4 * Copyright (C) 2002, Linus Torvalds.
8 * 04Jul2002 akpm@zip.com.au
10 * 11Sep2002 janetinc@us.ibm.com
11 * added readv/writev support.
12 * 29Oct2002 akpm@zip.com.au
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 <asm/atomic.h>
41 * How many user pages to map in one call to get_user_pages(). This determines
42 * the size of a structure on the stack.
47 * This code generally works in units of "dio_blocks". A dio_block is
48 * somewhere between the hard sector size and the filesystem block size. it
49 * is determined on a per-invocation basis. When talking to the filesystem
50 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
51 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
52 * to bio_block quantities by shifting left by blkfactor.
54 * If blkfactor is zero then the user's request was aligned to the filesystem's
57 * lock_type is DIO_LOCKING for regular files on direct-IO-naive filesystems.
58 * This determines whether we need to do the fancy locking which prevents
59 * direct-IO from being able to read uninitialised disk blocks. If its zero
60 * (blockdev) this locking is not done, and if it is DIO_OWN_LOCKING i_mutex is
61 * not held for the entire direct write (taken briefly, initially, during a
62 * direct read though, but its never held for the duration of a direct-IO).
66 /* BIO submission state */
67 struct bio
*bio
; /* bio under assembly */
70 loff_t i_size
; /* i_size when submitted */
71 int lock_type
; /* doesn't change */
72 unsigned blkbits
; /* doesn't change */
73 unsigned blkfactor
; /* When we're using an alignment which
74 is finer than the filesystem's soft
75 blocksize, this specifies how much
76 finer. blkfactor=2 means 1/4-block
77 alignment. Does not change */
78 unsigned start_zero_done
; /* flag: sub-blocksize zeroing has
79 been performed at the start of a
81 int pages_in_io
; /* approximate total IO pages */
82 size_t size
; /* total request size (doesn't change)*/
83 sector_t block_in_file
; /* Current offset into the underlying
84 file in dio_block units. */
85 unsigned blocks_available
; /* At block_in_file. changes */
86 sector_t final_block_in_request
;/* doesn't change */
87 unsigned first_block_in_page
; /* doesn't change, Used only once */
88 int boundary
; /* prev block is at a boundary */
89 int reap_counter
; /* rate limit reaping */
90 get_block_t
*get_block
; /* block mapping function */
91 dio_iodone_t
*end_io
; /* IO completion function */
92 sector_t final_block_in_bio
; /* current final block in bio + 1 */
93 sector_t next_block_for_io
; /* next block to be put under IO,
94 in dio_blocks units */
95 struct buffer_head map_bh
; /* last get_block() result */
98 * Deferred addition of a page to the dio. These variables are
99 * private to dio_send_cur_page(), submit_page_section() and
100 * dio_bio_add_page().
102 struct page
*cur_page
; /* The page */
103 unsigned cur_page_offset
; /* Offset into it, in bytes */
104 unsigned cur_page_len
; /* Nr of bytes at cur_page_offset */
105 sector_t cur_page_block
; /* Where it starts */
108 * Page fetching state. These variables belong to dio_refill_pages().
110 int curr_page
; /* changes */
111 int total_pages
; /* doesn't change */
112 unsigned long curr_user_address
;/* changes */
115 * Page queue. These variables belong to dio_refill_pages() and
118 struct page
*pages
[DIO_PAGES
]; /* page buffer */
119 unsigned head
; /* next page to process */
120 unsigned tail
; /* last valid page + 1 */
121 int page_errors
; /* errno from get_user_pages() */
123 /* BIO completion state */
124 atomic_t refcount
; /* direct_io_worker() and bios */
125 spinlock_t bio_lock
; /* protects BIO fields below */
126 struct bio
*bio_list
; /* singly linked via bi_private */
127 struct task_struct
*waiter
; /* waiting task (NULL if none) */
129 /* AIO related stuff */
130 struct kiocb
*iocb
; /* kiocb */
131 int is_async
; /* is IO async ? */
132 int io_error
; /* IO error in completion path */
133 ssize_t result
; /* IO result */
137 * How many pages are in the queue?
139 static inline unsigned dio_pages_present(struct dio
*dio
)
141 return dio
->tail
- dio
->head
;
145 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
147 static int dio_refill_pages(struct dio
*dio
)
152 nr_pages
= min(dio
->total_pages
- dio
->curr_page
, DIO_PAGES
);
153 down_read(¤t
->mm
->mmap_sem
);
154 ret
= get_user_pages(
155 current
, /* Task for fault acounting */
156 current
->mm
, /* whose pages? */
157 dio
->curr_user_address
, /* Where from? */
158 nr_pages
, /* How many pages? */
159 dio
->rw
== READ
, /* Write to memory? */
163 up_read(¤t
->mm
->mmap_sem
);
165 if (ret
< 0 && dio
->blocks_available
&& (dio
->rw
& WRITE
)) {
166 struct page
*page
= ZERO_PAGE(dio
->curr_user_address
);
168 * A memory fault, but the filesystem has some outstanding
169 * mapped blocks. We need to use those blocks up to avoid
170 * leaking stale data in the file.
172 if (dio
->page_errors
== 0)
173 dio
->page_errors
= ret
;
174 page_cache_get(page
);
175 dio
->pages
[0] = page
;
183 dio
->curr_user_address
+= ret
* PAGE_SIZE
;
184 dio
->curr_page
+= ret
;
194 * Get another userspace page. Returns an ERR_PTR on error. Pages are
195 * buffered inside the dio so that we can call get_user_pages() against a
196 * decent number of pages, less frequently. To provide nicer use of the
199 static struct page
*dio_get_page(struct dio
*dio
)
201 if (dio_pages_present(dio
) == 0) {
204 ret
= dio_refill_pages(dio
);
207 BUG_ON(dio_pages_present(dio
) == 0);
209 return dio
->pages
[dio
->head
++];
213 * dio_complete() - called when all DIO BIO I/O has been completed
214 * @offset: the byte offset in the file of the completed operation
216 * This releases locks as dictated by the locking type, lets interested parties
217 * know that a DIO operation has completed, and calculates the resulting return
218 * code for the operation.
220 * It lets the filesystem know if it registered an interest earlier via
221 * get_block. Pass the private field of the map buffer_head so that
222 * filesystems can use it to hold additional state between get_block calls and
225 static int dio_complete(struct dio
*dio
, loff_t offset
, int ret
)
227 ssize_t transferred
= 0;
230 transferred
= dio
->result
;
232 /* Check for short read case */
233 if ((dio
->rw
== READ
) && ((offset
+ transferred
) > dio
->i_size
))
234 transferred
= dio
->i_size
- offset
;
237 if (dio
->end_io
&& dio
->result
)
238 dio
->end_io(dio
->iocb
, offset
, transferred
,
239 dio
->map_bh
.b_private
);
240 if (dio
->lock_type
== DIO_LOCKING
)
241 /* lockdep: non-owner release */
242 up_read_non_owner(&dio
->inode
->i_alloc_sem
);
245 ret
= dio
->page_errors
;
255 * Called when a BIO has been processed. If the count goes to zero then IO is
256 * complete and we can signal this to the AIO layer.
258 static void dio_complete_aio(struct dio
*dio
)
262 ret
= dio_complete(dio
, dio
->iocb
->ki_pos
, 0);
264 /* Complete AIO later if falling back to buffered i/o */
265 if (dio
->result
== dio
->size
||
266 ((dio
->rw
== READ
) && dio
->result
)) {
267 aio_complete(dio
->iocb
, ret
, 0);
272 static int dio_bio_complete(struct dio
*dio
, struct bio
*bio
);
274 * Asynchronous IO callback.
276 static int dio_bio_end_aio(struct bio
*bio
, unsigned int bytes_done
, int error
)
278 struct dio
*dio
= bio
->bi_private
;
279 int waiter_holds_ref
= 0;
285 /* cleanup the bio */
286 dio_bio_complete(dio
, bio
);
288 waiter_holds_ref
= !!dio
->waiter
;
289 remaining
= atomic_sub_return(1, (&dio
->refcount
));
290 if (remaining
== 1 && waiter_holds_ref
)
291 wake_up_process(dio
->waiter
);
294 dio_complete_aio(dio
);
300 * The BIO completion handler simply queues the BIO up for the process-context
303 * During I/O bi_private points at the dio. After I/O, bi_private is used to
304 * implement a singly-linked list of completed BIOs, at dio->bio_list.
306 static int dio_bio_end_io(struct bio
*bio
, unsigned int bytes_done
, int error
)
308 struct dio
*dio
= bio
->bi_private
;
314 spin_lock_irqsave(&dio
->bio_lock
, flags
);
315 bio
->bi_private
= dio
->bio_list
;
317 if ((atomic_sub_return(1, &dio
->refcount
) == 1) && dio
->waiter
)
318 wake_up_process(dio
->waiter
);
319 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
324 dio_bio_alloc(struct dio
*dio
, struct block_device
*bdev
,
325 sector_t first_sector
, int nr_vecs
)
329 bio
= bio_alloc(GFP_KERNEL
, nr_vecs
);
334 bio
->bi_sector
= first_sector
;
336 bio
->bi_end_io
= dio_bio_end_aio
;
338 bio
->bi_end_io
= dio_bio_end_io
;
345 * In the AIO read case we speculatively dirty the pages before starting IO.
346 * During IO completion, any of these pages which happen to have been written
347 * back will be redirtied by bio_check_pages_dirty().
349 * bios hold a dio reference between submit_bio and ->end_io.
351 static void dio_bio_submit(struct dio
*dio
)
353 struct bio
*bio
= dio
->bio
;
355 bio
->bi_private
= dio
;
356 atomic_inc(&dio
->refcount
);
357 if (dio
->is_async
&& dio
->rw
== READ
)
358 bio_set_pages_dirty(bio
);
359 submit_bio(dio
->rw
, bio
);
366 * Release any resources in case of a failure
368 static void dio_cleanup(struct dio
*dio
)
370 while (dio_pages_present(dio
))
371 page_cache_release(dio_get_page(dio
));
374 static int wait_for_more_bios(struct dio
*dio
)
376 assert_spin_locked(&dio
->bio_lock
);
378 return (atomic_read(&dio
->refcount
) > 1) && (dio
->bio_list
== NULL
);
382 * Wait for the next BIO to complete. Remove it and return it. NULL is
383 * returned once all BIOs have been completed. This must only be called once
384 * all bios have been issued so that dio->refcount can only decrease. This
385 * requires that that the caller hold a reference on the dio.
387 static struct bio
*dio_await_one(struct dio
*dio
)
390 struct bio
*bio
= NULL
;
392 spin_lock_irqsave(&dio
->bio_lock
, flags
);
393 while (wait_for_more_bios(dio
)) {
394 set_current_state(TASK_UNINTERRUPTIBLE
);
395 if (wait_for_more_bios(dio
)) {
396 dio
->waiter
= current
;
397 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
399 spin_lock_irqsave(&dio
->bio_lock
, flags
);
402 set_current_state(TASK_RUNNING
);
406 dio
->bio_list
= bio
->bi_private
;
408 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
413 * Process one completed BIO. No locks are held.
415 static int dio_bio_complete(struct dio
*dio
, struct bio
*bio
)
417 const int uptodate
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
);
418 struct bio_vec
*bvec
= bio
->bi_io_vec
;
422 dio
->io_error
= -EIO
;
424 if (dio
->is_async
&& dio
->rw
== READ
) {
425 bio_check_pages_dirty(bio
); /* transfers ownership */
427 for (page_no
= 0; page_no
< bio
->bi_vcnt
; page_no
++) {
428 struct page
*page
= bvec
[page_no
].bv_page
;
430 if (dio
->rw
== READ
&& !PageCompound(page
))
431 set_page_dirty_lock(page
);
432 page_cache_release(page
);
436 return uptodate
? 0 : -EIO
;
440 * Wait on and process all in-flight BIOs. This must only be called once
441 * all bios have been issued so that the refcount can only decrease.
442 * This just waits for all bios to make it through dio_bio_complete. IO
443 * errors are propogated through dio->io_error and should be propogated via
446 static void dio_await_completion(struct dio
*dio
)
450 bio
= dio_await_one(dio
);
452 dio_bio_complete(dio
, bio
);
457 * A really large O_DIRECT read or write can generate a lot of BIOs. So
458 * to keep the memory consumption sane we periodically reap any completed BIOs
459 * during the BIO generation phase.
461 * This also helps to limit the peak amount of pinned userspace memory.
463 static int dio_bio_reap(struct dio
*dio
)
467 if (dio
->reap_counter
++ >= 64) {
468 while (dio
->bio_list
) {
473 spin_lock_irqsave(&dio
->bio_lock
, flags
);
475 dio
->bio_list
= bio
->bi_private
;
476 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
477 ret2
= dio_bio_complete(dio
, bio
);
481 dio
->reap_counter
= 0;
487 * Call into the fs to map some more disk blocks. We record the current number
488 * of available blocks at dio->blocks_available. These are in units of the
489 * fs blocksize, (1 << inode->i_blkbits).
491 * The fs is allowed to map lots of blocks at once. If it wants to do that,
492 * it uses the passed inode-relative block number as the file offset, as usual.
494 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
495 * has remaining to do. The fs should not map more than this number of blocks.
497 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
498 * indicate how much contiguous disk space has been made available at
501 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
502 * This isn't very efficient...
504 * In the case of filesystem holes: the fs may return an arbitrarily-large
505 * hole by returning an appropriate value in b_size and by clearing
506 * buffer_mapped(). However the direct-io code will only process holes one
507 * block at a time - it will repeatedly call get_block() as it walks the hole.
509 static int get_more_blocks(struct dio
*dio
)
512 struct buffer_head
*map_bh
= &dio
->map_bh
;
513 sector_t fs_startblk
; /* Into file, in filesystem-sized blocks */
514 unsigned long fs_count
; /* Number of filesystem-sized blocks */
515 unsigned long dio_count
;/* Number of dio_block-sized blocks */
516 unsigned long blkmask
;
520 * If there was a memory error and we've overwritten all the
521 * mapped blocks then we can now return that memory error
523 ret
= dio
->page_errors
;
525 BUG_ON(dio
->block_in_file
>= dio
->final_block_in_request
);
526 fs_startblk
= dio
->block_in_file
>> dio
->blkfactor
;
527 dio_count
= dio
->final_block_in_request
- dio
->block_in_file
;
528 fs_count
= dio_count
>> dio
->blkfactor
;
529 blkmask
= (1 << dio
->blkfactor
) - 1;
530 if (dio_count
& blkmask
)
534 map_bh
->b_size
= fs_count
<< dio
->inode
->i_blkbits
;
536 create
= dio
->rw
& WRITE
;
537 if (dio
->lock_type
== DIO_LOCKING
) {
538 if (dio
->block_in_file
< (i_size_read(dio
->inode
) >>
541 } else if (dio
->lock_type
== DIO_NO_LOCKING
) {
546 * For writes inside i_size we forbid block creations: only
547 * overwrites are permitted. We fall back to buffered writes
548 * at a higher level for inside-i_size block-instantiating
551 ret
= (*dio
->get_block
)(dio
->inode
, fs_startblk
,
558 * There is no bio. Make one now.
560 static int dio_new_bio(struct dio
*dio
, sector_t start_sector
)
565 ret
= dio_bio_reap(dio
);
568 sector
= start_sector
<< (dio
->blkbits
- 9);
569 nr_pages
= min(dio
->pages_in_io
, bio_get_nr_vecs(dio
->map_bh
.b_bdev
));
570 BUG_ON(nr_pages
<= 0);
571 ret
= dio_bio_alloc(dio
, dio
->map_bh
.b_bdev
, sector
, nr_pages
);
578 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
579 * that was successful then update final_block_in_bio and take a ref against
580 * the just-added page.
582 * Return zero on success. Non-zero means the caller needs to start a new BIO.
584 static int dio_bio_add_page(struct dio
*dio
)
588 ret
= bio_add_page(dio
->bio
, dio
->cur_page
,
589 dio
->cur_page_len
, dio
->cur_page_offset
);
590 if (ret
== dio
->cur_page_len
) {
592 * Decrement count only, if we are done with this page
594 if ((dio
->cur_page_len
+ dio
->cur_page_offset
) == PAGE_SIZE
)
596 page_cache_get(dio
->cur_page
);
597 dio
->final_block_in_bio
= dio
->cur_page_block
+
598 (dio
->cur_page_len
>> dio
->blkbits
);
607 * Put cur_page under IO. The section of cur_page which is described by
608 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
609 * starts on-disk at cur_page_block.
611 * We take a ref against the page here (on behalf of its presence in the bio).
613 * The caller of this function is responsible for removing cur_page from the
614 * dio, and for dropping the refcount which came from that presence.
616 static int dio_send_cur_page(struct dio
*dio
)
622 * See whether this new request is contiguous with the old
624 if (dio
->final_block_in_bio
!= dio
->cur_page_block
)
627 * Submit now if the underlying fs is about to perform a
634 if (dio
->bio
== NULL
) {
635 ret
= dio_new_bio(dio
, dio
->cur_page_block
);
640 if (dio_bio_add_page(dio
) != 0) {
642 ret
= dio_new_bio(dio
, dio
->cur_page_block
);
644 ret
= dio_bio_add_page(dio
);
653 * An autonomous function to put a chunk of a page under deferred IO.
655 * The caller doesn't actually know (or care) whether this piece of page is in
656 * a BIO, or is under IO or whatever. We just take care of all possible
657 * situations here. The separation between the logic of do_direct_IO() and
658 * that of submit_page_section() is important for clarity. Please don't break.
660 * The chunk of page starts on-disk at blocknr.
662 * We perform deferred IO, by recording the last-submitted page inside our
663 * private part of the dio structure. If possible, we just expand the IO
664 * across that page here.
666 * If that doesn't work out then we put the old page into the bio and add this
667 * page to the dio instead.
670 submit_page_section(struct dio
*dio
, struct page
*page
,
671 unsigned offset
, unsigned len
, sector_t blocknr
)
675 if (dio
->rw
& WRITE
) {
677 * Read accounting is performed in submit_bio()
679 task_io_account_write(len
);
683 * Can we just grow the current page's presence in the dio?
685 if ( (dio
->cur_page
== page
) &&
686 (dio
->cur_page_offset
+ dio
->cur_page_len
== offset
) &&
687 (dio
->cur_page_block
+
688 (dio
->cur_page_len
>> dio
->blkbits
) == blocknr
)) {
689 dio
->cur_page_len
+= len
;
692 * If dio->boundary then we want to schedule the IO now to
693 * avoid metadata seeks.
696 ret
= dio_send_cur_page(dio
);
697 page_cache_release(dio
->cur_page
);
698 dio
->cur_page
= NULL
;
704 * If there's a deferred page already there then send it.
707 ret
= dio_send_cur_page(dio
);
708 page_cache_release(dio
->cur_page
);
709 dio
->cur_page
= NULL
;
714 page_cache_get(page
); /* It is in dio */
715 dio
->cur_page
= page
;
716 dio
->cur_page_offset
= offset
;
717 dio
->cur_page_len
= len
;
718 dio
->cur_page_block
= blocknr
;
724 * Clean any dirty buffers in the blockdev mapping which alias newly-created
725 * file blocks. Only called for S_ISREG files - blockdevs do not set
728 static void clean_blockdev_aliases(struct dio
*dio
)
733 nblocks
= dio
->map_bh
.b_size
>> dio
->inode
->i_blkbits
;
735 for (i
= 0; i
< nblocks
; i
++) {
736 unmap_underlying_metadata(dio
->map_bh
.b_bdev
,
737 dio
->map_bh
.b_blocknr
+ i
);
742 * If we are not writing the entire block and get_block() allocated
743 * the block for us, we need to fill-in the unused portion of the
744 * block with zeros. This happens only if user-buffer, fileoffset or
745 * io length is not filesystem block-size multiple.
747 * `end' is zero if we're doing the start of the IO, 1 at the end of the
750 static void dio_zero_block(struct dio
*dio
, int end
)
752 unsigned dio_blocks_per_fs_block
;
753 unsigned this_chunk_blocks
; /* In dio_blocks */
754 unsigned this_chunk_bytes
;
757 dio
->start_zero_done
= 1;
758 if (!dio
->blkfactor
|| !buffer_new(&dio
->map_bh
))
761 dio_blocks_per_fs_block
= 1 << dio
->blkfactor
;
762 this_chunk_blocks
= dio
->block_in_file
& (dio_blocks_per_fs_block
- 1);
764 if (!this_chunk_blocks
)
768 * We need to zero out part of an fs block. It is either at the
769 * beginning or the end of the fs block.
772 this_chunk_blocks
= dio_blocks_per_fs_block
- this_chunk_blocks
;
774 this_chunk_bytes
= this_chunk_blocks
<< dio
->blkbits
;
776 page
= ZERO_PAGE(dio
->curr_user_address
);
777 if (submit_page_section(dio
, page
, 0, this_chunk_bytes
,
778 dio
->next_block_for_io
))
781 dio
->next_block_for_io
+= this_chunk_blocks
;
785 * Walk the user pages, and the file, mapping blocks to disk and generating
786 * a sequence of (page,offset,len,block) mappings. These mappings are injected
787 * into submit_page_section(), which takes care of the next stage of submission
789 * Direct IO against a blockdev is different from a file. Because we can
790 * happily perform page-sized but 512-byte aligned IOs. It is important that
791 * blockdev IO be able to have fine alignment and large sizes.
793 * So what we do is to permit the ->get_block function to populate bh.b_size
794 * with the size of IO which is permitted at this offset and this i_blkbits.
796 * For best results, the blockdev should be set up with 512-byte i_blkbits and
797 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
798 * fine alignment but still allows this function to work in PAGE_SIZE units.
800 static int do_direct_IO(struct dio
*dio
)
802 const unsigned blkbits
= dio
->blkbits
;
803 const unsigned blocks_per_page
= PAGE_SIZE
>> blkbits
;
805 unsigned block_in_page
;
806 struct buffer_head
*map_bh
= &dio
->map_bh
;
809 /* The I/O can start at any block offset within the first page */
810 block_in_page
= dio
->first_block_in_page
;
812 while (dio
->block_in_file
< dio
->final_block_in_request
) {
813 page
= dio_get_page(dio
);
819 while (block_in_page
< blocks_per_page
) {
820 unsigned offset_in_page
= block_in_page
<< blkbits
;
821 unsigned this_chunk_bytes
; /* # of bytes mapped */
822 unsigned this_chunk_blocks
; /* # of blocks */
825 if (dio
->blocks_available
== 0) {
827 * Need to go and map some more disk
829 unsigned long blkmask
;
830 unsigned long dio_remainder
;
832 ret
= get_more_blocks(dio
);
834 page_cache_release(page
);
837 if (!buffer_mapped(map_bh
))
840 dio
->blocks_available
=
841 map_bh
->b_size
>> dio
->blkbits
;
842 dio
->next_block_for_io
=
843 map_bh
->b_blocknr
<< dio
->blkfactor
;
844 if (buffer_new(map_bh
))
845 clean_blockdev_aliases(dio
);
850 blkmask
= (1 << dio
->blkfactor
) - 1;
851 dio_remainder
= (dio
->block_in_file
& blkmask
);
854 * If we are at the start of IO and that IO
855 * starts partway into a fs-block,
856 * dio_remainder will be non-zero. If the IO
857 * is a read then we can simply advance the IO
858 * cursor to the first block which is to be
859 * read. But if the IO is a write and the
860 * block was newly allocated we cannot do that;
861 * the start of the fs block must be zeroed out
864 if (!buffer_new(map_bh
))
865 dio
->next_block_for_io
+= dio_remainder
;
866 dio
->blocks_available
-= dio_remainder
;
870 if (!buffer_mapped(map_bh
)) {
872 loff_t i_size_aligned
;
874 /* AKPM: eargh, -ENOTBLK is a hack */
875 if (dio
->rw
& WRITE
) {
876 page_cache_release(page
);
881 * Be sure to account for a partial block as the
882 * last block in the file
884 i_size_aligned
= ALIGN(i_size_read(dio
->inode
),
886 if (dio
->block_in_file
>=
887 i_size_aligned
>> blkbits
) {
889 page_cache_release(page
);
892 kaddr
= kmap_atomic(page
, KM_USER0
);
893 memset(kaddr
+ (block_in_page
<< blkbits
),
895 flush_dcache_page(page
);
896 kunmap_atomic(kaddr
, KM_USER0
);
897 dio
->block_in_file
++;
903 * If we're performing IO which has an alignment which
904 * is finer than the underlying fs, go check to see if
905 * we must zero out the start of this block.
907 if (unlikely(dio
->blkfactor
&& !dio
->start_zero_done
))
908 dio_zero_block(dio
, 0);
911 * Work out, in this_chunk_blocks, how much disk we
912 * can add to this page
914 this_chunk_blocks
= dio
->blocks_available
;
915 u
= (PAGE_SIZE
- offset_in_page
) >> blkbits
;
916 if (this_chunk_blocks
> u
)
917 this_chunk_blocks
= u
;
918 u
= dio
->final_block_in_request
- dio
->block_in_file
;
919 if (this_chunk_blocks
> u
)
920 this_chunk_blocks
= u
;
921 this_chunk_bytes
= this_chunk_blocks
<< blkbits
;
922 BUG_ON(this_chunk_bytes
== 0);
924 dio
->boundary
= buffer_boundary(map_bh
);
925 ret
= submit_page_section(dio
, page
, offset_in_page
,
926 this_chunk_bytes
, dio
->next_block_for_io
);
928 page_cache_release(page
);
931 dio
->next_block_for_io
+= this_chunk_blocks
;
933 dio
->block_in_file
+= this_chunk_blocks
;
934 block_in_page
+= this_chunk_blocks
;
935 dio
->blocks_available
-= this_chunk_blocks
;
937 BUG_ON(dio
->block_in_file
> dio
->final_block_in_request
);
938 if (dio
->block_in_file
== dio
->final_block_in_request
)
942 /* Drop the ref which was taken in get_user_pages() */
943 page_cache_release(page
);
951 * Releases both i_mutex and i_alloc_sem
954 direct_io_worker(int rw
, struct kiocb
*iocb
, struct inode
*inode
,
955 const struct iovec
*iov
, loff_t offset
, unsigned long nr_segs
,
956 unsigned blkbits
, get_block_t get_block
, dio_iodone_t end_io
,
959 unsigned long user_addr
;
968 dio
->blkbits
= blkbits
;
969 dio
->blkfactor
= inode
->i_blkbits
- blkbits
;
970 dio
->start_zero_done
= 0;
972 dio
->block_in_file
= offset
>> blkbits
;
973 dio
->blocks_available
= 0;
974 dio
->cur_page
= NULL
;
977 dio
->reap_counter
= 0;
978 dio
->get_block
= get_block
;
979 dio
->end_io
= end_io
;
980 dio
->map_bh
.b_private
= NULL
;
981 dio
->final_block_in_bio
= -1;
982 dio
->next_block_for_io
= -1;
984 dio
->page_errors
= 0;
988 dio
->i_size
= i_size_read(inode
);
990 atomic_set(&dio
->refcount
, 1);
991 spin_lock_init(&dio
->bio_lock
);
992 dio
->bio_list
= NULL
;
996 * In case of non-aligned buffers, we may need 2 more
997 * pages since we need to zero out first and last block.
999 if (unlikely(dio
->blkfactor
))
1000 dio
->pages_in_io
= 2;
1002 dio
->pages_in_io
= 0;
1004 for (seg
= 0; seg
< nr_segs
; seg
++) {
1005 user_addr
= (unsigned long)iov
[seg
].iov_base
;
1007 ((user_addr
+iov
[seg
].iov_len
+PAGE_SIZE
-1)/PAGE_SIZE
1008 - user_addr
/PAGE_SIZE
);
1011 for (seg
= 0; seg
< nr_segs
; seg
++) {
1012 user_addr
= (unsigned long)iov
[seg
].iov_base
;
1013 dio
->size
+= bytes
= iov
[seg
].iov_len
;
1015 /* Index into the first page of the first block */
1016 dio
->first_block_in_page
= (user_addr
& ~PAGE_MASK
) >> blkbits
;
1017 dio
->final_block_in_request
= dio
->block_in_file
+
1019 /* Page fetching state */
1024 dio
->total_pages
= 0;
1025 if (user_addr
& (PAGE_SIZE
-1)) {
1027 bytes
-= PAGE_SIZE
- (user_addr
& (PAGE_SIZE
- 1));
1029 dio
->total_pages
+= (bytes
+ PAGE_SIZE
- 1) / PAGE_SIZE
;
1030 dio
->curr_user_address
= user_addr
;
1032 ret
= do_direct_IO(dio
);
1034 dio
->result
+= iov
[seg
].iov_len
-
1035 ((dio
->final_block_in_request
- dio
->block_in_file
) <<
1042 } /* end iovec loop */
1044 if (ret
== -ENOTBLK
&& (rw
& WRITE
)) {
1046 * The remaining part of the request will be
1047 * be handled by buffered I/O when we return
1052 * There may be some unwritten disk at the end of a part-written
1053 * fs-block-sized block. Go zero that now.
1055 dio_zero_block(dio
, 1);
1057 if (dio
->cur_page
) {
1058 ret2
= dio_send_cur_page(dio
);
1061 page_cache_release(dio
->cur_page
);
1062 dio
->cur_page
= NULL
;
1065 dio_bio_submit(dio
);
1067 /* All IO is now issued, send it on its way */
1068 blk_run_address_space(inode
->i_mapping
);
1071 * It is possible that, we return short IO due to end of file.
1072 * In that case, we need to release all the pages we got hold on.
1077 * All block lookups have been performed. For READ requests
1078 * we can let i_mutex go now that its achieved its purpose
1079 * of protecting us from looking up uninitialized blocks.
1081 if ((rw
== READ
) && (dio
->lock_type
== DIO_LOCKING
))
1082 mutex_unlock(&dio
->inode
->i_mutex
);
1085 * OK, all BIOs are submitted, so we can decrement bio_count to truly
1086 * reflect the number of to-be-processed BIOs.
1088 if (dio
->is_async
) {
1089 int should_wait
= 0;
1091 if (dio
->result
< dio
->size
&& (rw
& WRITE
)) {
1092 dio
->waiter
= current
;
1099 dio_await_completion(dio
);
1101 /* this can free the dio */
1102 if (atomic_dec_and_test(&dio
->refcount
))
1103 dio_complete_aio(dio
);
1108 dio_await_completion(dio
);
1110 ret
= dio_complete(dio
, offset
, ret
);
1112 /* We could have also come here on an AIO file extend */
1113 if (!is_sync_kiocb(iocb
) && (rw
& WRITE
) &&
1114 ret
>= 0 && dio
->result
== dio
->size
)
1116 * For AIO writes where we have completed the
1117 * i/o, we have to mark the the aio complete.
1119 aio_complete(iocb
, ret
, 0);
1121 if (atomic_dec_and_test(&dio
->refcount
))
1130 * This is a library function for use by filesystem drivers.
1131 * The locking rules are governed by the dio_lock_type parameter.
1133 * DIO_NO_LOCKING (no locking, for raw block device access)
1134 * For writes, i_mutex is not held on entry; it is never taken.
1136 * DIO_LOCKING (simple locking for regular files)
1137 * For writes we are called under i_mutex and return with i_mutex held, even
1138 * though it is internally dropped.
1139 * For reads, i_mutex is not held on entry, but it is taken and dropped before
1142 * DIO_OWN_LOCKING (filesystem provides synchronisation and handling of
1143 * uninitialised data, allowing parallel direct readers and writers)
1144 * For writes we are called without i_mutex, return without it, never touch it.
1145 * For reads we are called under i_mutex and return with i_mutex held, even
1146 * though it may be internally dropped.
1148 * Additional i_alloc_sem locking requirements described inline below.
1151 __blockdev_direct_IO(int rw
, struct kiocb
*iocb
, struct inode
*inode
,
1152 struct block_device
*bdev
, const struct iovec
*iov
, loff_t offset
,
1153 unsigned long nr_segs
, get_block_t get_block
, dio_iodone_t end_io
,
1159 unsigned blkbits
= inode
->i_blkbits
;
1160 unsigned bdev_blkbits
= 0;
1161 unsigned blocksize_mask
= (1 << blkbits
) - 1;
1162 ssize_t retval
= -EINVAL
;
1163 loff_t end
= offset
;
1165 int release_i_mutex
= 0;
1166 int acquire_i_mutex
= 0;
1172 bdev_blkbits
= blksize_bits(bdev_hardsect_size(bdev
));
1174 if (offset
& blocksize_mask
) {
1176 blkbits
= bdev_blkbits
;
1177 blocksize_mask
= (1 << blkbits
) - 1;
1178 if (offset
& blocksize_mask
)
1182 /* Check the memory alignment. Blocks cannot straddle pages */
1183 for (seg
= 0; seg
< nr_segs
; seg
++) {
1184 addr
= (unsigned long)iov
[seg
].iov_base
;
1185 size
= iov
[seg
].iov_len
;
1187 if ((addr
& blocksize_mask
) || (size
& blocksize_mask
)) {
1189 blkbits
= bdev_blkbits
;
1190 blocksize_mask
= (1 << blkbits
) - 1;
1191 if ((addr
& blocksize_mask
) || (size
& blocksize_mask
))
1196 dio
= kmalloc(sizeof(*dio
), GFP_KERNEL
);
1202 * For block device access DIO_NO_LOCKING is used,
1203 * neither readers nor writers do any locking at all
1204 * For regular files using DIO_LOCKING,
1205 * readers need to grab i_mutex and i_alloc_sem
1206 * writers need to grab i_alloc_sem only (i_mutex is already held)
1207 * For regular files using DIO_OWN_LOCKING,
1208 * neither readers nor writers take any locks here
1210 dio
->lock_type
= dio_lock_type
;
1211 if (dio_lock_type
!= DIO_NO_LOCKING
) {
1212 /* watch out for a 0 len io from a tricksy fs */
1213 if (rw
== READ
&& end
> offset
) {
1214 struct address_space
*mapping
;
1216 mapping
= iocb
->ki_filp
->f_mapping
;
1217 if (dio_lock_type
!= DIO_OWN_LOCKING
) {
1218 mutex_lock(&inode
->i_mutex
);
1219 release_i_mutex
= 1;
1222 retval
= filemap_write_and_wait_range(mapping
, offset
,
1229 if (dio_lock_type
== DIO_OWN_LOCKING
) {
1230 mutex_unlock(&inode
->i_mutex
);
1231 acquire_i_mutex
= 1;
1235 if (dio_lock_type
== DIO_LOCKING
)
1236 /* lockdep: not the owner will release it */
1237 down_read_non_owner(&inode
->i_alloc_sem
);
1241 * For file extending writes updating i_size before data
1242 * writeouts complete can expose uninitialized blocks. So
1243 * even for AIO, we need to wait for i/o to complete before
1244 * returning in this case.
1246 dio
->is_async
= !is_sync_kiocb(iocb
) && !((rw
& WRITE
) &&
1247 (end
> i_size_read(inode
)));
1249 retval
= direct_io_worker(rw
, iocb
, inode
, iov
, offset
,
1250 nr_segs
, blkbits
, get_block
, end_io
, dio
);
1252 if (rw
== READ
&& dio_lock_type
== DIO_LOCKING
)
1253 release_i_mutex
= 0;
1256 if (release_i_mutex
)
1257 mutex_unlock(&inode
->i_mutex
);
1258 else if (acquire_i_mutex
)
1259 mutex_lock(&inode
->i_mutex
);
1262 EXPORT_SYMBOL(__blockdev_direct_IO
);