninitialized blocks in dumb filesystems.
1216 * In that case we need to wait for I/O completion even if asked
1217 * for an asynchronous write.
1219 if (is_sync_kiocb(iocb
))
1220 dio
->is_async
= false;
1221 else if (!(dio
->flags
& DIO_ASYNC_EXTEND
) &&
1222 iov_iter_rw(iter
) == WRITE
&& end
> i_size_read(inode
))
1223 dio
->is_async
= false;
1225 dio
->is_async
= true;
1228 if (iov_iter_rw(iter
) == WRITE
) {
1229 dio
->op
= REQ_OP_WRITE
;
1230 dio
->op_flags
= REQ_SYNC
| REQ_IDLE
;
1231 if (iocb
->ki_flags
& IOCB_NOWAIT
)
1232 dio
->op_flags
|= REQ_NOWAIT
;
1234 dio
->op
= REQ_OP_READ
;
1238 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1239 * so that we can call ->fsync.
1241 if (dio
->is_async
&& iov_iter_rw(iter
) == WRITE
) {
1243 if ((iocb
->ki_filp
->f_flags
& O_DSYNC
) ||
1244 IS_SYNC(iocb
->ki_filp
->f_mapping
->host
))
1245 retval
= dio_set_defer_completion(dio
);
1246 else if (!dio
->inode
->i_sb
->s_dio_done_wq
) {
1248 * In case of AIO write racing with buffered read we
1249 * need to defer completion. We can't decide this now,
1250 * however the workqueue needs to be initialized here.
1252 retval
= sb_init_dio_done_wq(dio
->inode
->i_sb
);
1256 * We grab i_mutex only for reads so we don't have
1257 * to release it here
1259 kmem_cache_free(dio_cache
, dio
);
1265 * Will be decremented at I/O completion time.
1267 if (!(dio
->flags
& DIO_SKIP_DIO_COUNT
))
1268 inode_dio_begin(inode
);
1271 sdio
.blkbits
= blkbits
;
1272 sdio
.blkfactor
= i_blkbits
- blkbits
;
1273 sdio
.block_in_file
= offset
>> blkbits
;
1275 sdio
.get_block
= get_block
;
1276 dio
->end_io
= end_io
;
1277 sdio
.submit_io
= submit_io
;
1278 sdio
.final_block_in_bio
= -1;
1279 sdio
.next_block_for_io
= -1;
1283 spin_lock_init(&dio
->bio_lock
);
1286 dio
->should_dirty
= (iter
->type
== ITER_IOVEC
);
1288 sdio
.final_block_in_request
=
1289 (offset
+ iov_iter_count(iter
)) >> blkbits
;
1292 * In case of non-aligned buffers, we may need 2 more
1293 * pages since we need to zero out first and last block.
1295 if (unlikely(sdio
.blkfactor
))
1296 sdio
.pages_in_io
= 2;
1298 sdio
.pages_in_io
+= iov_iter_npages(iter
, INT_MAX
);
1300 blk_start_plug(&plug
);
1302 retval
= do_direct_IO(dio
, &sdio
, &map_bh
);
1304 dio_cleanup(dio
, &sdio
);
1306 if (retval
== -ENOTBLK
) {
1308 * The remaining part of the request will be
1309 * be handled by buffered I/O when we return
1314 * There may be some unwritten disk at the end of a part-written
1315 * fs-block-sized block. Go zero that now.
1317 dio_zero_block(dio
, &sdio
, 1, &map_bh
);
1319 if (sdio
.cur_page
) {
1322 ret2
= dio_send_cur_page(dio
, &sdio
, &map_bh
);
1325 put_page(sdio
.cur_page
);
1326 sdio
.cur_page
= NULL
;
1329 dio_bio_submit(dio
, &sdio
);
1331 blk_finish_plug(&plug
);
1334 * It is possible that, we return short IO due to end of file.
1335 * In that case, we need to release all the pages we got hold on.
1337 dio_cleanup(dio
, &sdio
);
1340 * All block lookups have been performed. For READ requests
1341 * we can let i_mutex go now that its achieved its purpose
1342 * of protecting us from looking up uninitialized blocks.
1344 if (iov_iter_rw(iter
) == READ
&& (dio
->flags
& DIO_LOCKING
))
1345 inode_unlock(dio
->inode
);
1348 * The only time we want to leave bios in flight is when a successful
1349 * partial aio read or full aio write have been setup. In that case
1350 * bio completion will call aio_complete. The only time it's safe to
1351 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1352 * This had *better* be the only place that raises -EIOCBQUEUED.
1354 BUG_ON(retval
== -EIOCBQUEUED
);
1355 if (dio
->is_async
&& retval
== 0 && dio
->result
&&
1356 (iov_iter_rw(iter
) == READ
|| dio
->result
== count
))
1357 retval
= -EIOCBQUEUED
;
1359 dio_await_completion(dio
);
1361 if (drop_refcount(dio
) == 0) {
1362 retval
= dio_complete(dio
, retval
, false);
1364 BUG_ON(retval
!= -EIOCBQUEUED
);
1370 ssize_t
__blockdev_direct_IO(struct kiocb
*iocb
, struct inode
*inode
,
1371 struct block_device
*bdev
, struct iov_iter
*iter
,
1372 get_block_t get_block
,
1373 dio_iodone_t end_io
, dio_submit_t submit_io
,
1377 * The block device state is needed in the end to finally
1378 * submit everything. Since it's likely to be cache cold
1379 * prefetch it here as first thing to hide some of the
1382 * Attempt to prefetch the pieces we likely need later.
1384 prefetch(&bdev
->bd_disk
->part_tbl
);
1385 prefetch(bdev
->bd_queue
);
1386 prefetch((char *)bdev
->bd_queue
+ SMP_CACHE_BYTES
);
1388 return do_blockdev_direct_IO(iocb
, inode
, bdev
, iter
, get_block
,
1389 end_io
, submit_io
, flags
);
1392 EXPORT_SYMBOL(__blockdev_direct_IO
);
1394 static __init
int dio_init(void)
1396 dio_cache
= KMEM_CACHE(dio
, SLAB_PANIC
);
1399 module_init(dio_init
)