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 int io_error
; /* IO error in completion path */
131 unsigned long refcount
; /* direct_io_worker() and bios */
132 struct bio
*bio_list
; /* singly linked via bi_private */
133 struct task_struct
*waiter
; /* waiting task (NULL if none) */
135 /* AIO related stuff */
136 struct kiocb
*iocb
; /* kiocb */
137 ssize_t result
; /* IO result */
140 * pages[] (and any fields placed after it) are not zeroed out at
141 * allocation time. Don't add new fields after pages[] unless you
142 * wish that they not be zeroed.
144 struct page
*pages
[DIO_PAGES
]; /* page buffer */
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
)
165 nr_pages
= min(sdio
->total_pages
- sdio
->curr_page
, DIO_PAGES
);
166 ret
= get_user_pages_fast(
167 sdio
->curr_user_address
, /* Where from? */
168 nr_pages
, /* How many pages? */
169 dio
->rw
== READ
, /* Write to memory? */
170 &dio
->pages
[0]); /* Put results here */
172 if (ret
< 0 && sdio
->blocks_available
&& (dio
->rw
& WRITE
)) {
173 struct page
*page
= ZERO_PAGE(0);
175 * A memory fault, but the filesystem has some outstanding
176 * mapped blocks. We need to use those blocks up to avoid
177 * leaking stale data in the file.
179 if (dio
->page_errors
== 0)
180 dio
->page_errors
= ret
;
181 page_cache_get(page
);
182 dio
->pages
[0] = page
;
190 sdio
->curr_user_address
+= ret
* PAGE_SIZE
;
191 sdio
->curr_page
+= ret
;
201 * Get another userspace page. Returns an ERR_PTR on error. Pages are
202 * buffered inside the dio so that we can call get_user_pages() against a
203 * decent number of pages, less frequently. To provide nicer use of the
206 static inline struct page
*dio_get_page(struct dio
*dio
,
207 struct dio_submit
*sdio
)
209 if (dio_pages_present(sdio
) == 0) {
212 ret
= dio_refill_pages(dio
, sdio
);
215 BUG_ON(dio_pages_present(sdio
) == 0);
217 return dio
->pages
[sdio
->head
++];
221 * dio_complete() - called when all DIO BIO I/O has been completed
222 * @offset: the byte offset in the file of the completed operation
224 * This releases locks as dictated by the locking type, lets interested parties
225 * know that a DIO operation has completed, and calculates the resulting return
226 * code for the operation.
228 * It lets the filesystem know if it registered an interest earlier via
229 * get_block. Pass the private field of the map buffer_head so that
230 * filesystems can use it to hold additional state between get_block calls and
233 static ssize_t
dio_complete(struct dio
*dio
, loff_t offset
, ssize_t ret
, bool is_async
)
235 ssize_t transferred
= 0;
238 * AIO submission can race with bio completion to get here while
239 * expecting to have the last io completed by bio completion.
240 * In that case -EIOCBQUEUED is in fact not an error we want
241 * to preserve through this call.
243 if (ret
== -EIOCBQUEUED
)
247 transferred
= dio
->result
;
249 /* Check for short read case */
250 if ((dio
->rw
== READ
) && ((offset
+ transferred
) > dio
->i_size
))
251 transferred
= dio
->i_size
- offset
;
255 ret
= dio
->page_errors
;
261 if (dio
->end_io
&& dio
->result
) {
262 dio
->end_io(dio
->iocb
, offset
, transferred
,
263 dio
->private, ret
, is_async
);
265 inode_dio_done(dio
->inode
);
267 aio_complete(dio
->iocb
, ret
, 0);
273 static int dio_bio_complete(struct dio
*dio
, struct bio
*bio
);
275 * Asynchronous IO callback.
277 static void dio_bio_end_aio(struct bio
*bio
, int error
)
279 struct dio
*dio
= bio
->bi_private
;
280 unsigned long remaining
;
283 /* cleanup the bio */
284 dio_bio_complete(dio
, bio
);
286 spin_lock_irqsave(&dio
->bio_lock
, flags
);
287 remaining
= --dio
->refcount
;
288 if (remaining
== 1 && dio
->waiter
)
289 wake_up_process(dio
->waiter
);
290 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
292 if (remaining
== 0) {
293 dio_complete(dio
, dio
->iocb
->ki_pos
, 0, true);
294 kmem_cache_free(dio_cache
, dio
);
299 * The BIO completion handler simply queues the BIO up for the process-context
302 * During I/O bi_private points at the dio. After I/O, bi_private is used to
303 * implement a singly-linked list of completed BIOs, at dio->bio_list.
305 static void dio_bio_end_io(struct bio
*bio
, int error
)
307 struct dio
*dio
= bio
->bi_private
;
310 spin_lock_irqsave(&dio
->bio_lock
, flags
);
311 bio
->bi_private
= dio
->bio_list
;
313 if (--dio
->refcount
== 1 && dio
->waiter
)
314 wake_up_process(dio
->waiter
);
315 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
319 * dio_end_io - handle the end io action for the given bio
320 * @bio: The direct io bio thats being completed
321 * @error: Error if there was one
323 * This is meant to be called by any filesystem that uses their own dio_submit_t
324 * so that the DIO specific endio actions are dealt with after the filesystem
325 * has done it's completion work.
327 void dio_end_io(struct bio
*bio
, int error
)
329 struct dio
*dio
= bio
->bi_private
;
332 dio_bio_end_aio(bio
, error
);
334 dio_bio_end_io(bio
, error
);
336 EXPORT_SYMBOL_GPL(dio_end_io
);
339 dio_bio_alloc(struct dio
*dio
, struct dio_submit
*sdio
,
340 struct block_device
*bdev
,
341 sector_t first_sector
, int nr_vecs
)
346 * bio_alloc() is guaranteed to return a bio when called with
347 * __GFP_WAIT and we request a valid number of vectors.
349 bio
= bio_alloc(GFP_KERNEL
, nr_vecs
);
352 bio
->bi_sector
= first_sector
;
354 bio
->bi_end_io
= dio_bio_end_aio
;
356 bio
->bi_end_io
= dio_bio_end_io
;
359 sdio
->logical_offset_in_bio
= sdio
->cur_page_fs_offset
;
363 * In the AIO read case we speculatively dirty the pages before starting IO.
364 * During IO completion, any of these pages which happen to have been written
365 * back will be redirtied by bio_check_pages_dirty().
367 * bios hold a dio reference between submit_bio and ->end_io.
369 static inline void dio_bio_submit(struct dio
*dio
, struct dio_submit
*sdio
)
371 struct bio
*bio
= sdio
->bio
;
374 bio
->bi_private
= dio
;
376 spin_lock_irqsave(&dio
->bio_lock
, flags
);
378 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
380 if (dio
->is_async
&& dio
->rw
== READ
)
381 bio_set_pages_dirty(bio
);
384 sdio
->submit_io(dio
->rw
, bio
, dio
->inode
,
385 sdio
->logical_offset_in_bio
);
387 submit_bio(dio
->rw
, bio
);
391 sdio
->logical_offset_in_bio
= 0;
395 * Release any resources in case of a failure
397 static inline void dio_cleanup(struct dio
*dio
, struct dio_submit
*sdio
)
399 while (dio_pages_present(sdio
))
400 page_cache_release(dio_get_page(dio
, sdio
));
404 * Wait for the next BIO to complete. Remove it and return it. NULL is
405 * returned once all BIOs have been completed. This must only be called once
406 * all bios have been issued so that dio->refcount can only decrease. This
407 * requires that that the caller hold a reference on the dio.
409 static struct bio
*dio_await_one(struct dio
*dio
)
412 struct bio
*bio
= NULL
;
414 spin_lock_irqsave(&dio
->bio_lock
, flags
);
417 * Wait as long as the list is empty and there are bios in flight. bio
418 * completion drops the count, maybe adds to the list, and wakes while
419 * holding the bio_lock so we don't need set_current_state()'s barrier
420 * and can call it after testing our condition.
422 while (dio
->refcount
> 1 && dio
->bio_list
== NULL
) {
423 __set_current_state(TASK_UNINTERRUPTIBLE
);
424 dio
->waiter
= current
;
425 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
427 /* wake up sets us TASK_RUNNING */
428 spin_lock_irqsave(&dio
->bio_lock
, flags
);
433 dio
->bio_list
= bio
->bi_private
;
435 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
440 * Process one completed BIO. No locks are held.
442 static int dio_bio_complete(struct dio
*dio
, struct bio
*bio
)
444 const int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
445 struct bio_vec
*bvec
;
449 dio
->io_error
= -EIO
;
451 if (dio
->is_async
&& dio
->rw
== READ
) {
452 bio_check_pages_dirty(bio
); /* transfers ownership */
454 bio_for_each_segment_all(bvec
, bio
, i
) {
455 struct page
*page
= bvec
->bv_page
;
457 if (dio
->rw
== READ
&& !PageCompound(page
))
458 set_page_dirty_lock(page
);
459 page_cache_release(page
);
463 return uptodate
? 0 : -EIO
;
467 * Wait on and process all in-flight BIOs. This must only be called once
468 * all bios have been issued so that the refcount can only decrease.
469 * This just waits for all bios to make it through dio_bio_complete. IO
470 * errors are propagated through dio->io_error and should be propagated via
473 static void dio_await_completion(struct dio
*dio
)
477 bio
= dio_await_one(dio
);
479 dio_bio_complete(dio
, bio
);
484 * A really large O_DIRECT read or write can generate a lot of BIOs. So
485 * to keep the memory consumption sane we periodically reap any completed BIOs
486 * during the BIO generation phase.
488 * This also helps to limit the peak amount of pinned userspace memory.
490 static inline int dio_bio_reap(struct dio
*dio
, struct dio_submit
*sdio
)
494 if (sdio
->reap_counter
++ >= 64) {
495 while (dio
->bio_list
) {
500 spin_lock_irqsave(&dio
->bio_lock
, flags
);
502 dio
->bio_list
= bio
->bi_private
;
503 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
504 ret2
= dio_bio_complete(dio
, bio
);
508 sdio
->reap_counter
= 0;
514 * Call into the fs to map some more disk blocks. We record the current number
515 * of available blocks at sdio->blocks_available. These are in units of the
516 * fs blocksize, (1 << inode->i_blkbits).
518 * The fs is allowed to map lots of blocks at once. If it wants to do that,
519 * it uses the passed inode-relative block number as the file offset, as usual.
521 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
522 * has remaining to do. The fs should not map more than this number of blocks.
524 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
525 * indicate how much contiguous disk space has been made available at
528 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
529 * This isn't very efficient...
531 * In the case of filesystem holes: the fs may return an arbitrarily-large
532 * hole by returning an appropriate value in b_size and by clearing
533 * buffer_mapped(). However the direct-io code will only process holes one
534 * block at a time - it will repeatedly call get_block() as it walks the hole.
536 static int get_more_blocks(struct dio
*dio
, struct dio_submit
*sdio
,
537 struct buffer_head
*map_bh
)
540 sector_t fs_startblk
; /* Into file, in filesystem-sized blocks */
541 sector_t fs_endblk
; /* Into file, in filesystem-sized blocks */
542 unsigned long fs_count
; /* Number of filesystem-sized blocks */
544 unsigned int i_blkbits
= sdio
->blkbits
+ sdio
->blkfactor
;
547 * If there was a memory error and we've overwritten all the
548 * mapped blocks then we can now return that memory error
550 ret
= dio
->page_errors
;
552 BUG_ON(sdio
->block_in_file
>= sdio
->final_block_in_request
);
553 fs_startblk
= sdio
->block_in_file
>> sdio
->blkfactor
;
554 fs_endblk
= (sdio
->final_block_in_request
- 1) >>
556 fs_count
= fs_endblk
- fs_startblk
+ 1;
559 map_bh
->b_size
= fs_count
<< i_blkbits
;
562 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we
563 * forbid block creations: only overwrites are permitted.
564 * We will return early to the caller once we see an
565 * unmapped buffer head returned, and the caller will fall
566 * back to buffered I/O.
568 * Otherwise the decision is left to the get_blocks method,
569 * which may decide to handle it or also return an unmapped
572 create
= dio
->rw
& WRITE
;
573 if (dio
->flags
& DIO_SKIP_HOLES
) {
574 if (sdio
->block_in_file
< (i_size_read(dio
->inode
) >>
579 ret
= (*sdio
->get_block
)(dio
->inode
, fs_startblk
,
582 /* Store for completion */
583 dio
->private = map_bh
->b_private
;
589 * There is no bio. Make one now.
591 static inline int dio_new_bio(struct dio
*dio
, struct dio_submit
*sdio
,
592 sector_t start_sector
, struct buffer_head
*map_bh
)
597 ret
= dio_bio_reap(dio
, sdio
);
600 sector
= start_sector
<< (sdio
->blkbits
- 9);
601 nr_pages
= min(sdio
->pages_in_io
, bio_get_nr_vecs(map_bh
->b_bdev
));
602 nr_pages
= min(nr_pages
, BIO_MAX_PAGES
);
603 BUG_ON(nr_pages
<= 0);
604 dio_bio_alloc(dio
, sdio
, map_bh
->b_bdev
, sector
, nr_pages
);
611 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
612 * that was successful then update final_block_in_bio and take a ref against
613 * the just-added page.
615 * Return zero on success. Non-zero means the caller needs to start a new BIO.
617 static inline int dio_bio_add_page(struct dio_submit
*sdio
)
621 ret
= bio_add_page(sdio
->bio
, sdio
->cur_page
,
622 sdio
->cur_page_len
, sdio
->cur_page_offset
);
623 if (ret
== sdio
->cur_page_len
) {
625 * Decrement count only, if we are done with this page
627 if ((sdio
->cur_page_len
+ sdio
->cur_page_offset
) == PAGE_SIZE
)
629 page_cache_get(sdio
->cur_page
);
630 sdio
->final_block_in_bio
= sdio
->cur_page_block
+
631 (sdio
->cur_page_len
>> sdio
->blkbits
);
640 * Put cur_page under IO. The section of cur_page which is described by
641 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
642 * starts on-disk at cur_page_block.
644 * We take a ref against the page here (on behalf of its presence in the bio).
646 * The caller of this function is responsible for removing cur_page from the
647 * dio, and for dropping the refcount which came from that presence.
649 static inline int dio_send_cur_page(struct dio
*dio
, struct dio_submit
*sdio
,
650 struct buffer_head
*map_bh
)
655 loff_t cur_offset
= sdio
->cur_page_fs_offset
;
656 loff_t bio_next_offset
= sdio
->logical_offset_in_bio
+
660 * See whether this new request is contiguous with the old.
662 * Btrfs cannot handle having logically non-contiguous requests
663 * submitted. For example if you have
665 * Logical: [0-4095][HOLE][8192-12287]
666 * Physical: [0-4095] [4096-8191]
668 * We cannot submit those pages together as one BIO. So if our
669 * current logical offset in the file does not equal what would
670 * be the next logical offset in the bio, submit the bio we
673 if (sdio
->final_block_in_bio
!= sdio
->cur_page_block
||
674 cur_offset
!= bio_next_offset
)
675 dio_bio_submit(dio
, sdio
);
678 if (sdio
->bio
== NULL
) {
679 ret
= dio_new_bio(dio
, sdio
, sdio
->cur_page_block
, map_bh
);
684 if (dio_bio_add_page(sdio
) != 0) {
685 dio_bio_submit(dio
, sdio
);
686 ret
= dio_new_bio(dio
, sdio
, sdio
->cur_page_block
, map_bh
);
688 ret
= dio_bio_add_page(sdio
);
697 * An autonomous function to put a chunk of a page under deferred IO.
699 * The caller doesn't actually know (or care) whether this piece of page is in
700 * a BIO, or is under IO or whatever. We just take care of all possible
701 * situations here. The separation between the logic of do_direct_IO() and
702 * that of submit_page_section() is important for clarity. Please don't break.
704 * The chunk of page starts on-disk at blocknr.
706 * We perform deferred IO, by recording the last-submitted page inside our
707 * private part of the dio structure. If possible, we just expand the IO
708 * across that page here.
710 * If that doesn't work out then we put the old page into the bio and add this
711 * page to the dio instead.
714 submit_page_section(struct dio
*dio
, struct dio_submit
*sdio
, struct page
*page
,
715 unsigned offset
, unsigned len
, sector_t blocknr
,
716 struct buffer_head
*map_bh
)
720 if (dio
->rw
& WRITE
) {
722 * Read accounting is performed in submit_bio()
724 task_io_account_write(len
);
728 * Can we just grow the current page's presence in the dio?
730 if (sdio
->cur_page
== page
&&
731 sdio
->cur_page_offset
+ sdio
->cur_page_len
== offset
&&
732 sdio
->cur_page_block
+
733 (sdio
->cur_page_len
>> sdio
->blkbits
) == blocknr
) {
734 sdio
->cur_page_len
+= len
;
739 * If there's a deferred page already there then send it.
741 if (sdio
->cur_page
) {
742 ret
= dio_send_cur_page(dio
, sdio
, map_bh
);
743 page_cache_release(sdio
->cur_page
);
744 sdio
->cur_page
= NULL
;
749 page_cache_get(page
); /* It is in dio */
750 sdio
->cur_page
= page
;
751 sdio
->cur_page_offset
= offset
;
752 sdio
->cur_page_len
= len
;
753 sdio
->cur_page_block
= blocknr
;
754 sdio
->cur_page_fs_offset
= sdio
->block_in_file
<< sdio
->blkbits
;
757 * If sdio->boundary then we want to schedule the IO now to
758 * avoid metadata seeks.
760 if (sdio
->boundary
) {
761 ret
= dio_send_cur_page(dio
, sdio
, map_bh
);
763 dio_bio_submit(dio
, sdio
);
764 page_cache_release(sdio
->cur_page
);
765 sdio
->cur_page
= NULL
;
771 * Clean any dirty buffers in the blockdev mapping which alias newly-created
772 * file blocks. Only called for S_ISREG files - blockdevs do not set
775 static void clean_blockdev_aliases(struct dio
*dio
, struct buffer_head
*map_bh
)
780 nblocks
= map_bh
->b_size
>> dio
->inode
->i_blkbits
;
782 for (i
= 0; i
< nblocks
; i
++) {
783 unmap_underlying_metadata(map_bh
->b_bdev
,
784 map_bh
->b_blocknr
+ i
);
789 * If we are not writing the entire block and get_block() allocated
790 * the block for us, we need to fill-in the unused portion of the
791 * block with zeros. This happens only if user-buffer, fileoffset or
792 * io length is not filesystem block-size multiple.
794 * `end' is zero if we're doing the start of the IO, 1 at the end of the
797 static inline void dio_zero_block(struct dio
*dio
, struct dio_submit
*sdio
,
798 int end
, struct buffer_head
*map_bh
)
800 unsigned dio_blocks_per_fs_block
;
801 unsigned this_chunk_blocks
; /* In dio_blocks */
802 unsigned this_chunk_bytes
;
805 sdio
->start_zero_done
= 1;
806 if (!sdio
->blkfactor
|| !buffer_new(map_bh
))
809 dio_blocks_per_fs_block
= 1 << sdio
->blkfactor
;
810 this_chunk_blocks
= sdio
->block_in_file
& (dio_blocks_per_fs_block
- 1);
812 if (!this_chunk_blocks
)
816 * We need to zero out part of an fs block. It is either at the
817 * beginning or the end of the fs block.
820 this_chunk_blocks
= dio_blocks_per_fs_block
- this_chunk_blocks
;
822 this_chunk_bytes
= this_chunk_blocks
<< sdio
->blkbits
;
825 if (submit_page_section(dio
, sdio
, page
, 0, this_chunk_bytes
,
826 sdio
->next_block_for_io
, map_bh
))
829 sdio
->next_block_for_io
+= this_chunk_blocks
;
833 * Walk the user pages, and the file, mapping blocks to disk and generating
834 * a sequence of (page,offset,len,block) mappings. These mappings are injected
835 * into submit_page_section(), which takes care of the next stage of submission
837 * Direct IO against a blockdev is different from a file. Because we can
838 * happily perform page-sized but 512-byte aligned IOs. It is important that
839 * blockdev IO be able to have fine alignment and large sizes.
841 * So what we do is to permit the ->get_block function to populate bh.b_size
842 * with the size of IO which is permitted at this offset and this i_blkbits.
844 * For best results, the blockdev should be set up with 512-byte i_blkbits and
845 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
846 * fine alignment but still allows this function to work in PAGE_SIZE units.
848 static int do_direct_IO(struct dio
*dio
, struct dio_submit
*sdio
,
849 struct buffer_head
*map_bh
)
851 const unsigned blkbits
= sdio
->blkbits
;
852 const unsigned blocks_per_page
= PAGE_SIZE
>> blkbits
;
854 unsigned block_in_page
;
857 /* The I/O can start at any block offset within the first page */
858 block_in_page
= sdio
->first_block_in_page
;
860 while (sdio
->block_in_file
< sdio
->final_block_in_request
) {
861 page
= dio_get_page(dio
, sdio
);
867 while (block_in_page
< blocks_per_page
) {
868 unsigned offset_in_page
= block_in_page
<< blkbits
;
869 unsigned this_chunk_bytes
; /* # of bytes mapped */
870 unsigned this_chunk_blocks
; /* # of blocks */
873 if (sdio
->blocks_available
== 0) {
875 * Need to go and map some more disk
877 unsigned long blkmask
;
878 unsigned long dio_remainder
;
880 ret
= get_more_blocks(dio
, sdio
, map_bh
);
882 page_cache_release(page
);
885 if (!buffer_mapped(map_bh
))
888 sdio
->blocks_available
=
889 map_bh
->b_size
>> sdio
->blkbits
;
890 sdio
->next_block_for_io
=
891 map_bh
->b_blocknr
<< sdio
->blkfactor
;
892 if (buffer_new(map_bh
))
893 clean_blockdev_aliases(dio
, map_bh
);
895 if (!sdio
->blkfactor
)
898 blkmask
= (1 << sdio
->blkfactor
) - 1;
899 dio_remainder
= (sdio
->block_in_file
& blkmask
);
902 * If we are at the start of IO and that IO
903 * starts partway into a fs-block,
904 * dio_remainder will be non-zero. If the IO
905 * is a read then we can simply advance the IO
906 * cursor to the first block which is to be
907 * read. But if the IO is a write and the
908 * block was newly allocated we cannot do that;
909 * the start of the fs block must be zeroed out
912 if (!buffer_new(map_bh
))
913 sdio
->next_block_for_io
+= dio_remainder
;
914 sdio
->blocks_available
-= dio_remainder
;
918 if (!buffer_mapped(map_bh
)) {
919 loff_t i_size_aligned
;
921 /* AKPM: eargh, -ENOTBLK is a hack */
922 if (dio
->rw
& WRITE
) {
923 page_cache_release(page
);
928 * Be sure to account for a partial block as the
929 * last block in the file
931 i_size_aligned
= ALIGN(i_size_read(dio
->inode
),
933 if (sdio
->block_in_file
>=
934 i_size_aligned
>> blkbits
) {
936 page_cache_release(page
);
937 dio_cleanup(dio
, sdio
);
940 zero_user(page
, block_in_page
<< blkbits
,
942 sdio
->block_in_file
++;
948 * If we're performing IO which has an alignment which
949 * is finer than the underlying fs, go check to see if
950 * we must zero out the start of this block.
952 if (unlikely(sdio
->blkfactor
&& !sdio
->start_zero_done
))
953 dio_zero_block(dio
, sdio
, 0, map_bh
);
956 * Work out, in this_chunk_blocks, how much disk we
957 * can add to this page
959 this_chunk_blocks
= sdio
->blocks_available
;
960 u
= (PAGE_SIZE
- offset_in_page
) >> blkbits
;
961 if (this_chunk_blocks
> u
)
962 this_chunk_blocks
= u
;
963 u
= sdio
->final_block_in_request
- sdio
->block_in_file
;
964 if (this_chunk_blocks
> u
)
965 this_chunk_blocks
= u
;
966 this_chunk_bytes
= this_chunk_blocks
<< blkbits
;
967 BUG_ON(this_chunk_bytes
== 0);
969 if (this_chunk_blocks
== sdio
->blocks_available
)
970 sdio
->boundary
= buffer_boundary(map_bh
);
971 ret
= submit_page_section(dio
, sdio
, page
,
974 sdio
->next_block_for_io
,
977 page_cache_release(page
);
980 sdio
->next_block_for_io
+= this_chunk_blocks
;
982 sdio
->block_in_file
+= this_chunk_blocks
;
983 block_in_page
+= this_chunk_blocks
;
984 sdio
->blocks_available
-= this_chunk_blocks
;
986 BUG_ON(sdio
->block_in_file
> sdio
->final_block_in_request
);
987 if (sdio
->block_in_file
== sdio
->final_block_in_request
)
991 /* Drop the ref which was taken in get_user_pages() */
992 page_cache_release(page
);
999 static inline int drop_refcount(struct dio
*dio
)
1002 unsigned long flags
;
1005 * Sync will always be dropping the final ref and completing the
1006 * operation. AIO can if it was a broken operation described above or
1007 * in fact if all the bios race to complete before we get here. In
1008 * that case dio_complete() translates the EIOCBQUEUED into the proper
1009 * return code that the caller will hand to aio_complete().
1011 * This is managed by the bio_lock instead of being an atomic_t so that
1012 * completion paths can drop their ref and use the remaining count to
1013 * decide to wake the submission path atomically.
1015 spin_lock_irqsave(&dio
->bio_lock
, flags
);
1016 ret2
= --dio
->refcount
;
1017 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
1022 * This is a library function for use by filesystem drivers.
1024 * The locking rules are governed by the flags parameter:
1025 * - if the flags value contains DIO_LOCKING we use a fancy locking
1026 * scheme for dumb filesystems.
1027 * For writes this function is called under i_mutex and returns with
1028 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1029 * taken and dropped again before returning.
1030 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1031 * internal locking but rather rely on the filesystem to synchronize
1032 * direct I/O reads/writes versus each other and truncate.
1034 * To help with locking against truncate we incremented the i_dio_count
1035 * counter before starting direct I/O, and decrement it once we are done.
1036 * Truncate can wait for it to reach zero to provide exclusion. It is
1037 * expected that filesystem provide exclusion between new direct I/O
1038 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1039 * but other filesystems need to take care of this on their own.
1041 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1042 * is always inlined. Otherwise gcc is unable to split the structure into
1043 * individual fields and will generate much worse code. This is important
1044 * for the whole file.
1046 static inline ssize_t
1047 do_blockdev_direct_IO(int rw
, struct kiocb
*iocb
, struct inode
*inode
,
1048 struct block_device
*bdev
, const struct iovec
*iov
, loff_t offset
,
1049 unsigned long nr_segs
, get_block_t get_block
, dio_iodone_t end_io
,
1050 dio_submit_t submit_io
, int flags
)
1055 unsigned i_blkbits
= ACCESS_ONCE(inode
->i_blkbits
);
1056 unsigned blkbits
= i_blkbits
;
1057 unsigned blocksize_mask
= (1 << blkbits
) - 1;
1058 ssize_t retval
= -EINVAL
;
1059 loff_t end
= offset
;
1061 struct dio_submit sdio
= { 0, };
1062 unsigned long user_addr
;
1064 struct buffer_head map_bh
= { 0, };
1065 struct blk_plug plug
;
1071 * Avoid references to bdev if not absolutely needed to give
1072 * the early prefetch in the caller enough time.
1075 if (offset
& blocksize_mask
) {
1077 blkbits
= blksize_bits(bdev_logical_block_size(bdev
));
1078 blocksize_mask
= (1 << blkbits
) - 1;
1079 if (offset
& blocksize_mask
)
1083 /* Check the memory alignment. Blocks cannot straddle pages */
1084 for (seg
= 0; seg
< nr_segs
; seg
++) {
1085 addr
= (unsigned long)iov
[seg
].iov_base
;
1086 size
= iov
[seg
].iov_len
;
1088 if (unlikely((addr
& blocksize_mask
) ||
1089 (size
& blocksize_mask
))) {
1091 blkbits
= blksize_bits(
1092 bdev_logical_block_size(bdev
));
1093 blocksize_mask
= (1 << blkbits
) - 1;
1094 if ((addr
& blocksize_mask
) || (size
& blocksize_mask
))
1099 /* watch out for a 0 len io from a tricksy fs */
1100 if (rw
== READ
&& end
== offset
)
1103 dio
= kmem_cache_alloc(dio_cache
, GFP_KERNEL
);
1108 * Believe it or not, zeroing out the page array caused a .5%
1109 * performance regression in a database benchmark. So, we take
1110 * care to only zero out what's needed.
1112 memset(dio
, 0, offsetof(struct dio
, pages
));
1115 if (dio
->flags
& DIO_LOCKING
) {
1117 struct address_space
*mapping
=
1118 iocb
->ki_filp
->f_mapping
;
1120 /* will be released by direct_io_worker */
1121 mutex_lock(&inode
->i_mutex
);
1123 retval
= filemap_write_and_wait_range(mapping
, offset
,
1126 mutex_unlock(&inode
->i_mutex
);
1127 kmem_cache_free(dio_cache
, dio
);
1134 * Will be decremented at I/O completion time.
1136 atomic_inc(&inode
->i_dio_count
);
1139 * For file extending writes updating i_size before data
1140 * writeouts complete can expose uninitialized blocks. So
1141 * even for AIO, we need to wait for i/o to complete before
1142 * returning in this case.
1144 dio
->is_async
= !is_sync_kiocb(iocb
) && !((rw
& WRITE
) &&
1145 (end
> i_size_read(inode
)));
1151 sdio
.blkbits
= blkbits
;
1152 sdio
.blkfactor
= i_blkbits
- blkbits
;
1153 sdio
.block_in_file
= offset
>> blkbits
;
1155 sdio
.get_block
= get_block
;
1156 dio
->end_io
= end_io
;
1157 sdio
.submit_io
= submit_io
;
1158 sdio
.final_block_in_bio
= -1;
1159 sdio
.next_block_for_io
= -1;
1162 dio
->i_size
= i_size_read(inode
);
1164 spin_lock_init(&dio
->bio_lock
);
1168 * In case of non-aligned buffers, we may need 2 more
1169 * pages since we need to zero out first and last block.
1171 if (unlikely(sdio
.blkfactor
))
1172 sdio
.pages_in_io
= 2;
1174 for (seg
= 0; seg
< nr_segs
; seg
++) {
1175 user_addr
= (unsigned long)iov
[seg
].iov_base
;
1177 ((user_addr
+ iov
[seg
].iov_len
+ PAGE_SIZE
-1) /
1178 PAGE_SIZE
- user_addr
/ PAGE_SIZE
);
1181 blk_start_plug(&plug
);
1183 for (seg
= 0; seg
< nr_segs
; seg
++) {
1184 user_addr
= (unsigned long)iov
[seg
].iov_base
;
1185 sdio
.size
+= bytes
= iov
[seg
].iov_len
;
1187 /* Index into the first page of the first block */
1188 sdio
.first_block_in_page
= (user_addr
& ~PAGE_MASK
) >> blkbits
;
1189 sdio
.final_block_in_request
= sdio
.block_in_file
+
1191 /* Page fetching state */
1196 sdio
.total_pages
= 0;
1197 if (user_addr
& (PAGE_SIZE
-1)) {
1199 bytes
-= PAGE_SIZE
- (user_addr
& (PAGE_SIZE
- 1));
1201 sdio
.total_pages
+= (bytes
+ PAGE_SIZE
- 1) / PAGE_SIZE
;
1202 sdio
.curr_user_address
= user_addr
;
1204 retval
= do_direct_IO(dio
, &sdio
, &map_bh
);
1206 dio
->result
+= iov
[seg
].iov_len
-
1207 ((sdio
.final_block_in_request
- sdio
.block_in_file
) <<
1211 dio_cleanup(dio
, &sdio
);
1214 } /* end iovec loop */
1216 if (retval
== -ENOTBLK
) {
1218 * The remaining part of the request will be
1219 * be handled by buffered I/O when we return
1224 * There may be some unwritten disk at the end of a part-written
1225 * fs-block-sized block. Go zero that now.
1227 dio_zero_block(dio
, &sdio
, 1, &map_bh
);
1229 if (sdio
.cur_page
) {
1232 ret2
= dio_send_cur_page(dio
, &sdio
, &map_bh
);
1235 page_cache_release(sdio
.cur_page
);
1236 sdio
.cur_page
= NULL
;
1239 dio_bio_submit(dio
, &sdio
);
1241 blk_finish_plug(&plug
);
1244 * It is possible that, we return short IO due to end of file.
1245 * In that case, we need to release all the pages we got hold on.
1247 dio_cleanup(dio
, &sdio
);
1250 * All block lookups have been performed. For READ requests
1251 * we can let i_mutex go now that its achieved its purpose
1252 * of protecting us from looking up uninitialized blocks.
1254 if (rw
== READ
&& (dio
->flags
& DIO_LOCKING
))
1255 mutex_unlock(&dio
->inode
->i_mutex
);
1258 * The only time we want to leave bios in flight is when a successful
1259 * partial aio read or full aio write have been setup. In that case
1260 * bio completion will call aio_complete. The only time it's safe to
1261 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1262 * This had *better* be the only place that raises -EIOCBQUEUED.
1264 BUG_ON(retval
== -EIOCBQUEUED
);
1265 if (dio
->is_async
&& retval
== 0 && dio
->result
&&
1266 ((rw
== READ
) || (dio
->result
== sdio
.size
)))
1267 retval
= -EIOCBQUEUED
;
1269 if (retval
!= -EIOCBQUEUED
)
1270 dio_await_completion(dio
);
1272 if (drop_refcount(dio
) == 0) {
1273 retval
= dio_complete(dio
, offset
, retval
, false);
1274 kmem_cache_free(dio_cache
, dio
);
1276 BUG_ON(retval
!= -EIOCBQUEUED
);
1283 __blockdev_direct_IO(int rw
, struct kiocb
*iocb
, struct inode
*inode
,
1284 struct block_device
*bdev
, const struct iovec
*iov
, loff_t offset
,
1285 unsigned long nr_segs
, get_block_t get_block
, dio_iodone_t end_io
,
1286 dio_submit_t submit_io
, int flags
)
1289 * The block device state is needed in the end to finally
1290 * submit everything. Since it's likely to be cache cold
1291 * prefetch it here as first thing to hide some of the
1294 * Attempt to prefetch the pieces we likely need later.
1296 prefetch(&bdev
->bd_disk
->part_tbl
);
1297 prefetch(bdev
->bd_queue
);
1298 prefetch((char *)bdev
->bd_queue
+ SMP_CACHE_BYTES
);
1300 return do_blockdev_direct_IO(rw
, iocb
, inode
, bdev
, iov
, offset
,
1301 nr_segs
, get_block
, end_io
,
1305 EXPORT_SYMBOL(__blockdev_direct_IO
);
1307 static __init
int dio_init(void)
1309 dio_cache
= KMEM_CACHE(dio
, SLAB_PANIC
);
1312 module_init(dio_init
)