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>
40 #include <linux/aio.h>
43 * How many user pages to map in one call to get_user_pages(). This determines
44 * the size of a structure in the slab cache
49 * This code generally works in units of "dio_blocks". A dio_block is
50 * somewhere between the hard sector size and the filesystem block size. it
51 * is determined on a per-invocation basis. When talking to the filesystem
52 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
53 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
54 * to bio_block quantities by shifting left by blkfactor.
56 * If blkfactor is zero then the user's request was aligned to the filesystem's
60 /* dio_state only used in the submission path */
63 struct bio
*bio
; /* bio under assembly */
64 unsigned blkbits
; /* doesn't change */
65 unsigned blkfactor
; /* When we're using an alignment which
66 is finer than the filesystem's soft
67 blocksize, this specifies how much
68 finer. blkfactor=2 means 1/4-block
69 alignment. Does not change */
70 unsigned start_zero_done
; /* flag: sub-blocksize zeroing has
71 been performed at the start of a
73 int pages_in_io
; /* approximate total IO pages */
74 size_t size
; /* total request size (doesn't change)*/
75 sector_t block_in_file
; /* Current offset into the underlying
76 file in dio_block units. */
77 unsigned blocks_available
; /* At block_in_file. changes */
78 int reap_counter
; /* rate limit reaping */
79 sector_t final_block_in_request
;/* doesn't change */
80 unsigned first_block_in_page
; /* doesn't change, Used only once */
81 int boundary
; /* prev block is at a boundary */
82 get_block_t
*get_block
; /* block mapping function */
83 dio_submit_t
*submit_io
; /* IO submition function */
85 loff_t logical_offset_in_bio
; /* current first logical block in bio */
86 sector_t final_block_in_bio
; /* current final block in bio + 1 */
87 sector_t next_block_for_io
; /* next block to be put under IO,
88 in dio_blocks units */
91 * Deferred addition of a page to the dio. These variables are
92 * private to dio_send_cur_page(), submit_page_section() and
95 struct page
*cur_page
; /* The page */
96 unsigned cur_page_offset
; /* Offset into it, in bytes */
97 unsigned cur_page_len
; /* Nr of bytes at cur_page_offset */
98 sector_t cur_page_block
; /* Where it starts */
99 loff_t cur_page_fs_offset
; /* Offset in file */
102 * Page fetching state. These variables belong to dio_refill_pages().
104 int curr_page
; /* changes */
105 int total_pages
; /* doesn't change */
106 unsigned long curr_user_address
;/* changes */
109 * Page queue. These variables belong to dio_refill_pages() and
112 unsigned head
; /* next page to process */
113 unsigned tail
; /* last valid page + 1 */
116 /* dio_state communicated between submission path and end_io */
118 int flags
; /* doesn't change */
121 loff_t i_size
; /* i_size when submitted */
122 dio_iodone_t
*end_io
; /* IO completion function */
124 void *private; /* copy from map_bh.b_private */
126 /* BIO completion state */
127 spinlock_t bio_lock
; /* protects BIO fields below */
128 int page_errors
; /* errno from get_user_pages() */
129 int is_async
; /* is IO async ? */
130 bool defer_completion
; /* defer AIO completion to workqueue? */
131 int io_error
; /* IO error in completion path */
132 unsigned long refcount
; /* direct_io_worker() and bios */
133 struct bio
*bio_list
; /* singly linked via bi_private */
134 struct task_struct
*waiter
; /* waiting task (NULL if none) */
136 /* AIO related stuff */
137 struct kiocb
*iocb
; /* kiocb */
138 ssize_t result
; /* IO result */
141 * pages[] (and any fields placed after it) are not zeroed out at
142 * allocation time. Don't add new fields after pages[] unless you
143 * wish that they not be zeroed.
146 struct page
*pages
[DIO_PAGES
]; /* page buffer */
147 struct work_struct complete_work
;/* deferred AIO completion */
149 } ____cacheline_aligned_in_smp
;
151 static struct kmem_cache
*dio_cache __read_mostly
;
154 * How many pages are in the queue?
156 static inline unsigned dio_pages_present(struct dio_submit
*sdio
)
158 return sdio
->tail
- sdio
->head
;
162 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
164 static inline int dio_refill_pages(struct dio
*dio
, struct dio_submit
*sdio
)
169 nr_pages
= min(sdio
->total_pages
- sdio
->curr_page
, DIO_PAGES
);
170 ret
= get_user_pages_fast(
171 sdio
->curr_user_address
, /* Where from? */
172 nr_pages
, /* How many pages? */
173 dio
->rw
== READ
, /* Write to memory? */
174 &dio
->pages
[0]); /* Put results here */
176 if (ret
< 0 && sdio
->blocks_available
&& (dio
->rw
& WRITE
)) {
177 struct page
*page
= ZERO_PAGE(0);
179 * A memory fault, but the filesystem has some outstanding
180 * mapped blocks. We need to use those blocks up to avoid
181 * leaking stale data in the file.
183 if (dio
->page_errors
== 0)
184 dio
->page_errors
= ret
;
185 page_cache_get(page
);
186 dio
->pages
[0] = page
;
194 sdio
->curr_user_address
+= ret
* PAGE_SIZE
;
195 sdio
->curr_page
+= ret
;
205 * Get another userspace page. Returns an ERR_PTR on error. Pages are
206 * buffered inside the dio so that we can call get_user_pages() against a
207 * decent number of pages, less frequently. To provide nicer use of the
210 static inline struct page
*dio_get_page(struct dio
*dio
,
211 struct dio_submit
*sdio
)
213 if (dio_pages_present(sdio
) == 0) {
216 ret
= dio_refill_pages(dio
, sdio
);
219 BUG_ON(dio_pages_present(sdio
) == 0);
221 return dio
->pages
[sdio
->head
++];
225 * dio_complete() - called when all DIO BIO I/O has been completed
226 * @offset: the byte offset in the file of the completed operation
228 * This drops i_dio_count, lets interested parties know that a DIO operation
229 * has completed, and calculates the resulting return code for the operation.
231 * It lets the filesystem know if it registered an interest earlier via
232 * get_block. Pass the private field of the map buffer_head so that
233 * filesystems can use it to hold additional state between get_block calls and
236 static ssize_t
dio_complete(struct dio
*dio
, loff_t offset
, ssize_t ret
,
239 ssize_t transferred
= 0;
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.
247 if (ret
== -EIOCBQUEUED
)
251 transferred
= dio
->result
;
253 /* Check for short read case */
254 if ((dio
->rw
== READ
) && ((offset
+ transferred
) > dio
->i_size
))
255 transferred
= dio
->i_size
- offset
;
259 ret
= dio
->page_errors
;
265 if (dio
->end_io
&& dio
->result
)
266 dio
->end_io(dio
->iocb
, offset
, transferred
, dio
->private);
268 inode_dio_done(dio
->inode
);
270 if (dio
->rw
& WRITE
) {
273 err
= generic_write_sync(dio
->iocb
->ki_filp
, offset
,
275 if (err
< 0 && ret
> 0)
279 aio_complete(dio
->iocb
, ret
, 0);
282 kmem_cache_free(dio_cache
, dio
);
286 static void dio_aio_complete_work(struct work_struct
*work
)
288 struct dio
*dio
= container_of(work
, struct dio
, complete_work
);
290 dio_complete(dio
, dio
->iocb
->ki_pos
, 0, true);
293 static int dio_bio_complete(struct dio
*dio
, struct bio
*bio
);
296 * Asynchronous IO callback.
298 static void dio_bio_end_aio(struct bio
*bio
, int error
)
300 struct dio
*dio
= bio
->bi_private
;
301 unsigned long remaining
;
304 /* cleanup the bio */
305 dio_bio_complete(dio
, bio
);
307 spin_lock_irqsave(&dio
->bio_lock
, flags
);
308 remaining
= --dio
->refcount
;
309 if (remaining
== 1 && dio
->waiter
)
310 wake_up_process(dio
->waiter
);
311 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
313 if (remaining
== 0) {
314 if (dio
->result
&& dio
->defer_completion
) {
315 INIT_WORK(&dio
->complete_work
, dio_aio_complete_work
);
316 queue_work(dio
->inode
->i_sb
->s_dio_done_wq
,
317 &dio
->complete_work
);
319 dio_complete(dio
, dio
->iocb
->ki_pos
, 0, true);
325 * The BIO completion handler simply queues the BIO up for the process-context
328 * During I/O bi_private points at the dio. After I/O, bi_private is used to
329 * implement a singly-linked list of completed BIOs, at dio->bio_list.
331 static void dio_bio_end_io(struct bio
*bio
, int error
)
333 struct dio
*dio
= bio
->bi_private
;
336 spin_lock_irqsave(&dio
->bio_lock
, flags
);
337 bio
->bi_private
= dio
->bio_list
;
339 if (--dio
->refcount
== 1 && dio
->waiter
)
340 wake_up_process(dio
->waiter
);
341 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
345 * dio_end_io - handle the end io action for the given bio
346 * @bio: The direct io bio thats being completed
347 * @error: Error if there was one
349 * This is meant to be called by any filesystem that uses their own dio_submit_t
350 * so that the DIO specific endio actions are dealt with after the filesystem
351 * has done it's completion work.
353 void dio_end_io(struct bio
*bio
, int error
)
355 struct dio
*dio
= bio
->bi_private
;
358 dio_bio_end_aio(bio
, error
);
360 dio_bio_end_io(bio
, error
);
362 EXPORT_SYMBOL_GPL(dio_end_io
);
365 dio_bio_alloc(struct dio
*dio
, struct dio_submit
*sdio
,
366 struct block_device
*bdev
,
367 sector_t first_sector
, int nr_vecs
)
372 * bio_alloc() is guaranteed to return a bio when called with
373 * __GFP_WAIT and we request a valid number of vectors.
375 bio
= bio_alloc(GFP_KERNEL
, nr_vecs
);
378 bio
->bi_sector
= first_sector
;
380 bio
->bi_end_io
= dio_bio_end_aio
;
382 bio
->bi_end_io
= dio_bio_end_io
;
385 sdio
->logical_offset_in_bio
= sdio
->cur_page_fs_offset
;
389 * In the AIO read case we speculatively dirty the pages before starting IO.
390 * During IO completion, any of these pages which happen to have been written
391 * back will be redirtied by bio_check_pages_dirty().
393 * bios hold a dio reference between submit_bio and ->end_io.
395 static inline void dio_bio_submit(struct dio
*dio
, struct dio_submit
*sdio
)
397 struct bio
*bio
= sdio
->bio
;
400 bio
->bi_private
= dio
;
402 spin_lock_irqsave(&dio
->bio_lock
, flags
);
404 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
406 if (dio
->is_async
&& dio
->rw
== READ
)
407 bio_set_pages_dirty(bio
);
410 sdio
->submit_io(dio
->rw
, bio
, dio
->inode
,
411 sdio
->logical_offset_in_bio
);
413 submit_bio(dio
->rw
, bio
);
417 sdio
->logical_offset_in_bio
= 0;
421 * Release any resources in case of a failure
423 static inline void dio_cleanup(struct dio
*dio
, struct dio_submit
*sdio
)
425 while (dio_pages_present(sdio
))
426 page_cache_release(dio_get_page(dio
, sdio
));
430 * Wait for the next BIO to complete. Remove it and return it. NULL is
431 * returned once all BIOs have been completed. This must only be called once
432 * all bios have been issued so that dio->refcount can only decrease. This
433 * requires that that the caller hold a reference on the dio.
435 static struct bio
*dio_await_one(struct dio
*dio
)
438 struct bio
*bio
= NULL
;
440 spin_lock_irqsave(&dio
->bio_lock
, flags
);
443 * Wait as long as the list is empty and there are bios in flight. bio
444 * completion drops the count, maybe adds to the list, and wakes while
445 * holding the bio_lock so we don't need set_current_state()'s barrier
446 * and can call it after testing our condition.
448 while (dio
->refcount
> 1 && dio
->bio_list
== NULL
) {
449 __set_current_state(TASK_UNINTERRUPTIBLE
);
450 dio
->waiter
= current
;
451 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
453 /* wake up sets us TASK_RUNNING */
454 spin_lock_irqsave(&dio
->bio_lock
, flags
);
459 dio
->bio_list
= bio
->bi_private
;
461 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
466 * Process one completed BIO. No locks are held.
468 static int dio_bio_complete(struct dio
*dio
, struct bio
*bio
)
470 const int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
471 struct bio_vec
*bvec
;
475 dio
->io_error
= -EIO
;
477 if (dio
->is_async
&& dio
->rw
== READ
) {
478 bio_check_pages_dirty(bio
); /* transfers ownership */
480 bio_for_each_segment_all(bvec
, bio
, i
) {
481 struct page
*page
= bvec
->bv_page
;
483 if (dio
->rw
== READ
&& !PageCompound(page
))
484 set_page_dirty_lock(page
);
485 page_cache_release(page
);
489 return uptodate
? 0 : -EIO
;
493 * Wait on and process all in-flight BIOs. This must only be called once
494 * all bios have been issued so that the refcount can only decrease.
495 * This just waits for all bios to make it through dio_bio_complete. IO
496 * errors are propagated through dio->io_error and should be propagated via
499 static void dio_await_completion(struct dio
*dio
)
503 bio
= dio_await_one(dio
);
505 dio_bio_complete(dio
, bio
);
510 * A really large O_DIRECT read or write can generate a lot of BIOs. So
511 * to keep the memory consumption sane we periodically reap any completed BIOs
512 * during the BIO generation phase.
514 * This also helps to limit the peak amount of pinned userspace memory.
516 static inline int dio_bio_reap(struct dio
*dio
, struct dio_submit
*sdio
)
520 if (sdio
->reap_counter
++ >= 64) {
521 while (dio
->bio_list
) {
526 spin_lock_irqsave(&dio
->bio_lock
, flags
);
528 dio
->bio_list
= bio
->bi_private
;
529 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
530 ret2
= dio_bio_complete(dio
, bio
);
534 sdio
->reap_counter
= 0;
540 * Create workqueue for deferred direct IO completions. We allocate the
541 * workqueue when it's first needed. This avoids creating workqueue for
542 * filesystems that don't need it and also allows us to create the workqueue
543 * late enough so the we can include s_id in the name of the workqueue.
545 static int sb_init_dio_done_wq(struct super_block
*sb
)
547 struct workqueue_struct
*old
;
548 struct workqueue_struct
*wq
= alloc_workqueue("dio/%s",
554 * This has to be atomic as more DIOs can race to create the workqueue
556 old
= cmpxchg(&sb
->s_dio_done_wq
, NULL
, wq
);
557 /* Someone created workqueue before us? Free ours... */
559 destroy_workqueue(wq
);
563 static int dio_set_defer_completion(struct dio
*dio
)
565 struct super_block
*sb
= dio
->inode
->i_sb
;
567 if (dio
->defer_completion
)
569 dio
->defer_completion
= true;
570 if (!sb
->s_dio_done_wq
)
571 return sb_init_dio_done_wq(sb
);
576 * Call into the fs to map some more disk blocks. We record the current number
577 * of available blocks at sdio->blocks_available. These are in units of the
578 * fs blocksize, (1 << inode->i_blkbits).
580 * The fs is allowed to map lots of blocks at once. If it wants to do that,
581 * it uses the passed inode-relative block number as the file offset, as usual.
583 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
584 * has remaining to do. The fs should not map more than this number of blocks.
586 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
587 * indicate how much contiguous disk space has been made available at
590 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
591 * This isn't very efficient...
593 * In the case of filesystem holes: the fs may return an arbitrarily-large
594 * hole by returning an appropriate value in b_size and by clearing
595 * buffer_mapped(). However the direct-io code will only process holes one
596 * block at a time - it will repeatedly call get_block() as it walks the hole.
598 static int get_more_blocks(struct dio
*dio
, struct dio_submit
*sdio
,
599 struct buffer_head
*map_bh
)
602 sector_t fs_startblk
; /* Into file, in filesystem-sized blocks */
603 sector_t fs_endblk
; /* Into file, in filesystem-sized blocks */
604 unsigned long fs_count
; /* Number of filesystem-sized blocks */
606 unsigned int i_blkbits
= sdio
->blkbits
+ sdio
->blkfactor
;
609 * If there was a memory error and we've overwritten all the
610 * mapped blocks then we can now return that memory error
612 ret
= dio
->page_errors
;
614 BUG_ON(sdio
->block_in_file
>= sdio
->final_block_in_request
);
615 fs_startblk
= sdio
->block_in_file
>> sdio
->blkfactor
;
616 fs_endblk
= (sdio
->final_block_in_request
- 1) >>
618 fs_count
= fs_endblk
- fs_startblk
+ 1;
621 map_bh
->b_size
= fs_count
<< i_blkbits
;
624 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we
625 * forbid block creations: only overwrites are permitted.
626 * We will return early to the caller once we see an
627 * unmapped buffer head returned, and the caller will fall
628 * back to buffered I/O.
630 * Otherwise the decision is left to the get_blocks method,
631 * which may decide to handle it or also return an unmapped
634 create
= dio
->rw
& WRITE
;
635 if (dio
->flags
& DIO_SKIP_HOLES
) {
636 if (sdio
->block_in_file
< (i_size_read(dio
->inode
) >>
641 ret
= (*sdio
->get_block
)(dio
->inode
, fs_startblk
,
644 /* Store for completion */
645 dio
->private = map_bh
->b_private
;
647 if (ret
== 0 && buffer_defer_completion(map_bh
))
648 ret
= dio_set_defer_completion(dio
);
654 * There is no bio. Make one now.
656 static inline int dio_new_bio(struct dio
*dio
, struct dio_submit
*sdio
,
657 sector_t start_sector
, struct buffer_head
*map_bh
)
662 ret
= dio_bio_reap(dio
, sdio
);
665 sector
= start_sector
<< (sdio
->blkbits
- 9);
666 nr_pages
= min(sdio
->pages_in_io
, bio_get_nr_vecs(map_bh
->b_bdev
));
667 nr_pages
= min(nr_pages
, BIO_MAX_PAGES
);
668 BUG_ON(nr_pages
<= 0);
669 dio_bio_alloc(dio
, sdio
, map_bh
->b_bdev
, sector
, nr_pages
);
676 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
677 * that was successful then update final_block_in_bio and take a ref against
678 * the just-added page.
680 * Return zero on success. Non-zero means the caller needs to start a new BIO.
682 static inline int dio_bio_add_page(struct dio_submit
*sdio
)
686 ret
= bio_add_page(sdio
->bio
, sdio
->cur_page
,
687 sdio
->cur_page_len
, sdio
->cur_page_offset
);
688 if (ret
== sdio
->cur_page_len
) {
690 * Decrement count only, if we are done with this page
692 if ((sdio
->cur_page_len
+ sdio
->cur_page_offset
) == PAGE_SIZE
)
694 page_cache_get(sdio
->cur_page
);
695 sdio
->final_block_in_bio
= sdio
->cur_page_block
+
696 (sdio
->cur_page_len
>> sdio
->blkbits
);
705 * Put cur_page under IO. The section of cur_page which is described by
706 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
707 * starts on-disk at cur_page_block.
709 * We take a ref against the page here (on behalf of its presence in the bio).
711 * The caller of this function is responsible for removing cur_page from the
712 * dio, and for dropping the refcount which came from that presence.
714 static inline int dio_send_cur_page(struct dio
*dio
, struct dio_submit
*sdio
,
715 struct buffer_head
*map_bh
)
720 loff_t cur_offset
= sdio
->cur_page_fs_offset
;
721 loff_t bio_next_offset
= sdio
->logical_offset_in_bio
+
725 * See whether this new request is contiguous with the old.
727 * Btrfs cannot handle having logically non-contiguous requests
728 * submitted. For example if you have
730 * Logical: [0-4095][HOLE][8192-12287]
731 * Physical: [0-4095] [4096-8191]
733 * We cannot submit those pages together as one BIO. So if our
734 * current logical offset in the file does not equal what would
735 * be the next logical offset in the bio, submit the bio we
738 if (sdio
->final_block_in_bio
!= sdio
->cur_page_block
||
739 cur_offset
!= bio_next_offset
)
740 dio_bio_submit(dio
, sdio
);
743 if (sdio
->bio
== NULL
) {
744 ret
= dio_new_bio(dio
, sdio
, sdio
->cur_page_block
, map_bh
);
749 if (dio_bio_add_page(sdio
) != 0) {
750 dio_bio_submit(dio
, sdio
);
751 ret
= dio_new_bio(dio
, sdio
, sdio
->cur_page_block
, map_bh
);
753 ret
= dio_bio_add_page(sdio
);
762 * An autonomous function to put a chunk of a page under deferred IO.
764 * The caller doesn't actually know (or care) whether this piece of page is in
765 * a BIO, or is under IO or whatever. We just take care of all possible
766 * situations here. The separation between the logic of do_direct_IO() and
767 * that of submit_page_section() is important for clarity. Please don't break.
769 * The chunk of page starts on-disk at blocknr.
771 * We perform deferred IO, by recording the last-submitted page inside our
772 * private part of the dio structure. If possible, we just expand the IO
773 * across that page here.
775 * If that doesn't work out then we put the old page into the bio and add this
776 * page to the dio instead.
779 submit_page_section(struct dio
*dio
, struct dio_submit
*sdio
, struct page
*page
,
780 unsigned offset
, unsigned len
, sector_t blocknr
,
781 struct buffer_head
*map_bh
)
785 if (dio
->rw
& WRITE
) {
787 * Read accounting is performed in submit_bio()
789 task_io_account_write(len
);
793 * Can we just grow the current page's presence in the dio?
795 if (sdio
->cur_page
== page
&&
796 sdio
->cur_page_offset
+ sdio
->cur_page_len
== offset
&&
797 sdio
->cur_page_block
+
798 (sdio
->cur_page_len
>> sdio
->blkbits
) == blocknr
) {
799 sdio
->cur_page_len
+= len
;
804 * If there's a deferred page already there then send it.
806 if (sdio
->cur_page
) {
807 ret
= dio_send_cur_page(dio
, sdio
, map_bh
);
808 page_cache_release(sdio
->cur_page
);
809 sdio
->cur_page
= NULL
;
814 page_cache_get(page
); /* It is in dio */
815 sdio
->cur_page
= page
;
816 sdio
->cur_page_offset
= offset
;
817 sdio
->cur_page_len
= len
;
818 sdio
->cur_page_block
= blocknr
;
819 sdio
->cur_page_fs_offset
= sdio
->block_in_file
<< sdio
->blkbits
;
822 * If sdio->boundary then we want to schedule the IO now to
823 * avoid metadata seeks.
825 if (sdio
->boundary
) {
826 ret
= dio_send_cur_page(dio
, sdio
, map_bh
);
827 dio_bio_submit(dio
, sdio
);
828 page_cache_release(sdio
->cur_page
);
829 sdio
->cur_page
= NULL
;
835 * Clean any dirty buffers in the blockdev mapping which alias newly-created
836 * file blocks. Only called for S_ISREG files - blockdevs do not set
839 static void clean_blockdev_aliases(struct dio
*dio
, struct buffer_head
*map_bh
)
844 nblocks
= map_bh
->b_size
>> dio
->inode
->i_blkbits
;
846 for (i
= 0; i
< nblocks
; i
++) {
847 unmap_underlying_metadata(map_bh
->b_bdev
,
848 map_bh
->b_blocknr
+ i
);
853 * If we are not writing the entire block and get_block() allocated
854 * the block for us, we need to fill-in the unused portion of the
855 * block with zeros. This happens only if user-buffer, fileoffset or
856 * io length is not filesystem block-size multiple.
858 * `end' is zero if we're doing the start of the IO, 1 at the end of the
861 static inline void dio_zero_block(struct dio
*dio
, struct dio_submit
*sdio
,
862 int end
, struct buffer_head
*map_bh
)
864 unsigned dio_blocks_per_fs_block
;
865 unsigned this_chunk_blocks
; /* In dio_blocks */
866 unsigned this_chunk_bytes
;
869 sdio
->start_zero_done
= 1;
870 if (!sdio
->blkfactor
|| !buffer_new(map_bh
))
873 dio_blocks_per_fs_block
= 1 << sdio
->blkfactor
;
874 this_chunk_blocks
= sdio
->block_in_file
& (dio_blocks_per_fs_block
- 1);
876 if (!this_chunk_blocks
)
880 * We need to zero out part of an fs block. It is either at the
881 * beginning or the end of the fs block.
884 this_chunk_blocks
= dio_blocks_per_fs_block
- this_chunk_blocks
;
886 this_chunk_bytes
= this_chunk_blocks
<< sdio
->blkbits
;
889 if (submit_page_section(dio
, sdio
, page
, 0, this_chunk_bytes
,
890 sdio
->next_block_for_io
, map_bh
))
893 sdio
->next_block_for_io
+= this_chunk_blocks
;
897 * Walk the user pages, and the file, mapping blocks to disk and generating
898 * a sequence of (page,offset,len,block) mappings. These mappings are injected
899 * into submit_page_section(), which takes care of the next stage of submission
901 * Direct IO against a blockdev is different from a file. Because we can
902 * happily perform page-sized but 512-byte aligned IOs. It is important that
903 * blockdev IO be able to have fine alignment and large sizes.
905 * So what we do is to permit the ->get_block function to populate bh.b_size
906 * with the size of IO which is permitted at this offset and this i_blkbits.
908 * For best results, the blockdev should be set up with 512-byte i_blkbits and
909 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
910 * fine alignment but still allows this function to work in PAGE_SIZE units.
912 static int do_direct_IO(struct dio
*dio
, struct dio_submit
*sdio
,
913 struct buffer_head
*map_bh
)
915 const unsigned blkbits
= sdio
->blkbits
;
916 const unsigned blocks_per_page
= PAGE_SIZE
>> blkbits
;
918 unsigned block_in_page
;
921 /* The I/O can start at any block offset within the first page */
922 block_in_page
= sdio
->first_block_in_page
;
924 while (sdio
->block_in_file
< sdio
->final_block_in_request
) {
925 page
= dio_get_page(dio
, sdio
);
931 while (block_in_page
< blocks_per_page
) {
932 unsigned offset_in_page
= block_in_page
<< blkbits
;
933 unsigned this_chunk_bytes
; /* # of bytes mapped */
934 unsigned this_chunk_blocks
; /* # of blocks */
937 if (sdio
->blocks_available
== 0) {
939 * Need to go and map some more disk
941 unsigned long blkmask
;
942 unsigned long dio_remainder
;
944 ret
= get_more_blocks(dio
, sdio
, map_bh
);
946 page_cache_release(page
);
949 if (!buffer_mapped(map_bh
))
952 sdio
->blocks_available
=
953 map_bh
->b_size
>> sdio
->blkbits
;
954 sdio
->next_block_for_io
=
955 map_bh
->b_blocknr
<< sdio
->blkfactor
;
956 if (buffer_new(map_bh
))
957 clean_blockdev_aliases(dio
, map_bh
);
959 if (!sdio
->blkfactor
)
962 blkmask
= (1 << sdio
->blkfactor
) - 1;
963 dio_remainder
= (sdio
->block_in_file
& blkmask
);
966 * If we are at the start of IO and that IO
967 * starts partway into a fs-block,
968 * dio_remainder will be non-zero. If the IO
969 * is a read then we can simply advance the IO
970 * cursor to the first block which is to be
971 * read. But if the IO is a write and the
972 * block was newly allocated we cannot do that;
973 * the start of the fs block must be zeroed out
976 if (!buffer_new(map_bh
))
977 sdio
->next_block_for_io
+= dio_remainder
;
978 sdio
->blocks_available
-= dio_remainder
;
982 if (!buffer_mapped(map_bh
)) {
983 loff_t i_size_aligned
;
985 /* AKPM: eargh, -ENOTBLK is a hack */
986 if (dio
->rw
& WRITE
) {
987 page_cache_release(page
);
992 * Be sure to account for a partial block as the
993 * last block in the file
995 i_size_aligned
= ALIGN(i_size_read(dio
->inode
),
997 if (sdio
->block_in_file
>=
998 i_size_aligned
>> blkbits
) {
1000 page_cache_release(page
);
1003 zero_user(page
, block_in_page
<< blkbits
,
1005 sdio
->block_in_file
++;
1011 * If we're performing IO which has an alignment which
1012 * is finer than the underlying fs, go check to see if
1013 * we must zero out the start of this block.
1015 if (unlikely(sdio
->blkfactor
&& !sdio
->start_zero_done
))
1016 dio_zero_block(dio
, sdio
, 0, map_bh
);
1019 * Work out, in this_chunk_blocks, how much disk we
1020 * can add to this page
1022 this_chunk_blocks
= sdio
->blocks_available
;
1023 u
= (PAGE_SIZE
- offset_in_page
) >> blkbits
;
1024 if (this_chunk_blocks
> u
)
1025 this_chunk_blocks
= u
;
1026 u
= sdio
->final_block_in_request
- sdio
->block_in_file
;
1027 if (this_chunk_blocks
> u
)
1028 this_chunk_blocks
= u
;
1029 this_chunk_bytes
= this_chunk_blocks
<< blkbits
;
1030 BUG_ON(this_chunk_bytes
== 0);
1032 if (this_chunk_blocks
== sdio
->blocks_available
)
1033 sdio
->boundary
= buffer_boundary(map_bh
);
1034 ret
= submit_page_section(dio
, sdio
, page
,
1037 sdio
->next_block_for_io
,
1040 page_cache_release(page
);
1043 sdio
->next_block_for_io
+= this_chunk_blocks
;
1045 sdio
->block_in_file
+= this_chunk_blocks
;
1046 block_in_page
+= this_chunk_blocks
;
1047 sdio
->blocks_available
-= this_chunk_blocks
;
1049 BUG_ON(sdio
->block_in_file
> sdio
->final_block_in_request
);
1050 if (sdio
->block_in_file
== sdio
->final_block_in_request
)
1054 /* Drop the ref which was taken in get_user_pages() */
1055 page_cache_release(page
);
1062 static inline int drop_refcount(struct dio
*dio
)
1065 unsigned long flags
;
1068 * Sync will always be dropping the final ref and completing the
1069 * operation. AIO can if it was a broken operation described above or
1070 * in fact if all the bios race to complete before we get here. In
1071 * that case dio_complete() translates the EIOCBQUEUED into the proper
1072 * return code that the caller will hand to aio_complete().
1074 * This is managed by the bio_lock instead of being an atomic_t so that
1075 * completion paths can drop their ref and use the remaining count to
1076 * decide to wake the submission path atomically.
1078 spin_lock_irqsave(&dio
->bio_lock
, flags
);
1079 ret2
= --dio
->refcount
;
1080 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
1085 * This is a library function for use by filesystem drivers.
1087 * The locking rules are governed by the flags parameter:
1088 * - if the flags value contains DIO_LOCKING we use a fancy locking
1089 * scheme for dumb filesystems.
1090 * For writes this function is called under i_mutex and returns with
1091 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1092 * taken and dropped again before returning.
1093 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1094 * internal locking but rather rely on the filesystem to synchronize
1095 * direct I/O reads/writes versus each other and truncate.
1097 * To help with locking against truncate we incremented the i_dio_count
1098 * counter before starting direct I/O, and decrement it once we are done.
1099 * Truncate can wait for it to reach zero to provide exclusion. It is
1100 * expected that filesystem provide exclusion between new direct I/O
1101 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1102 * but other filesystems need to take care of this on their own.
1104 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1105 * is always inlined. Otherwise gcc is unable to split the structure into
1106 * individual fields and will generate much worse code. This is important
1107 * for the whole file.
1109 static inline ssize_t
1110 do_blockdev_direct_IO(int rw
, struct kiocb
*iocb
, struct inode
*inode
,
1111 struct block_device
*bdev
, const struct iovec
*iov
, loff_t offset
,
1112 unsigned long nr_segs
, get_block_t get_block
, dio_iodone_t end_io
,
1113 dio_submit_t submit_io
, int flags
)
1118 unsigned i_blkbits
= ACCESS_ONCE(inode
->i_blkbits
);
1119 unsigned blkbits
= i_blkbits
;
1120 unsigned blocksize_mask
= (1 << blkbits
) - 1;
1121 ssize_t retval
= -EINVAL
;
1122 loff_t end
= offset
;
1124 struct dio_submit sdio
= { 0, };
1125 unsigned long user_addr
;
1127 struct buffer_head map_bh
= { 0, };
1128 struct blk_plug plug
;
1134 * Avoid references to bdev if not absolutely needed to give
1135 * the early prefetch in the caller enough time.
1138 if (offset
& blocksize_mask
) {
1140 blkbits
= blksize_bits(bdev_logical_block_size(bdev
));
1141 blocksize_mask
= (1 << blkbits
) - 1;
1142 if (offset
& blocksize_mask
)
1146 /* Check the memory alignment. Blocks cannot straddle pages */
1147 for (seg
= 0; seg
< nr_segs
; seg
++) {
1148 addr
= (unsigned long)iov
[seg
].iov_base
;
1149 size
= iov
[seg
].iov_len
;
1151 if (unlikely((addr
& blocksize_mask
) ||
1152 (size
& blocksize_mask
))) {
1154 blkbits
= blksize_bits(
1155 bdev_logical_block_size(bdev
));
1156 blocksize_mask
= (1 << blkbits
) - 1;
1157 if ((addr
& blocksize_mask
) || (size
& blocksize_mask
))
1162 /* watch out for a 0 len io from a tricksy fs */
1163 if (rw
== READ
&& end
== offset
)
1166 dio
= kmem_cache_alloc(dio_cache
, GFP_KERNEL
);
1171 * Believe it or not, zeroing out the page array caused a .5%
1172 * performance regression in a database benchmark. So, we take
1173 * care to only zero out what's needed.
1175 memset(dio
, 0, offsetof(struct dio
, pages
));
1178 if (dio
->flags
& DIO_LOCKING
) {
1180 struct address_space
*mapping
=
1181 iocb
->ki_filp
->f_mapping
;
1183 /* will be released by direct_io_worker */
1184 mutex_lock(&inode
->i_mutex
);
1186 retval
= filemap_write_and_wait_range(mapping
, offset
,
1189 mutex_unlock(&inode
->i_mutex
);
1190 kmem_cache_free(dio_cache
, dio
);
1197 * For file extending writes updating i_size before data
1198 * writeouts complete can expose uninitialized blocks. So
1199 * even for AIO, we need to wait for i/o to complete before
1200 * returning in this case.
1202 dio
->is_async
= !is_sync_kiocb(iocb
) && !((rw
& WRITE
) &&
1203 (end
> i_size_read(inode
)));
1208 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1209 * so that we can call ->fsync.
1211 if (dio
->is_async
&& (rw
& WRITE
) &&
1212 ((iocb
->ki_filp
->f_flags
& O_DSYNC
) ||
1213 IS_SYNC(iocb
->ki_filp
->f_mapping
->host
))) {
1214 retval
= dio_set_defer_completion(dio
);
1217 * We grab i_mutex only for reads so we don't have
1218 * to release it here
1220 kmem_cache_free(dio_cache
, dio
);
1226 * Will be decremented at I/O completion time.
1228 atomic_inc(&inode
->i_dio_count
);
1231 sdio
.blkbits
= blkbits
;
1232 sdio
.blkfactor
= i_blkbits
- blkbits
;
1233 sdio
.block_in_file
= offset
>> blkbits
;
1235 sdio
.get_block
= get_block
;
1236 dio
->end_io
= end_io
;
1237 sdio
.submit_io
= submit_io
;
1238 sdio
.final_block_in_bio
= -1;
1239 sdio
.next_block_for_io
= -1;
1242 dio
->i_size
= i_size_read(inode
);
1244 spin_lock_init(&dio
->bio_lock
);
1248 * In case of non-aligned buffers, we may need 2 more
1249 * pages since we need to zero out first and last block.
1251 if (unlikely(sdio
.blkfactor
))
1252 sdio
.pages_in_io
= 2;
1254 for (seg
= 0; seg
< nr_segs
; seg
++) {
1255 user_addr
= (unsigned long)iov
[seg
].iov_base
;
1257 ((user_addr
+ iov
[seg
].iov_len
+ PAGE_SIZE
-1) /
1258 PAGE_SIZE
- user_addr
/ PAGE_SIZE
);
1261 blk_start_plug(&plug
);
1263 for (seg
= 0; seg
< nr_segs
; seg
++) {
1264 user_addr
= (unsigned long)iov
[seg
].iov_base
;
1265 sdio
.size
+= bytes
= iov
[seg
].iov_len
;
1267 /* Index into the first page of the first block */
1268 sdio
.first_block_in_page
= (user_addr
& ~PAGE_MASK
) >> blkbits
;
1269 sdio
.final_block_in_request
= sdio
.block_in_file
+
1271 /* Page fetching state */
1276 sdio
.total_pages
= 0;
1277 if (user_addr
& (PAGE_SIZE
-1)) {
1279 bytes
-= PAGE_SIZE
- (user_addr
& (PAGE_SIZE
- 1));
1281 sdio
.total_pages
+= (bytes
+ PAGE_SIZE
- 1) / PAGE_SIZE
;
1282 sdio
.curr_user_address
= user_addr
;
1284 retval
= do_direct_IO(dio
, &sdio
, &map_bh
);
1286 dio
->result
+= iov
[seg
].iov_len
-
1287 ((sdio
.final_block_in_request
- sdio
.block_in_file
) <<
1291 dio_cleanup(dio
, &sdio
);
1294 } /* end iovec loop */
1296 if (retval
== -ENOTBLK
) {
1298 * The remaining part of the request will be
1299 * be handled by buffered I/O when we return
1304 * There may be some unwritten disk at the end of a part-written
1305 * fs-block-sized block. Go zero that now.
1307 dio_zero_block(dio
, &sdio
, 1, &map_bh
);
1309 if (sdio
.cur_page
) {
1312 ret2
= dio_send_cur_page(dio
, &sdio
, &map_bh
);
1315 page_cache_release(sdio
.cur_page
);
1316 sdio
.cur_page
= NULL
;
1319 dio_bio_submit(dio
, &sdio
);
1321 blk_finish_plug(&plug
);
1324 * It is possible that, we return short IO due to end of file.
1325 * In that case, we need to release all the pages we got hold on.
1327 dio_cleanup(dio
, &sdio
);
1330 * All block lookups have been performed. For READ requests
1331 * we can let i_mutex go now that its achieved its purpose
1332 * of protecting us from looking up uninitialized blocks.
1334 if (rw
== READ
&& (dio
->flags
& DIO_LOCKING
))
1335 mutex_unlock(&dio
->inode
->i_mutex
);
1338 * The only time we want to leave bios in flight is when a successful
1339 * partial aio read or full aio write have been setup. In that case
1340 * bio completion will call aio_complete. The only time it's safe to
1341 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1342 * This had *better* be the only place that raises -EIOCBQUEUED.
1344 BUG_ON(retval
== -EIOCBQUEUED
);
1345 if (dio
->is_async
&& retval
== 0 && dio
->result
&&
1346 ((rw
== READ
) || (dio
->result
== sdio
.size
)))
1347 retval
= -EIOCBQUEUED
;
1349 if (retval
!= -EIOCBQUEUED
)
1350 dio_await_completion(dio
);
1352 if (drop_refcount(dio
) == 0) {
1353 retval
= dio_complete(dio
, offset
, retval
, false);
1355 BUG_ON(retval
!= -EIOCBQUEUED
);
1362 __blockdev_direct_IO(int rw
, struct kiocb
*iocb
, struct inode
*inode
,
1363 struct block_device
*bdev
, const struct iovec
*iov
, loff_t offset
,
1364 unsigned long nr_segs
, get_block_t get_block
, dio_iodone_t end_io
,
1365 dio_submit_t submit_io
, int flags
)
1368 * The block device state is needed in the end to finally
1369 * submit everything. Since it's likely to be cache cold
1370 * prefetch it here as first thing to hide some of the
1373 * Attempt to prefetch the pieces we likely need later.
1375 prefetch(&bdev
->bd_disk
->part_tbl
);
1376 prefetch(bdev
->bd_queue
);
1377 prefetch((char *)bdev
->bd_queue
+ SMP_CACHE_BYTES
);
1379 return do_blockdev_direct_IO(rw
, iocb
, inode
, bdev
, iov
, offset
,
1380 nr_segs
, get_block
, end_io
,
1384 EXPORT_SYMBOL(__blockdev_direct_IO
);
1386 static __init
int dio_init(void)
1388 dio_cache
= KMEM_CACHE(dio
, SLAB_PANIC
);
1391 module_init(dio_init
)