4 * Copyright (C) 2002, Linus Torvalds.
8 * 04Jul2002 Andrew Morton
10 * 11Sep2002 janetinc@us.ibm.com
11 * added readv/writev support.
12 * 29Oct2002 Andrew Morton
13 * rewrote bio_add_page() support.
14 * 30Oct2002 pbadari@us.ibm.com
15 * added support for non-aligned IO.
16 * 06Nov2002 pbadari@us.ibm.com
17 * added asynchronous IO support.
18 * 21Jul2003 nathans@sgi.com
19 * added IO completion notifier.
22 #include <linux/kernel.h>
23 #include <linux/module.h>
24 #include <linux/types.h>
27 #include <linux/slab.h>
28 #include <linux/highmem.h>
29 #include <linux/pagemap.h>
30 #include <linux/task_io_accounting_ops.h>
31 #include <linux/bio.h>
32 #include <linux/wait.h>
33 #include <linux/err.h>
34 #include <linux/blkdev.h>
35 #include <linux/buffer_head.h>
36 #include <linux/rwsem.h>
37 #include <linux/uio.h>
38 #include <linux/atomic.h>
39 #include <linux/prefetch.h>
42 * How many user pages to map in one call to get_user_pages(). This determines
43 * the size of a structure in the slab cache
48 * Flags for dio_complete()
50 #define DIO_COMPLETE_ASYNC 0x01 /* This is async IO */
51 #define DIO_COMPLETE_INVALIDATE 0x02 /* Can invalidate pages */
54 * This code generally works in units of "dio_blocks". A dio_block is
55 * somewhere between the hard sector size and the filesystem block size. it
56 * is determined on a per-invocation basis. When talking to the filesystem
57 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
58 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
59 * to bio_block quantities by shifting left by blkfactor.
61 * If blkfactor is zero then the user's request was aligned to the filesystem's
65 /* dio_state only used in the submission path */
68 struct bio
*bio
; /* bio under assembly */
69 unsigned blkbits
; /* doesn't change */
70 unsigned blkfactor
; /* When we're using an alignment which
71 is finer than the filesystem's soft
72 blocksize, this specifies how much
73 finer. blkfactor=2 means 1/4-block
74 alignment. Does not change */
75 unsigned start_zero_done
; /* flag: sub-blocksize zeroing has
76 been performed at the start of a
78 int pages_in_io
; /* approximate total IO pages */
79 sector_t block_in_file
; /* Current offset into the underlying
80 file in dio_block units. */
81 unsigned blocks_available
; /* At block_in_file. changes */
82 int reap_counter
; /* rate limit reaping */
83 sector_t final_block_in_request
;/* doesn't change */
84 int boundary
; /* prev block is at a boundary */
85 get_block_t
*get_block
; /* block mapping function */
86 dio_submit_t
*submit_io
; /* IO submition function */
88 loff_t logical_offset_in_bio
; /* current first logical block in bio */
89 sector_t final_block_in_bio
; /* current final block in bio + 1 */
90 sector_t next_block_for_io
; /* next block to be put under IO,
91 in dio_blocks units */
94 * Deferred addition of a page to the dio. These variables are
95 * private to dio_send_cur_page(), submit_page_section() and
98 struct page
*cur_page
; /* The page */
99 unsigned cur_page_offset
; /* Offset into it, in bytes */
100 unsigned cur_page_len
; /* Nr of bytes at cur_page_offset */
101 sector_t cur_page_block
; /* Where it starts */
102 loff_t cur_page_fs_offset
; /* Offset in file */
104 struct iov_iter
*iter
;
106 * Page queue. These variables belong to dio_refill_pages() and
109 unsigned head
; /* next page to process */
110 unsigned tail
; /* last valid page + 1 */
114 /* dio_state communicated between submission path and end_io */
116 int flags
; /* doesn't change */
120 struct gendisk
*bio_disk
;
122 loff_t i_size
; /* i_size when submitted */
123 dio_iodone_t
*end_io
; /* IO completion function */
125 void *private; /* copy from map_bh.b_private */
127 /* BIO completion state */
128 spinlock_t bio_lock
; /* protects BIO fields below */
129 int page_errors
; /* errno from get_user_pages() */
130 int is_async
; /* is IO async ? */
131 bool defer_completion
; /* defer AIO completion to workqueue? */
132 bool should_dirty
; /* if pages should be dirtied */
133 int io_error
; /* IO error in completion path */
134 unsigned long refcount
; /* direct_io_worker() and bios */
135 struct bio
*bio_list
; /* singly linked via bi_private */
136 struct task_struct
*waiter
; /* waiting task (NULL if none) */
138 /* AIO related stuff */
139 struct kiocb
*iocb
; /* kiocb */
140 ssize_t result
; /* IO result */
143 * pages[] (and any fields placed after it) are not zeroed out at
144 * allocation time. Don't add new fields after pages[] unless you
145 * wish that they not be zeroed.
148 struct page
*pages
[DIO_PAGES
]; /* page buffer */
149 struct work_struct complete_work
;/* deferred AIO completion */
151 } ____cacheline_aligned_in_smp
;
153 static struct kmem_cache
*dio_cache __read_mostly
;
156 * How many pages are in the queue?
158 static inline unsigned dio_pages_present(struct dio_submit
*sdio
)
160 return sdio
->tail
- sdio
->head
;
164 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
166 static inline int dio_refill_pages(struct dio
*dio
, struct dio_submit
*sdio
)
170 ret
= iov_iter_get_pages(sdio
->iter
, dio
->pages
, LONG_MAX
, DIO_PAGES
,
173 if (ret
< 0 && sdio
->blocks_available
&& (dio
->op
== REQ_OP_WRITE
)) {
174 struct page
*page
= ZERO_PAGE(0);
176 * A memory fault, but the filesystem has some outstanding
177 * mapped blocks. We need to use those blocks up to avoid
178 * leaking stale data in the file.
180 if (dio
->page_errors
== 0)
181 dio
->page_errors
= ret
;
183 dio
->pages
[0] = page
;
187 sdio
->to
= PAGE_SIZE
;
192 iov_iter_advance(sdio
->iter
, ret
);
195 sdio
->tail
= (ret
+ PAGE_SIZE
- 1) / PAGE_SIZE
;
196 sdio
->to
= ((ret
- 1) & (PAGE_SIZE
- 1)) + 1;
203 * Get another userspace page. Returns an ERR_PTR on error. Pages are
204 * buffered inside the dio so that we can call get_user_pages() against a
205 * decent number of pages, less frequently. To provide nicer use of the
208 static inline struct page
*dio_get_page(struct dio
*dio
,
209 struct dio_submit
*sdio
)
211 if (dio_pages_present(sdio
) == 0) {
214 ret
= dio_refill_pages(dio
, sdio
);
217 BUG_ON(dio_pages_present(sdio
) == 0);
219 return dio
->pages
[sdio
->head
];
223 * Warn about a page cache invalidation failure during a direct io write.
225 void dio_warn_stale_pagecache(struct file
*filp
)
227 static DEFINE_RATELIMIT_STATE(_rs
, 86400 * HZ
, DEFAULT_RATELIMIT_BURST
);
229 struct inode
*inode
= file_inode(filp
);
232 errseq_set(&inode
->i_mapping
->wb_err
, -EIO
);
233 if (__ratelimit(&_rs
)) {
234 path
= file_path(filp
, pathname
, sizeof(pathname
));
237 pr_crit("Page cache invalidation failure on direct I/O. Possible data corruption due to collision with buffered I/O!\n");
238 pr_crit("File: %s PID: %d Comm: %.20s\n", path
, current
->pid
,
244 * dio_complete() - called when all DIO BIO I/O has been completed
245 * @offset: the byte offset in the file of the completed operation
247 * This drops i_dio_count, lets interested parties know that a DIO operation
248 * has completed, and calculates the resulting return code for the operation.
250 * It lets the filesystem know if it registered an interest earlier via
251 * get_block. Pass the private field of the map buffer_head so that
252 * filesystems can use it to hold additional state between get_block calls and
255 static ssize_t
dio_complete(struct dio
*dio
, ssize_t ret
, unsigned int flags
)
257 loff_t offset
= dio
->iocb
->ki_pos
;
258 ssize_t transferred
= 0;
262 * AIO submission can race with bio completion to get here while
263 * expecting to have the last io completed by bio completion.
264 * In that case -EIOCBQUEUED is in fact not an error we want
265 * to preserve through this call.
267 if (ret
== -EIOCBQUEUED
)
271 transferred
= dio
->result
;
273 /* Check for short read case */
274 if ((dio
->op
== REQ_OP_READ
) &&
275 ((offset
+ transferred
) > dio
->i_size
))
276 transferred
= dio
->i_size
- offset
;
277 /* ignore EFAULT if some IO has been done */
278 if (unlikely(ret
== -EFAULT
) && transferred
)
283 ret
= dio
->page_errors
;
291 err
= dio
->end_io(dio
->iocb
, offset
, ret
, dio
->private);
297 * Try again to invalidate clean pages which might have been cached by
298 * non-direct readahead, or faulted in by get_user_pages() if the source
299 * of the write was an mmap'ed region of the file we're writing. Either
300 * one is a pretty crazy thing to do, so we don't support it 100%. If
301 * this invalidation fails, tough, the write still worked...
303 * And this page cache invalidation has to be after dio->end_io(), as
304 * some filesystems convert unwritten extents to real allocations in
305 * end_io() when necessary, otherwise a racing buffer read would cache
306 * zeros from unwritten extents.
308 if (flags
& DIO_COMPLETE_INVALIDATE
&&
309 ret
> 0 && dio
->op
== REQ_OP_WRITE
&&
310 dio
->inode
->i_mapping
->nrpages
) {
311 err
= invalidate_inode_pages2_range(dio
->inode
->i_mapping
,
312 offset
>> PAGE_SHIFT
,
313 (offset
+ ret
- 1) >> PAGE_SHIFT
);
315 dio_warn_stale_pagecache(dio
->iocb
->ki_filp
);
318 inode_dio_end(dio
->inode
);
320 if (flags
& DIO_COMPLETE_ASYNC
) {
322 * generic_write_sync expects ki_pos to have been updated
323 * already, but the submission path only does this for
326 dio
->iocb
->ki_pos
+= transferred
;
328 if (dio
->op
== REQ_OP_WRITE
)
329 ret
= generic_write_sync(dio
->iocb
, transferred
);
330 dio
->iocb
->ki_complete(dio
->iocb
, ret
, 0);
333 kmem_cache_free(dio_cache
, dio
);
337 static void dio_aio_complete_work(struct work_struct
*work
)
339 struct dio
*dio
= container_of(work
, struct dio
, complete_work
);
341 dio_complete(dio
, 0, DIO_COMPLETE_ASYNC
| DIO_COMPLETE_INVALIDATE
);
344 static blk_status_t
dio_bio_complete(struct dio
*dio
, struct bio
*bio
);
347 * Asynchronous IO callback.
349 static void dio_bio_end_aio(struct bio
*bio
)
351 struct dio
*dio
= bio
->bi_private
;
352 unsigned long remaining
;
354 bool defer_completion
= false;
356 /* cleanup the bio */
357 dio_bio_complete(dio
, bio
);
359 spin_lock_irqsave(&dio
->bio_lock
, flags
);
360 remaining
= --dio
->refcount
;
361 if (remaining
== 1 && dio
->waiter
)
362 wake_up_process(dio
->waiter
);
363 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
365 if (remaining
== 0) {
367 * Defer completion when defer_completion is set or
368 * when the inode has pages mapped and this is AIO write.
369 * We need to invalidate those pages because there is a
370 * chance they contain stale data in the case buffered IO
371 * went in between AIO submission and completion into the
375 defer_completion
= dio
->defer_completion
||
376 (dio
->op
== REQ_OP_WRITE
&&
377 dio
->inode
->i_mapping
->nrpages
);
378 if (defer_completion
) {
379 INIT_WORK(&dio
->complete_work
, dio_aio_complete_work
);
380 queue_work(dio
->inode
->i_sb
->s_dio_done_wq
,
381 &dio
->complete_work
);
383 dio_complete(dio
, 0, DIO_COMPLETE_ASYNC
);
389 * The BIO completion handler simply queues the BIO up for the process-context
392 * During I/O bi_private points at the dio. After I/O, bi_private is used to
393 * implement a singly-linked list of completed BIOs, at dio->bio_list.
395 static void dio_bio_end_io(struct bio
*bio
)
397 struct dio
*dio
= bio
->bi_private
;
400 spin_lock_irqsave(&dio
->bio_lock
, flags
);
401 bio
->bi_private
= dio
->bio_list
;
403 if (--dio
->refcount
== 1 && dio
->waiter
)
404 wake_up_process(dio
->waiter
);
405 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
409 * dio_end_io - handle the end io action for the given bio
410 * @bio: The direct io bio thats being completed
412 * This is meant to be called by any filesystem that uses their own dio_submit_t
413 * so that the DIO specific endio actions are dealt with after the filesystem
414 * has done it's completion work.
416 void dio_end_io(struct bio
*bio
)
418 struct dio
*dio
= bio
->bi_private
;
421 dio_bio_end_aio(bio
);
425 EXPORT_SYMBOL_GPL(dio_end_io
);
428 dio_bio_alloc(struct dio
*dio
, struct dio_submit
*sdio
,
429 struct block_device
*bdev
,
430 sector_t first_sector
, int nr_vecs
)
435 * bio_alloc() is guaranteed to return a bio when allowed to sleep and
436 * we request a valid number of vectors.
438 bio
= bio_alloc(GFP_KERNEL
, nr_vecs
);
440 bio_set_dev(bio
, bdev
);
441 bio
->bi_iter
.bi_sector
= first_sector
;
442 bio_set_op_attrs(bio
, dio
->op
, dio
->op_flags
);
444 bio
->bi_end_io
= dio_bio_end_aio
;
446 bio
->bi_end_io
= dio_bio_end_io
;
448 bio
->bi_write_hint
= dio
->iocb
->ki_hint
;
451 sdio
->logical_offset_in_bio
= sdio
->cur_page_fs_offset
;
455 * In the AIO read case we speculatively dirty the pages before starting IO.
456 * During IO completion, any of these pages which happen to have been written
457 * back will be redirtied by bio_check_pages_dirty().
459 * bios hold a dio reference between submit_bio and ->end_io.
461 static inline void dio_bio_submit(struct dio
*dio
, struct dio_submit
*sdio
)
463 struct bio
*bio
= sdio
->bio
;
466 bio
->bi_private
= dio
;
468 spin_lock_irqsave(&dio
->bio_lock
, flags
);
470 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
472 if (dio
->is_async
&& dio
->op
== REQ_OP_READ
&& dio
->should_dirty
)
473 bio_set_pages_dirty(bio
);
475 dio
->bio_disk
= bio
->bi_disk
;
477 if (sdio
->submit_io
) {
478 sdio
->submit_io(bio
, dio
->inode
, sdio
->logical_offset_in_bio
);
479 dio
->bio_cookie
= BLK_QC_T_NONE
;
481 dio
->bio_cookie
= submit_bio(bio
);
485 sdio
->logical_offset_in_bio
= 0;
489 * Release any resources in case of a failure
491 static inline void dio_cleanup(struct dio
*dio
, struct dio_submit
*sdio
)
493 while (sdio
->head
< sdio
->tail
)
494 put_page(dio
->pages
[sdio
->head
++]);
498 * Wait for the next BIO to complete. Remove it and return it. NULL is
499 * returned once all BIOs have been completed. This must only be called once
500 * all bios have been issued so that dio->refcount can only decrease. This
501 * requires that that the caller hold a reference on the dio.
503 static struct bio
*dio_await_one(struct dio
*dio
)
506 struct bio
*bio
= NULL
;
508 spin_lock_irqsave(&dio
->bio_lock
, flags
);
511 * Wait as long as the list is empty and there are bios in flight. bio
512 * completion drops the count, maybe adds to the list, and wakes while
513 * holding the bio_lock so we don't need set_current_state()'s barrier
514 * and can call it after testing our condition.
516 while (dio
->refcount
> 1 && dio
->bio_list
== NULL
) {
517 __set_current_state(TASK_UNINTERRUPTIBLE
);
518 dio
->waiter
= current
;
519 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
520 if (!(dio
->iocb
->ki_flags
& IOCB_HIPRI
) ||
521 !blk_poll(dio
->bio_disk
->queue
, dio
->bio_cookie
))
523 /* wake up sets us TASK_RUNNING */
524 spin_lock_irqsave(&dio
->bio_lock
, flags
);
529 dio
->bio_list
= bio
->bi_private
;
531 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
536 * Process one completed BIO. No locks are held.
538 static blk_status_t
dio_bio_complete(struct dio
*dio
, struct bio
*bio
)
540 struct bio_vec
*bvec
;
542 blk_status_t err
= bio
->bi_status
;
545 if (err
== BLK_STS_AGAIN
&& (bio
->bi_opf
& REQ_NOWAIT
))
546 dio
->io_error
= -EAGAIN
;
548 dio
->io_error
= -EIO
;
551 if (dio
->is_async
&& dio
->op
== REQ_OP_READ
&& dio
->should_dirty
) {
552 bio_check_pages_dirty(bio
); /* transfers ownership */
554 bio_for_each_segment_all(bvec
, bio
, i
) {
555 struct page
*page
= bvec
->bv_page
;
557 if (dio
->op
== REQ_OP_READ
&& !PageCompound(page
) &&
559 set_page_dirty_lock(page
);
568 * Wait on and process all in-flight BIOs. This must only be called once
569 * all bios have been issued so that the refcount can only decrease.
570 * This just waits for all bios to make it through dio_bio_complete. IO
571 * errors are propagated through dio->io_error and should be propagated via
574 static void dio_await_completion(struct dio
*dio
)
578 bio
= dio_await_one(dio
);
580 dio_bio_complete(dio
, bio
);
585 * A really large O_DIRECT read or write can generate a lot of BIOs. So
586 * to keep the memory consumption sane we periodically reap any completed BIOs
587 * during the BIO generation phase.
589 * This also helps to limit the peak amount of pinned userspace memory.
591 static inline int dio_bio_reap(struct dio
*dio
, struct dio_submit
*sdio
)
595 if (sdio
->reap_counter
++ >= 64) {
596 while (dio
->bio_list
) {
601 spin_lock_irqsave(&dio
->bio_lock
, flags
);
603 dio
->bio_list
= bio
->bi_private
;
604 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
605 ret2
= blk_status_to_errno(dio_bio_complete(dio
, bio
));
609 sdio
->reap_counter
= 0;
615 * Create workqueue for deferred direct IO completions. We allocate the
616 * workqueue when it's first needed. This avoids creating workqueue for
617 * filesystems that don't need it and also allows us to create the workqueue
618 * late enough so the we can include s_id in the name of the workqueue.
620 int sb_init_dio_done_wq(struct super_block
*sb
)
622 struct workqueue_struct
*old
;
623 struct workqueue_struct
*wq
= alloc_workqueue("dio/%s",
629 * This has to be atomic as more DIOs can race to create the workqueue
631 old
= cmpxchg(&sb
->s_dio_done_wq
, NULL
, wq
);
632 /* Someone created workqueue before us? Free ours... */
634 destroy_workqueue(wq
);
638 static int dio_set_defer_completion(struct dio
*dio
)
640 struct super_block
*sb
= dio
->inode
->i_sb
;
642 if (dio
->defer_completion
)
644 dio
->defer_completion
= true;
645 if (!sb
->s_dio_done_wq
)
646 return sb_init_dio_done_wq(sb
);
651 * Call into the fs to map some more disk blocks. We record the current number
652 * of available blocks at sdio->blocks_available. These are in units of the
653 * fs blocksize, i_blocksize(inode).
655 * The fs is allowed to map lots of blocks at once. If it wants to do that,
656 * it uses the passed inode-relative block number as the file offset, as usual.
658 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
659 * has remaining to do. The fs should not map more than this number of blocks.
661 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
662 * indicate how much contiguous disk space has been made available at
665 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
666 * This isn't very efficient...
668 * In the case of filesystem holes: the fs may return an arbitrarily-large
669 * hole by returning an appropriate value in b_size and by clearing
670 * buffer_mapped(). However the direct-io code will only process holes one
671 * block at a time - it will repeatedly call get_block() as it walks the hole.
673 static int get_more_blocks(struct dio
*dio
, struct dio_submit
*sdio
,
674 struct buffer_head
*map_bh
)
677 sector_t fs_startblk
; /* Into file, in filesystem-sized blocks */
678 sector_t fs_endblk
; /* Into file, in filesystem-sized blocks */
679 unsigned long fs_count
; /* Number of filesystem-sized blocks */
681 unsigned int i_blkbits
= sdio
->blkbits
+ sdio
->blkfactor
;
684 * If there was a memory error and we've overwritten all the
685 * mapped blocks then we can now return that memory error
687 ret
= dio
->page_errors
;
689 BUG_ON(sdio
->block_in_file
>= sdio
->final_block_in_request
);
690 fs_startblk
= sdio
->block_in_file
>> sdio
->blkfactor
;
691 fs_endblk
= (sdio
->final_block_in_request
- 1) >>
693 fs_count
= fs_endblk
- fs_startblk
+ 1;
696 map_bh
->b_size
= fs_count
<< i_blkbits
;
699 * For writes that could fill holes inside i_size on a
700 * DIO_SKIP_HOLES filesystem we forbid block creations: only
701 * overwrites are permitted. We will return early to the caller
702 * once we see an unmapped buffer head returned, and the caller
703 * will fall back to buffered I/O.
705 * Otherwise the decision is left to the get_blocks method,
706 * which may decide to handle it or also return an unmapped
709 create
= dio
->op
== REQ_OP_WRITE
;
710 if (dio
->flags
& DIO_SKIP_HOLES
) {
711 if (fs_startblk
<= ((i_size_read(dio
->inode
) - 1) >>
716 ret
= (*sdio
->get_block
)(dio
->inode
, fs_startblk
,
719 /* Store for completion */
720 dio
->private = map_bh
->b_private
;
722 if (ret
== 0 && buffer_defer_completion(map_bh
))
723 ret
= dio_set_defer_completion(dio
);
729 * There is no bio. Make one now.
731 static inline int dio_new_bio(struct dio
*dio
, struct dio_submit
*sdio
,
732 sector_t start_sector
, struct buffer_head
*map_bh
)
737 ret
= dio_bio_reap(dio
, sdio
);
740 sector
= start_sector
<< (sdio
->blkbits
- 9);
741 nr_pages
= min(sdio
->pages_in_io
, BIO_MAX_PAGES
);
742 BUG_ON(nr_pages
<= 0);
743 dio_bio_alloc(dio
, sdio
, map_bh
->b_bdev
, sector
, nr_pages
);
750 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
751 * that was successful then update final_block_in_bio and take a ref against
752 * the just-added page.
754 * Return zero on success. Non-zero means the caller needs to start a new BIO.
756 static inline int dio_bio_add_page(struct dio_submit
*sdio
)
760 ret
= bio_add_page(sdio
->bio
, sdio
->cur_page
,
761 sdio
->cur_page_len
, sdio
->cur_page_offset
);
762 if (ret
== sdio
->cur_page_len
) {
764 * Decrement count only, if we are done with this page
766 if ((sdio
->cur_page_len
+ sdio
->cur_page_offset
) == PAGE_SIZE
)
768 get_page(sdio
->cur_page
);
769 sdio
->final_block_in_bio
= sdio
->cur_page_block
+
770 (sdio
->cur_page_len
>> sdio
->blkbits
);
779 * Put cur_page under IO. The section of cur_page which is described by
780 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
781 * starts on-disk at cur_page_block.
783 * We take a ref against the page here (on behalf of its presence in the bio).
785 * The caller of this function is responsible for removing cur_page from the
786 * dio, and for dropping the refcount which came from that presence.
788 static inline int dio_send_cur_page(struct dio
*dio
, struct dio_submit
*sdio
,
789 struct buffer_head
*map_bh
)
794 loff_t cur_offset
= sdio
->cur_page_fs_offset
;
795 loff_t bio_next_offset
= sdio
->logical_offset_in_bio
+
796 sdio
->bio
->bi_iter
.bi_size
;
799 * See whether this new request is contiguous with the old.
801 * Btrfs cannot handle having logically non-contiguous requests
802 * submitted. For example if you have
804 * Logical: [0-4095][HOLE][8192-12287]
805 * Physical: [0-4095] [4096-8191]
807 * We cannot submit those pages together as one BIO. So if our
808 * current logical offset in the file does not equal what would
809 * be the next logical offset in the bio, submit the bio we
812 if (sdio
->final_block_in_bio
!= sdio
->cur_page_block
||
813 cur_offset
!= bio_next_offset
)
814 dio_bio_submit(dio
, sdio
);
817 if (sdio
->bio
== NULL
) {
818 ret
= dio_new_bio(dio
, sdio
, sdio
->cur_page_block
, map_bh
);
823 if (dio_bio_add_page(sdio
) != 0) {
824 dio_bio_submit(dio
, sdio
);
825 ret
= dio_new_bio(dio
, sdio
, sdio
->cur_page_block
, map_bh
);
827 ret
= dio_bio_add_page(sdio
);
836 * An autonomous function to put a chunk of a page under deferred IO.
838 * The caller doesn't actually know (or care) whether this piece of page is in
839 * a BIO, or is under IO or whatever. We just take care of all possible
840 * situations here. The separation between the logic of do_direct_IO() and
841 * that of submit_page_section() is important for clarity. Please don't break.
843 * The chunk of page starts on-disk at blocknr.
845 * We perform deferred IO, by recording the last-submitted page inside our
846 * private part of the dio structure. If possible, we just expand the IO
847 * across that page here.
849 * If that doesn't work out then we put the old page into the bio and add this
850 * page to the dio instead.
853 submit_page_section(struct dio
*dio
, struct dio_submit
*sdio
, struct page
*page
,
854 unsigned offset
, unsigned len
, sector_t blocknr
,
855 struct buffer_head
*map_bh
)
859 if (dio
->op
== REQ_OP_WRITE
) {
861 * Read accounting is performed in submit_bio()
863 task_io_account_write(len
);
867 * Can we just grow the current page's presence in the dio?
869 if (sdio
->cur_page
== page
&&
870 sdio
->cur_page_offset
+ sdio
->cur_page_len
== offset
&&
871 sdio
->cur_page_block
+
872 (sdio
->cur_page_len
>> sdio
->blkbits
) == blocknr
) {
873 sdio
->cur_page_len
+= len
;
878 * If there's a deferred page already there then send it.
880 if (sdio
->cur_page
) {
881 ret
= dio_send_cur_page(dio
, sdio
, map_bh
);
882 put_page(sdio
->cur_page
);
883 sdio
->cur_page
= NULL
;
888 get_page(page
); /* It is in dio */
889 sdio
->cur_page
= page
;
890 sdio
->cur_page_offset
= offset
;
891 sdio
->cur_page_len
= len
;
892 sdio
->cur_page_block
= blocknr
;
893 sdio
->cur_page_fs_offset
= sdio
->block_in_file
<< sdio
->blkbits
;
896 * If sdio->boundary then we want to schedule the IO now to
897 * avoid metadata seeks.
899 if (sdio
->boundary
) {
900 ret
= dio_send_cur_page(dio
, sdio
, map_bh
);
902 dio_bio_submit(dio
, sdio
);
903 put_page(sdio
->cur_page
);
904 sdio
->cur_page
= NULL
;
910 * If we are not writing the entire block and get_block() allocated
911 * the block for us, we need to fill-in the unused portion of the
912 * block with zeros. This happens only if user-buffer, fileoffset or
913 * io length is not filesystem block-size multiple.
915 * `end' is zero if we're doing the start of the IO, 1 at the end of the
918 static inline void dio_zero_block(struct dio
*dio
, struct dio_submit
*sdio
,
919 int end
, struct buffer_head
*map_bh
)
921 unsigned dio_blocks_per_fs_block
;
922 unsigned this_chunk_blocks
; /* In dio_blocks */
923 unsigned this_chunk_bytes
;
926 sdio
->start_zero_done
= 1;
927 if (!sdio
->blkfactor
|| !buffer_new(map_bh
))
930 dio_blocks_per_fs_block
= 1 << sdio
->blkfactor
;
931 this_chunk_blocks
= sdio
->block_in_file
& (dio_blocks_per_fs_block
- 1);
933 if (!this_chunk_blocks
)
937 * We need to zero out part of an fs block. It is either at the
938 * beginning or the end of the fs block.
941 this_chunk_blocks
= dio_blocks_per_fs_block
- this_chunk_blocks
;
943 this_chunk_bytes
= this_chunk_blocks
<< sdio
->blkbits
;
946 if (submit_page_section(dio
, sdio
, page
, 0, this_chunk_bytes
,
947 sdio
->next_block_for_io
, map_bh
))
950 sdio
->next_block_for_io
+= this_chunk_blocks
;
954 * Walk the user pages, and the file, mapping blocks to disk and generating
955 * a sequence of (page,offset,len,block) mappings. These mappings are injected
956 * into submit_page_section(), which takes care of the next stage of submission
958 * Direct IO against a blockdev is different from a file. Because we can
959 * happily perform page-sized but 512-byte aligned IOs. It is important that
960 * blockdev IO be able to have fine alignment and large sizes.
962 * So what we do is to permit the ->get_block function to populate bh.b_size
963 * with the size of IO which is permitted at this offset and this i_blkbits.
965 * For best results, the blockdev should be set up with 512-byte i_blkbits and
966 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
967 * fine alignment but still allows this function to work in PAGE_SIZE units.
969 static int do_direct_IO(struct dio
*dio
, struct dio_submit
*sdio
,
970 struct buffer_head
*map_bh
)
972 const unsigned blkbits
= sdio
->blkbits
;
973 const unsigned i_blkbits
= blkbits
+ sdio
->blkfactor
;
976 while (sdio
->block_in_file
< sdio
->final_block_in_request
) {
980 page
= dio_get_page(dio
, sdio
);
985 from
= sdio
->head
? 0 : sdio
->from
;
986 to
= (sdio
->head
== sdio
->tail
- 1) ? sdio
->to
: PAGE_SIZE
;
990 unsigned this_chunk_bytes
; /* # of bytes mapped */
991 unsigned this_chunk_blocks
; /* # of blocks */
994 if (sdio
->blocks_available
== 0) {
996 * Need to go and map some more disk
998 unsigned long blkmask
;
999 unsigned long dio_remainder
;
1001 ret
= get_more_blocks(dio
, sdio
, map_bh
);
1006 if (!buffer_mapped(map_bh
))
1009 sdio
->blocks_available
=
1010 map_bh
->b_size
>> blkbits
;
1011 sdio
->next_block_for_io
=
1012 map_bh
->b_blocknr
<< sdio
->blkfactor
;
1013 if (buffer_new(map_bh
)) {
1017 map_bh
->b_size
>> i_blkbits
);
1020 if (!sdio
->blkfactor
)
1023 blkmask
= (1 << sdio
->blkfactor
) - 1;
1024 dio_remainder
= (sdio
->block_in_file
& blkmask
);
1027 * If we are at the start of IO and that IO
1028 * starts partway into a fs-block,
1029 * dio_remainder will be non-zero. If the IO
1030 * is a read then we can simply advance the IO
1031 * cursor to the first block which is to be
1032 * read. But if the IO is a write and the
1033 * block was newly allocated we cannot do that;
1034 * the start of the fs block must be zeroed out
1037 if (!buffer_new(map_bh
))
1038 sdio
->next_block_for_io
+= dio_remainder
;
1039 sdio
->blocks_available
-= dio_remainder
;
1043 if (!buffer_mapped(map_bh
)) {
1044 loff_t i_size_aligned
;
1046 /* AKPM: eargh, -ENOTBLK is a hack */
1047 if (dio
->op
== REQ_OP_WRITE
) {
1053 * Be sure to account for a partial block as the
1054 * last block in the file
1056 i_size_aligned
= ALIGN(i_size_read(dio
->inode
),
1058 if (sdio
->block_in_file
>=
1059 i_size_aligned
>> blkbits
) {
1064 zero_user(page
, from
, 1 << blkbits
);
1065 sdio
->block_in_file
++;
1066 from
+= 1 << blkbits
;
1067 dio
->result
+= 1 << blkbits
;
1072 * If we're performing IO which has an alignment which
1073 * is finer than the underlying fs, go check to see if
1074 * we must zero out the start of this block.
1076 if (unlikely(sdio
->blkfactor
&& !sdio
->start_zero_done
))
1077 dio_zero_block(dio
, sdio
, 0, map_bh
);
1080 * Work out, in this_chunk_blocks, how much disk we
1081 * can add to this page
1083 this_chunk_blocks
= sdio
->blocks_available
;
1084 u
= (to
- from
) >> blkbits
;
1085 if (this_chunk_blocks
> u
)
1086 this_chunk_blocks
= u
;
1087 u
= sdio
->final_block_in_request
- sdio
->block_in_file
;
1088 if (this_chunk_blocks
> u
)
1089 this_chunk_blocks
= u
;
1090 this_chunk_bytes
= this_chunk_blocks
<< blkbits
;
1091 BUG_ON(this_chunk_bytes
== 0);
1093 if (this_chunk_blocks
== sdio
->blocks_available
)
1094 sdio
->boundary
= buffer_boundary(map_bh
);
1095 ret
= submit_page_section(dio
, sdio
, page
,
1098 sdio
->next_block_for_io
,
1104 sdio
->next_block_for_io
+= this_chunk_blocks
;
1106 sdio
->block_in_file
+= this_chunk_blocks
;
1107 from
+= this_chunk_bytes
;
1108 dio
->result
+= this_chunk_bytes
;
1109 sdio
->blocks_available
-= this_chunk_blocks
;
1111 BUG_ON(sdio
->block_in_file
> sdio
->final_block_in_request
);
1112 if (sdio
->block_in_file
== sdio
->final_block_in_request
)
1116 /* Drop the ref which was taken in get_user_pages() */
1123 static inline int drop_refcount(struct dio
*dio
)
1126 unsigned long flags
;
1129 * Sync will always be dropping the final ref and completing the
1130 * operation. AIO can if it was a broken operation described above or
1131 * in fact if all the bios race to complete before we get here. In
1132 * that case dio_complete() translates the EIOCBQUEUED into the proper
1133 * return code that the caller will hand to ->complete().
1135 * This is managed by the bio_lock instead of being an atomic_t so that
1136 * completion paths can drop their ref and use the remaining count to
1137 * decide to wake the submission path atomically.
1139 spin_lock_irqsave(&dio
->bio_lock
, flags
);
1140 ret2
= --dio
->refcount
;
1141 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
1146 * This is a library function for use by filesystem drivers.
1148 * The locking rules are governed by the flags parameter:
1149 * - if the flags value contains DIO_LOCKING we use a fancy locking
1150 * scheme for dumb filesystems.
1151 * For writes this function is called under i_mutex and returns with
1152 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1153 * taken and dropped again before returning.
1154 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1155 * internal locking but rather rely on the filesystem to synchronize
1156 * direct I/O reads/writes versus each other and truncate.
1158 * To help with locking against truncate we incremented the i_dio_count
1159 * counter before starting direct I/O, and decrement it once we are done.
1160 * Truncate can wait for it to reach zero to provide exclusion. It is
1161 * expected that filesystem provide exclusion between new direct I/O
1162 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1163 * but other filesystems need to take care of this on their own.
1165 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1166 * is always inlined. Otherwise gcc is unable to split the structure into
1167 * individual fields and will generate much worse code. This is important
1168 * for the whole file.
1170 static inline ssize_t
1171 do_blockdev_direct_IO(struct kiocb
*iocb
, struct inode
*inode
,
1172 struct block_device
*bdev
, struct iov_iter
*iter
,
1173 get_block_t get_block
, dio_iodone_t end_io
,
1174 dio_submit_t submit_io
, int flags
)
1176 unsigned i_blkbits
= READ_ONCE(inode
->i_blkbits
);
1177 unsigned blkbits
= i_blkbits
;
1178 unsigned blocksize_mask
= (1 << blkbits
) - 1;
1179 ssize_t retval
= -EINVAL
;
1180 const size_t count
= iov_iter_count(iter
);
1181 loff_t offset
= iocb
->ki_pos
;
1182 const loff_t end
= offset
+ count
;
1184 struct dio_submit sdio
= { 0, };
1185 struct buffer_head map_bh
= { 0, };
1186 struct blk_plug plug
;
1187 unsigned long align
= offset
| iov_iter_alignment(iter
);
1190 * Avoid references to bdev if not absolutely needed to give
1191 * the early prefetch in the caller enough time.
1194 if (align
& blocksize_mask
) {
1196 blkbits
= blksize_bits(bdev_logical_block_size(bdev
));
1197 blocksize_mask
= (1 << blkbits
) - 1;
1198 if (align
& blocksize_mask
)
1202 /* watch out for a 0 len io from a tricksy fs */
1203 if (iov_iter_rw(iter
) == READ
&& !count
)
1206 dio
= kmem_cache_alloc(dio_cache
, GFP_KERNEL
);
1211 * Believe it or not, zeroing out the page array caused a .5%
1212 * performance regression in a database benchmark. So, we take
1213 * care to only zero out what's needed.
1215 memset(dio
, 0, offsetof(struct dio
, pages
));
1218 if (dio
->flags
& DIO_LOCKING
) {
1219 if (iov_iter_rw(iter
) == READ
) {
1220 struct address_space
*mapping
=
1221 iocb
->ki_filp
->f_mapping
;
1223 /* will be released by direct_io_worker */
1226 retval
= filemap_write_and_wait_range(mapping
, offset
,
1229 inode_unlock(inode
);
1230 kmem_cache_free(dio_cache
, dio
);
1236 /* Once we sampled i_size check for reads beyond EOF */
1237 dio
->i_size
= i_size_read(inode
);
1238 if (iov_iter_rw(iter
) == READ
&& offset
>= dio
->i_size
) {
1239 if (dio
->flags
& DIO_LOCKING
)
1240 inode_unlock(inode
);
1241 kmem_cache_free(dio_cache
, dio
);
1247 * For file extending writes updating i_size before data writeouts
1248 * complete can expose uninitialized blocks in dumb filesystems.
1249 * In that case we need to wait for I/O completion even if asked
1250 * for an asynchronous write.
1252 if (is_sync_kiocb(iocb
))
1253 dio
->is_async
= false;
1254 else if (iov_iter_rw(iter
) == WRITE
&& end
> i_size_read(inode
))
1255 dio
->is_async
= false;
1257 dio
->is_async
= true;
1260 if (iov_iter_rw(iter
) == WRITE
) {
1261 dio
->op
= REQ_OP_WRITE
;
1262 dio
->op_flags
= REQ_SYNC
| REQ_IDLE
;
1263 if (iocb
->ki_flags
& IOCB_NOWAIT
)
1264 dio
->op_flags
|= REQ_NOWAIT
;
1266 dio
->op
= REQ_OP_READ
;
1270 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1271 * so that we can call ->fsync.
1273 if (dio
->is_async
&& iov_iter_rw(iter
) == WRITE
) {
1275 if (iocb
->ki_flags
& IOCB_DSYNC
)
1276 retval
= dio_set_defer_completion(dio
);
1277 else if (!dio
->inode
->i_sb
->s_dio_done_wq
) {
1279 * In case of AIO write racing with buffered read we
1280 * need to defer completion. We can't decide this now,
1281 * however the workqueue needs to be initialized here.
1283 retval
= sb_init_dio_done_wq(dio
->inode
->i_sb
);
1287 * We grab i_mutex only for reads so we don't have
1288 * to release it here
1290 kmem_cache_free(dio_cache
, dio
);
1296 * Will be decremented at I/O completion time.
1298 inode_dio_begin(inode
);
1301 sdio
.blkbits
= blkbits
;
1302 sdio
.blkfactor
= i_blkbits
- blkbits
;
1303 sdio
.block_in_file
= offset
>> blkbits
;
1305 sdio
.get_block
= get_block
;
1306 dio
->end_io
= end_io
;
1307 sdio
.submit_io
= submit_io
;
1308 sdio
.final_block_in_bio
= -1;
1309 sdio
.next_block_for_io
= -1;
1313 spin_lock_init(&dio
->bio_lock
);
1316 dio
->should_dirty
= iter_is_iovec(iter
) && iov_iter_rw(iter
) == READ
;
1318 sdio
.final_block_in_request
= end
>> blkbits
;
1321 * In case of non-aligned buffers, we may need 2 more
1322 * pages since we need to zero out first and last block.
1324 if (unlikely(sdio
.blkfactor
))
1325 sdio
.pages_in_io
= 2;
1327 sdio
.pages_in_io
+= iov_iter_npages(iter
, INT_MAX
);
1329 blk_start_plug(&plug
);
1331 retval
= do_direct_IO(dio
, &sdio
, &map_bh
);
1333 dio_cleanup(dio
, &sdio
);
1335 if (retval
== -ENOTBLK
) {
1337 * The remaining part of the request will be
1338 * be handled by buffered I/O when we return
1343 * There may be some unwritten disk at the end of a part-written
1344 * fs-block-sized block. Go zero that now.
1346 dio_zero_block(dio
, &sdio
, 1, &map_bh
);
1348 if (sdio
.cur_page
) {
1351 ret2
= dio_send_cur_page(dio
, &sdio
, &map_bh
);
1354 put_page(sdio
.cur_page
);
1355 sdio
.cur_page
= NULL
;
1358 dio_bio_submit(dio
, &sdio
);
1360 blk_finish_plug(&plug
);
1363 * It is possible that, we return short IO due to end of file.
1364 * In that case, we need to release all the pages we got hold on.
1366 dio_cleanup(dio
, &sdio
);
1369 * All block lookups have been performed. For READ requests
1370 * we can let i_mutex go now that its achieved its purpose
1371 * of protecting us from looking up uninitialized blocks.
1373 if (iov_iter_rw(iter
) == READ
&& (dio
->flags
& DIO_LOCKING
))
1374 inode_unlock(dio
->inode
);
1377 * The only time we want to leave bios in flight is when a successful
1378 * partial aio read or full aio write have been setup. In that case
1379 * bio completion will call aio_complete. The only time it's safe to
1380 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1381 * This had *better* be the only place that raises -EIOCBQUEUED.
1383 BUG_ON(retval
== -EIOCBQUEUED
);
1384 if (dio
->is_async
&& retval
== 0 && dio
->result
&&
1385 (iov_iter_rw(iter
) == READ
|| dio
->result
== count
))
1386 retval
= -EIOCBQUEUED
;
1388 dio_await_completion(dio
);
1390 if (drop_refcount(dio
) == 0) {
1391 retval
= dio_complete(dio
, retval
, DIO_COMPLETE_INVALIDATE
);
1393 BUG_ON(retval
!= -EIOCBQUEUED
);
1399 ssize_t
__blockdev_direct_IO(struct kiocb
*iocb
, struct inode
*inode
,
1400 struct block_device
*bdev
, struct iov_iter
*iter
,
1401 get_block_t get_block
,
1402 dio_iodone_t end_io
, dio_submit_t submit_io
,
1406 * The block device state is needed in the end to finally
1407 * submit everything. Since it's likely to be cache cold
1408 * prefetch it here as first thing to hide some of the
1411 * Attempt to prefetch the pieces we likely need later.
1413 prefetch(&bdev
->bd_disk
->part_tbl
);
1414 prefetch(bdev
->bd_queue
);
1415 prefetch((char *)bdev
->bd_queue
+ SMP_CACHE_BYTES
);
1417 return do_blockdev_direct_IO(iocb
, inode
, bdev
, iter
, get_block
,
1418 end_io
, submit_io
, flags
);
1421 EXPORT_SYMBOL(__blockdev_direct_IO
);
1423 static __init
int dio_init(void)
1425 dio_cache
= KMEM_CACHE(dio
, SLAB_PANIC
);
1428 module_init(dio_init
)