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/fscrypt.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 * Flags for dio_complete()
51 #define DIO_COMPLETE_ASYNC 0x01 /* This is async IO */
52 #define DIO_COMPLETE_INVALIDATE 0x02 /* Can invalidate pages */
55 * This code generally works in units of "dio_blocks". A dio_block is
56 * somewhere between the hard sector size and the filesystem block size. it
57 * is determined on a per-invocation basis. When talking to the filesystem
58 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
59 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
60 * to bio_block quantities by shifting left by blkfactor.
62 * If blkfactor is zero then the user's request was aligned to the filesystem's
66 /* dio_state only used in the submission path */
69 struct bio
*bio
; /* bio under assembly */
70 unsigned blkbits
; /* doesn't change */
71 unsigned blkfactor
; /* When we're using an alignment which
72 is finer than the filesystem's soft
73 blocksize, this specifies how much
74 finer. blkfactor=2 means 1/4-block
75 alignment. Does not change */
76 unsigned start_zero_done
; /* flag: sub-blocksize zeroing has
77 been performed at the start of a
79 int pages_in_io
; /* approximate total IO pages */
80 sector_t block_in_file
; /* Current offset into the underlying
81 file in dio_block units. */
82 unsigned blocks_available
; /* At block_in_file. changes */
83 int reap_counter
; /* rate limit reaping */
84 sector_t final_block_in_request
;/* doesn't change */
85 int boundary
; /* prev block is at a boundary */
86 get_block_t
*get_block
; /* block mapping function */
87 dio_submit_t
*submit_io
; /* IO submition function */
89 loff_t logical_offset_in_bio
; /* current first logical block in bio */
90 sector_t final_block_in_bio
; /* current final block in bio + 1 */
91 sector_t next_block_for_io
; /* next block to be put under IO,
92 in dio_blocks units */
95 * Deferred addition of a page to the dio. These variables are
96 * private to dio_send_cur_page(), submit_page_section() and
99 struct page
*cur_page
; /* The page */
100 unsigned cur_page_offset
; /* Offset into it, in bytes */
101 unsigned cur_page_len
; /* Nr of bytes at cur_page_offset */
102 sector_t cur_page_block
; /* Where it starts */
103 loff_t cur_page_fs_offset
; /* Offset in file */
105 struct iov_iter
*iter
;
107 * Page queue. These variables belong to dio_refill_pages() and
110 unsigned head
; /* next page to process */
111 unsigned tail
; /* last valid page + 1 */
115 /* dio_state communicated between submission path and end_io */
117 int flags
; /* doesn't change */
121 struct gendisk
*bio_disk
;
123 loff_t i_size
; /* i_size when submitted */
124 dio_iodone_t
*end_io
; /* IO completion function */
126 void *private; /* copy from map_bh.b_private */
128 /* BIO completion state */
129 spinlock_t bio_lock
; /* protects BIO fields below */
130 int page_errors
; /* errno from get_user_pages() */
131 int is_async
; /* is IO async ? */
132 bool defer_completion
; /* defer AIO completion to workqueue? */
133 bool should_dirty
; /* if pages should be dirtied */
134 int io_error
; /* IO error in completion path */
135 unsigned long refcount
; /* direct_io_worker() and bios */
136 struct bio
*bio_list
; /* singly linked via bi_private */
137 struct task_struct
*waiter
; /* waiting task (NULL if none) */
139 /* AIO related stuff */
140 struct kiocb
*iocb
; /* kiocb */
141 ssize_t result
; /* IO result */
144 * pages[] (and any fields placed after it) are not zeroed out at
145 * allocation time. Don't add new fields after pages[] unless you
146 * wish that they not be zeroed.
149 struct page
*pages
[DIO_PAGES
]; /* page buffer */
150 struct work_struct complete_work
;/* deferred AIO completion */
152 } ____cacheline_aligned_in_smp
;
154 static struct kmem_cache
*dio_cache __read_mostly
;
157 * How many pages are in the queue?
159 static inline unsigned dio_pages_present(struct dio_submit
*sdio
)
161 return sdio
->tail
- sdio
->head
;
165 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
167 static inline int dio_refill_pages(struct dio
*dio
, struct dio_submit
*sdio
)
171 ret
= iov_iter_get_pages(sdio
->iter
, dio
->pages
, LONG_MAX
, DIO_PAGES
,
174 if (ret
< 0 && sdio
->blocks_available
&& (dio
->op
== REQ_OP_WRITE
)) {
175 struct page
*page
= ZERO_PAGE(0);
177 * A memory fault, but the filesystem has some outstanding
178 * mapped blocks. We need to use those blocks up to avoid
179 * leaking stale data in the file.
181 if (dio
->page_errors
== 0)
182 dio
->page_errors
= ret
;
184 dio
->pages
[0] = page
;
188 sdio
->to
= PAGE_SIZE
;
193 iov_iter_advance(sdio
->iter
, ret
);
196 sdio
->tail
= (ret
+ PAGE_SIZE
- 1) / PAGE_SIZE
;
197 sdio
->to
= ((ret
- 1) & (PAGE_SIZE
- 1)) + 1;
204 * Get another userspace page. Returns an ERR_PTR on error. Pages are
205 * buffered inside the dio so that we can call get_user_pages() against a
206 * decent number of pages, less frequently. To provide nicer use of the
209 static inline struct page
*dio_get_page(struct dio
*dio
,
210 struct dio_submit
*sdio
)
212 if (dio_pages_present(sdio
) == 0) {
215 ret
= dio_refill_pages(dio
, sdio
);
218 BUG_ON(dio_pages_present(sdio
) == 0);
220 return dio
->pages
[sdio
->head
];
224 * dio_complete() - called when all DIO BIO I/O has been completed
225 * @offset: the byte offset in the file of the completed operation
227 * This drops i_dio_count, lets interested parties know that a DIO operation
228 * has completed, and calculates the resulting return code for the operation.
230 * It lets the filesystem know if it registered an interest earlier via
231 * get_block. Pass the private field of the map buffer_head so that
232 * filesystems can use it to hold additional state between get_block calls and
235 static ssize_t
dio_complete(struct dio
*dio
, ssize_t ret
, unsigned int flags
)
237 loff_t offset
= dio
->iocb
->ki_pos
;
238 ssize_t transferred
= 0;
242 * AIO submission can race with bio completion to get here while
243 * expecting to have the last io completed by bio completion.
244 * In that case -EIOCBQUEUED is in fact not an error we want
245 * to preserve through this call.
247 if (ret
== -EIOCBQUEUED
)
251 transferred
= dio
->result
;
253 /* Check for short read case */
254 if ((dio
->op
== REQ_OP_READ
) &&
255 ((offset
+ transferred
) > dio
->i_size
))
256 transferred
= dio
->i_size
- offset
;
257 /* ignore EFAULT if some IO has been done */
258 if (unlikely(ret
== -EFAULT
) && transferred
)
263 ret
= dio
->page_errors
;
271 err
= dio
->end_io(dio
->iocb
, offset
, ret
, dio
->private);
277 * Try again to invalidate clean pages which might have been cached by
278 * non-direct readahead, or faulted in by get_user_pages() if the source
279 * of the write was an mmap'ed region of the file we're writing. Either
280 * one is a pretty crazy thing to do, so we don't support it 100%. If
281 * this invalidation fails, tough, the write still worked...
283 * And this page cache invalidation has to be after dio->end_io(), as
284 * some filesystems convert unwritten extents to real allocations in
285 * end_io() when necessary, otherwise a racing buffer read would cache
286 * zeros from unwritten extents.
288 if (flags
& DIO_COMPLETE_INVALIDATE
&&
289 ret
> 0 && dio
->op
== REQ_OP_WRITE
&&
290 dio
->inode
->i_mapping
->nrpages
) {
291 err
= invalidate_inode_pages2_range(dio
->inode
->i_mapping
,
292 offset
>> PAGE_SHIFT
,
293 (offset
+ ret
- 1) >> PAGE_SHIFT
);
297 if (!(dio
->flags
& DIO_SKIP_DIO_COUNT
))
298 inode_dio_end(dio
->inode
);
300 if (flags
& DIO_COMPLETE_ASYNC
) {
302 * generic_write_sync expects ki_pos to have been updated
303 * already, but the submission path only does this for
306 dio
->iocb
->ki_pos
+= transferred
;
308 if (ret
> 0 && dio
->op
== REQ_OP_WRITE
)
309 ret
= generic_write_sync(dio
->iocb
, ret
);
310 dio
->iocb
->ki_complete(dio
->iocb
, ret
, 0);
313 kmem_cache_free(dio_cache
, dio
);
317 static void dio_aio_complete_work(struct work_struct
*work
)
319 struct dio
*dio
= container_of(work
, struct dio
, complete_work
);
321 dio_complete(dio
, 0, DIO_COMPLETE_ASYNC
| DIO_COMPLETE_INVALIDATE
);
324 static blk_status_t
dio_bio_complete(struct dio
*dio
, struct bio
*bio
);
327 * Asynchronous IO callback.
329 static void dio_bio_end_aio(struct bio
*bio
)
331 struct dio
*dio
= bio
->bi_private
;
332 unsigned long remaining
;
334 bool defer_completion
= false;
336 /* cleanup the bio */
337 dio_bio_complete(dio
, bio
);
339 spin_lock_irqsave(&dio
->bio_lock
, flags
);
340 remaining
= --dio
->refcount
;
341 if (remaining
== 1 && dio
->waiter
)
342 wake_up_process(dio
->waiter
);
343 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
345 if (remaining
== 0) {
347 * Defer completion when defer_completion is set or
348 * when the inode has pages mapped and this is AIO write.
349 * We need to invalidate those pages because there is a
350 * chance they contain stale data in the case buffered IO
351 * went in between AIO submission and completion into the
355 defer_completion
= dio
->defer_completion
||
356 (dio
->op
== REQ_OP_WRITE
&&
357 dio
->inode
->i_mapping
->nrpages
);
358 if (defer_completion
) {
359 INIT_WORK(&dio
->complete_work
, dio_aio_complete_work
);
360 queue_work(dio
->inode
->i_sb
->s_dio_done_wq
,
361 &dio
->complete_work
);
363 dio_complete(dio
, 0, DIO_COMPLETE_ASYNC
);
369 * The BIO completion handler simply queues the BIO up for the process-context
372 * During I/O bi_private points at the dio. After I/O, bi_private is used to
373 * implement a singly-linked list of completed BIOs, at dio->bio_list.
375 static void dio_bio_end_io(struct bio
*bio
)
377 struct dio
*dio
= bio
->bi_private
;
380 spin_lock_irqsave(&dio
->bio_lock
, flags
);
381 bio
->bi_private
= dio
->bio_list
;
383 if (--dio
->refcount
== 1 && dio
->waiter
)
384 wake_up_process(dio
->waiter
);
385 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
389 * dio_end_io - handle the end io action for the given bio
390 * @bio: The direct io bio thats being completed
392 * This is meant to be called by any filesystem that uses their own dio_submit_t
393 * so that the DIO specific endio actions are dealt with after the filesystem
394 * has done it's completion work.
396 void dio_end_io(struct bio
*bio
)
398 struct dio
*dio
= bio
->bi_private
;
401 dio_bio_end_aio(bio
);
405 EXPORT_SYMBOL_GPL(dio_end_io
);
408 dio_bio_alloc(struct dio
*dio
, struct dio_submit
*sdio
,
409 struct block_device
*bdev
,
410 sector_t first_sector
, int nr_vecs
)
415 * bio_alloc() is guaranteed to return a bio when called with
416 * __GFP_RECLAIM and we request a valid number of vectors.
418 bio
= bio_alloc(GFP_KERNEL
, nr_vecs
);
420 bio_set_dev(bio
, bdev
);
421 bio
->bi_iter
.bi_sector
= first_sector
;
422 bio_set_op_attrs(bio
, dio
->op
, dio
->op_flags
);
424 bio
->bi_end_io
= dio_bio_end_aio
;
426 bio
->bi_end_io
= dio_bio_end_io
;
428 bio
->bi_write_hint
= dio
->iocb
->ki_hint
;
431 sdio
->logical_offset_in_bio
= sdio
->cur_page_fs_offset
;
435 * In the AIO read case we speculatively dirty the pages before starting IO.
436 * During IO completion, any of these pages which happen to have been written
437 * back will be redirtied by bio_check_pages_dirty().
439 * bios hold a dio reference between submit_bio and ->end_io.
441 #ifdef CONFIG_CRYPTO_DISKCIPHER_DUN
442 static bool is_inode_filesystem_type(const struct inode
*inode
,
445 if (!inode
|| !fs_type
)
451 if (!inode
->i_sb
->s_type
)
454 return (strcmp(inode
->i_sb
->s_type
->name
, fs_type
) == 0);
458 static inline void dio_bio_submit(struct dio
*dio
, struct dio_submit
*sdio
)
460 struct bio
*bio
= sdio
->bio
;
463 bio
->bi_private
= dio
;
465 spin_lock_irqsave(&dio
->bio_lock
, flags
);
467 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
469 #if defined(CONFIG_CRYPTO_DISKCIPHER)
470 if (dio
->inode
&& fscrypt_has_encryption_key(dio
->inode
)) {
471 fscrypt_set_bio(dio
->inode
, bio
, 0);
472 crypto_diskcipher_debug(FS_DIO
, bio
->bi_opf
);
473 #if defined(CONFIG_CRYPTO_DISKCIPHER_DUN)
474 /* device unit number for iv sector */
475 #define PG_DUN(i,p) \
476 ((((i)->i_ino & 0xffffffff) << 32) | ((p) & 0xffffffff))
478 if (is_inode_filesystem_type(dio
->inode
, "f2fs"))
479 fscrypt_set_bio(dio
->inode
, bio
, PG_DUN(dio
->inode
,
480 (sdio
->logical_offset_in_bio
>> PAGE_SHIFT
)));
485 if (dio
->is_async
&& dio
->op
== REQ_OP_READ
&& dio
->should_dirty
)
486 bio_set_pages_dirty(bio
);
488 dio
->bio_disk
= bio
->bi_disk
;
490 if (sdio
->submit_io
) {
491 sdio
->submit_io(bio
, dio
->inode
, sdio
->logical_offset_in_bio
);
492 dio
->bio_cookie
= BLK_QC_T_NONE
;
494 dio
->bio_cookie
= submit_bio(bio
);
498 sdio
->logical_offset_in_bio
= 0;
502 * Release any resources in case of a failure
504 static inline void dio_cleanup(struct dio
*dio
, struct dio_submit
*sdio
)
506 while (sdio
->head
< sdio
->tail
)
507 put_page(dio
->pages
[sdio
->head
++]);
511 * Wait for the next BIO to complete. Remove it and return it. NULL is
512 * returned once all BIOs have been completed. This must only be called once
513 * all bios have been issued so that dio->refcount can only decrease. This
514 * requires that that the caller hold a reference on the dio.
516 static struct bio
*dio_await_one(struct dio
*dio
)
519 struct bio
*bio
= NULL
;
521 spin_lock_irqsave(&dio
->bio_lock
, flags
);
524 * Wait as long as the list is empty and there are bios in flight. bio
525 * completion drops the count, maybe adds to the list, and wakes while
526 * holding the bio_lock so we don't need set_current_state()'s barrier
527 * and can call it after testing our condition.
529 while (dio
->refcount
> 1 && dio
->bio_list
== NULL
) {
530 __set_current_state(TASK_UNINTERRUPTIBLE
);
531 dio
->waiter
= current
;
532 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
533 if (!(dio
->iocb
->ki_flags
& IOCB_HIPRI
) ||
534 !blk_mq_poll(dio
->bio_disk
->queue
, dio
->bio_cookie
))
536 /* wake up sets us TASK_RUNNING */
537 spin_lock_irqsave(&dio
->bio_lock
, flags
);
542 dio
->bio_list
= bio
->bi_private
;
544 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
549 * Process one completed BIO. No locks are held.
551 static blk_status_t
dio_bio_complete(struct dio
*dio
, struct bio
*bio
)
553 struct bio_vec
*bvec
;
555 blk_status_t err
= bio
->bi_status
;
558 if (err
== BLK_STS_AGAIN
&& (bio
->bi_opf
& REQ_NOWAIT
))
559 dio
->io_error
= -EAGAIN
;
561 dio
->io_error
= -EIO
;
564 if (dio
->is_async
&& dio
->op
== REQ_OP_READ
&& dio
->should_dirty
) {
565 bio_check_pages_dirty(bio
); /* transfers ownership */
567 bio_for_each_segment_all(bvec
, bio
, i
) {
568 struct page
*page
= bvec
->bv_page
;
570 if (dio
->op
== REQ_OP_READ
&& !PageCompound(page
) &&
572 set_page_dirty_lock(page
);
581 * Wait on and process all in-flight BIOs. This must only be called once
582 * all bios have been issued so that the refcount can only decrease.
583 * This just waits for all bios to make it through dio_bio_complete. IO
584 * errors are propagated through dio->io_error and should be propagated via
587 static void dio_await_completion(struct dio
*dio
)
591 bio
= dio_await_one(dio
);
593 dio_bio_complete(dio
, bio
);
598 * A really large O_DIRECT read or write can generate a lot of BIOs. So
599 * to keep the memory consumption sane we periodically reap any completed BIOs
600 * during the BIO generation phase.
602 * This also helps to limit the peak amount of pinned userspace memory.
604 static inline int dio_bio_reap(struct dio
*dio
, struct dio_submit
*sdio
)
608 if (sdio
->reap_counter
++ >= 64) {
609 while (dio
->bio_list
) {
614 spin_lock_irqsave(&dio
->bio_lock
, flags
);
616 dio
->bio_list
= bio
->bi_private
;
617 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
618 ret2
= blk_status_to_errno(dio_bio_complete(dio
, bio
));
622 sdio
->reap_counter
= 0;
628 * Create workqueue for deferred direct IO completions. We allocate the
629 * workqueue when it's first needed. This avoids creating workqueue for
630 * filesystems that don't need it and also allows us to create the workqueue
631 * late enough so the we can include s_id in the name of the workqueue.
633 int sb_init_dio_done_wq(struct super_block
*sb
)
635 struct workqueue_struct
*old
;
636 struct workqueue_struct
*wq
= alloc_workqueue("dio/%s",
642 * This has to be atomic as more DIOs can race to create the workqueue
644 old
= cmpxchg(&sb
->s_dio_done_wq
, NULL
, wq
);
645 /* Someone created workqueue before us? Free ours... */
647 destroy_workqueue(wq
);
651 static int dio_set_defer_completion(struct dio
*dio
)
653 struct super_block
*sb
= dio
->inode
->i_sb
;
655 if (dio
->defer_completion
)
657 dio
->defer_completion
= true;
658 if (!sb
->s_dio_done_wq
)
659 return sb_init_dio_done_wq(sb
);
664 * Call into the fs to map some more disk blocks. We record the current number
665 * of available blocks at sdio->blocks_available. These are in units of the
666 * fs blocksize, i_blocksize(inode).
668 * The fs is allowed to map lots of blocks at once. If it wants to do that,
669 * it uses the passed inode-relative block number as the file offset, as usual.
671 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
672 * has remaining to do. The fs should not map more than this number of blocks.
674 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
675 * indicate how much contiguous disk space has been made available at
678 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
679 * This isn't very efficient...
681 * In the case of filesystem holes: the fs may return an arbitrarily-large
682 * hole by returning an appropriate value in b_size and by clearing
683 * buffer_mapped(). However the direct-io code will only process holes one
684 * block at a time - it will repeatedly call get_block() as it walks the hole.
686 static int get_more_blocks(struct dio
*dio
, struct dio_submit
*sdio
,
687 struct buffer_head
*map_bh
)
690 sector_t fs_startblk
; /* Into file, in filesystem-sized blocks */
691 sector_t fs_endblk
; /* Into file, in filesystem-sized blocks */
692 unsigned long fs_count
; /* Number of filesystem-sized blocks */
694 unsigned int i_blkbits
= sdio
->blkbits
+ sdio
->blkfactor
;
698 * If there was a memory error and we've overwritten all the
699 * mapped blocks then we can now return that memory error
701 ret
= dio
->page_errors
;
703 BUG_ON(sdio
->block_in_file
>= sdio
->final_block_in_request
);
704 fs_startblk
= sdio
->block_in_file
>> sdio
->blkfactor
;
705 fs_endblk
= (sdio
->final_block_in_request
- 1) >>
707 fs_count
= fs_endblk
- fs_startblk
+ 1;
710 map_bh
->b_size
= fs_count
<< i_blkbits
;
713 * For writes that could fill holes inside i_size on a
714 * DIO_SKIP_HOLES filesystem we forbid block creations: only
715 * overwrites are permitted. We will return early to the caller
716 * once we see an unmapped buffer head returned, and the caller
717 * will fall back to buffered I/O.
719 * Otherwise the decision is left to the get_blocks method,
720 * which may decide to handle it or also return an unmapped
723 create
= dio
->op
== REQ_OP_WRITE
;
724 if (dio
->flags
& DIO_SKIP_HOLES
) {
725 i_size
= i_size_read(dio
->inode
);
726 if (i_size
&& fs_startblk
<= (i_size
- 1) >> i_blkbits
)
730 ret
= (*sdio
->get_block
)(dio
->inode
, fs_startblk
,
733 /* Store for completion */
734 dio
->private = map_bh
->b_private
;
736 if (ret
== 0 && buffer_defer_completion(map_bh
))
737 ret
= dio_set_defer_completion(dio
);
743 * There is no bio. Make one now.
745 static inline int dio_new_bio(struct dio
*dio
, struct dio_submit
*sdio
,
746 sector_t start_sector
, struct buffer_head
*map_bh
)
751 ret
= dio_bio_reap(dio
, sdio
);
754 sector
= start_sector
<< (sdio
->blkbits
- 9);
755 nr_pages
= min(sdio
->pages_in_io
, BIO_MAX_PAGES
);
756 BUG_ON(nr_pages
<= 0);
757 dio_bio_alloc(dio
, sdio
, map_bh
->b_bdev
, sector
, nr_pages
);
764 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
765 * that was successful then update final_block_in_bio and take a ref against
766 * the just-added page.
768 * Return zero on success. Non-zero means the caller needs to start a new BIO.
770 static inline int dio_bio_add_page(struct dio_submit
*sdio
)
774 ret
= bio_add_page(sdio
->bio
, sdio
->cur_page
,
775 sdio
->cur_page_len
, sdio
->cur_page_offset
);
776 if (ret
== sdio
->cur_page_len
) {
778 * Decrement count only, if we are done with this page
780 if ((sdio
->cur_page_len
+ sdio
->cur_page_offset
) == PAGE_SIZE
)
782 get_page(sdio
->cur_page
);
783 sdio
->final_block_in_bio
= sdio
->cur_page_block
+
784 (sdio
->cur_page_len
>> sdio
->blkbits
);
793 * Put cur_page under IO. The section of cur_page which is described by
794 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
795 * starts on-disk at cur_page_block.
797 * We take a ref against the page here (on behalf of its presence in the bio).
799 * The caller of this function is responsible for removing cur_page from the
800 * dio, and for dropping the refcount which came from that presence.
802 static inline int dio_send_cur_page(struct dio
*dio
, struct dio_submit
*sdio
,
803 struct buffer_head
*map_bh
)
808 loff_t cur_offset
= sdio
->cur_page_fs_offset
;
809 loff_t bio_next_offset
= sdio
->logical_offset_in_bio
+
810 sdio
->bio
->bi_iter
.bi_size
;
813 * See whether this new request is contiguous with the old.
815 * Btrfs cannot handle having logically non-contiguous requests
816 * submitted. For example if you have
818 * Logical: [0-4095][HOLE][8192-12287]
819 * Physical: [0-4095] [4096-8191]
821 * We cannot submit those pages together as one BIO. So if our
822 * current logical offset in the file does not equal what would
823 * be the next logical offset in the bio, submit the bio we
826 if (sdio
->final_block_in_bio
!= sdio
->cur_page_block
||
827 cur_offset
!= bio_next_offset
)
828 dio_bio_submit(dio
, sdio
);
831 if (sdio
->bio
== NULL
) {
832 ret
= dio_new_bio(dio
, sdio
, sdio
->cur_page_block
, map_bh
);
837 if (dio_bio_add_page(sdio
) != 0) {
838 dio_bio_submit(dio
, sdio
);
839 ret
= dio_new_bio(dio
, sdio
, sdio
->cur_page_block
, map_bh
);
841 ret
= dio_bio_add_page(sdio
);
850 * An autonomous function to put a chunk of a page under deferred IO.
852 * The caller doesn't actually know (or care) whether this piece of page is in
853 * a BIO, or is under IO or whatever. We just take care of all possible
854 * situations here. The separation between the logic of do_direct_IO() and
855 * that of submit_page_section() is important for clarity. Please don't break.
857 * The chunk of page starts on-disk at blocknr.
859 * We perform deferred IO, by recording the last-submitted page inside our
860 * private part of the dio structure. If possible, we just expand the IO
861 * across that page here.
863 * If that doesn't work out then we put the old page into the bio and add this
864 * page to the dio instead.
867 submit_page_section(struct dio
*dio
, struct dio_submit
*sdio
, struct page
*page
,
868 unsigned offset
, unsigned len
, sector_t blocknr
,
869 struct buffer_head
*map_bh
)
873 if (dio
->op
== REQ_OP_WRITE
) {
875 * Read accounting is performed in submit_bio()
877 task_io_account_write(len
);
881 * Can we just grow the current page's presence in the dio?
883 if (sdio
->cur_page
== page
&&
884 sdio
->cur_page_offset
+ sdio
->cur_page_len
== offset
&&
885 sdio
->cur_page_block
+
886 (sdio
->cur_page_len
>> sdio
->blkbits
) == blocknr
) {
887 sdio
->cur_page_len
+= len
;
892 * If there's a deferred page already there then send it.
894 if (sdio
->cur_page
) {
895 ret
= dio_send_cur_page(dio
, sdio
, map_bh
);
896 put_page(sdio
->cur_page
);
897 sdio
->cur_page
= NULL
;
902 get_page(page
); /* It is in dio */
903 sdio
->cur_page
= page
;
904 sdio
->cur_page_offset
= offset
;
905 sdio
->cur_page_len
= len
;
906 sdio
->cur_page_block
= blocknr
;
907 sdio
->cur_page_fs_offset
= sdio
->block_in_file
<< sdio
->blkbits
;
910 * If sdio->boundary then we want to schedule the IO now to
911 * avoid metadata seeks.
913 if (sdio
->boundary
) {
914 ret
= dio_send_cur_page(dio
, sdio
, map_bh
);
916 dio_bio_submit(dio
, sdio
);
917 put_page(sdio
->cur_page
);
918 sdio
->cur_page
= NULL
;
924 * If we are not writing the entire block and get_block() allocated
925 * the block for us, we need to fill-in the unused portion of the
926 * block with zeros. This happens only if user-buffer, fileoffset or
927 * io length is not filesystem block-size multiple.
929 * `end' is zero if we're doing the start of the IO, 1 at the end of the
932 static inline void dio_zero_block(struct dio
*dio
, struct dio_submit
*sdio
,
933 int end
, struct buffer_head
*map_bh
)
935 unsigned dio_blocks_per_fs_block
;
936 unsigned this_chunk_blocks
; /* In dio_blocks */
937 unsigned this_chunk_bytes
;
940 sdio
->start_zero_done
= 1;
941 if (!sdio
->blkfactor
|| !buffer_new(map_bh
))
944 dio_blocks_per_fs_block
= 1 << sdio
->blkfactor
;
945 this_chunk_blocks
= sdio
->block_in_file
& (dio_blocks_per_fs_block
- 1);
947 if (!this_chunk_blocks
)
951 * We need to zero out part of an fs block. It is either at the
952 * beginning or the end of the fs block.
955 this_chunk_blocks
= dio_blocks_per_fs_block
- this_chunk_blocks
;
957 this_chunk_bytes
= this_chunk_blocks
<< sdio
->blkbits
;
960 if (submit_page_section(dio
, sdio
, page
, 0, this_chunk_bytes
,
961 sdio
->next_block_for_io
, map_bh
))
964 sdio
->next_block_for_io
+= this_chunk_blocks
;
968 * Walk the user pages, and the file, mapping blocks to disk and generating
969 * a sequence of (page,offset,len,block) mappings. These mappings are injected
970 * into submit_page_section(), which takes care of the next stage of submission
972 * Direct IO against a blockdev is different from a file. Because we can
973 * happily perform page-sized but 512-byte aligned IOs. It is important that
974 * blockdev IO be able to have fine alignment and large sizes.
976 * So what we do is to permit the ->get_block function to populate bh.b_size
977 * with the size of IO which is permitted at this offset and this i_blkbits.
979 * For best results, the blockdev should be set up with 512-byte i_blkbits and
980 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
981 * fine alignment but still allows this function to work in PAGE_SIZE units.
983 static int do_direct_IO(struct dio
*dio
, struct dio_submit
*sdio
,
984 struct buffer_head
*map_bh
)
986 const unsigned blkbits
= sdio
->blkbits
;
987 const unsigned i_blkbits
= blkbits
+ sdio
->blkfactor
;
990 while (sdio
->block_in_file
< sdio
->final_block_in_request
) {
994 page
= dio_get_page(dio
, sdio
);
999 from
= sdio
->head
? 0 : sdio
->from
;
1000 to
= (sdio
->head
== sdio
->tail
- 1) ? sdio
->to
: PAGE_SIZE
;
1004 unsigned this_chunk_bytes
; /* # of bytes mapped */
1005 unsigned this_chunk_blocks
; /* # of blocks */
1008 if (sdio
->blocks_available
== 0) {
1010 * Need to go and map some more disk
1012 unsigned long blkmask
;
1013 unsigned long dio_remainder
;
1015 ret
= get_more_blocks(dio
, sdio
, map_bh
);
1020 if (!buffer_mapped(map_bh
))
1023 sdio
->blocks_available
=
1024 map_bh
->b_size
>> blkbits
;
1025 sdio
->next_block_for_io
=
1026 map_bh
->b_blocknr
<< sdio
->blkfactor
;
1027 if (buffer_new(map_bh
)) {
1031 map_bh
->b_size
>> i_blkbits
);
1034 if (!sdio
->blkfactor
)
1037 blkmask
= (1 << sdio
->blkfactor
) - 1;
1038 dio_remainder
= (sdio
->block_in_file
& blkmask
);
1041 * If we are at the start of IO and that IO
1042 * starts partway into a fs-block,
1043 * dio_remainder will be non-zero. If the IO
1044 * is a read then we can simply advance the IO
1045 * cursor to the first block which is to be
1046 * read. But if the IO is a write and the
1047 * block was newly allocated we cannot do that;
1048 * the start of the fs block must be zeroed out
1051 if (!buffer_new(map_bh
))
1052 sdio
->next_block_for_io
+= dio_remainder
;
1053 sdio
->blocks_available
-= dio_remainder
;
1057 if (!buffer_mapped(map_bh
)) {
1058 loff_t i_size_aligned
;
1060 /* AKPM: eargh, -ENOTBLK is a hack */
1061 if (dio
->op
== REQ_OP_WRITE
) {
1067 * Be sure to account for a partial block as the
1068 * last block in the file
1070 i_size_aligned
= ALIGN(i_size_read(dio
->inode
),
1072 if (sdio
->block_in_file
>=
1073 i_size_aligned
>> blkbits
) {
1078 zero_user(page
, from
, 1 << blkbits
);
1079 sdio
->block_in_file
++;
1080 from
+= 1 << blkbits
;
1081 dio
->result
+= 1 << blkbits
;
1086 * If we're performing IO which has an alignment which
1087 * is finer than the underlying fs, go check to see if
1088 * we must zero out the start of this block.
1090 if (unlikely(sdio
->blkfactor
&& !sdio
->start_zero_done
))
1091 dio_zero_block(dio
, sdio
, 0, map_bh
);
1094 * Work out, in this_chunk_blocks, how much disk we
1095 * can add to this page
1097 this_chunk_blocks
= sdio
->blocks_available
;
1098 u
= (to
- from
) >> blkbits
;
1099 if (this_chunk_blocks
> u
)
1100 this_chunk_blocks
= u
;
1101 u
= sdio
->final_block_in_request
- sdio
->block_in_file
;
1102 if (this_chunk_blocks
> u
)
1103 this_chunk_blocks
= u
;
1104 this_chunk_bytes
= this_chunk_blocks
<< blkbits
;
1105 BUG_ON(this_chunk_bytes
== 0);
1107 if (this_chunk_blocks
== sdio
->blocks_available
)
1108 sdio
->boundary
= buffer_boundary(map_bh
);
1109 ret
= submit_page_section(dio
, sdio
, page
,
1112 sdio
->next_block_for_io
,
1118 sdio
->next_block_for_io
+= this_chunk_blocks
;
1120 sdio
->block_in_file
+= this_chunk_blocks
;
1121 from
+= this_chunk_bytes
;
1122 dio
->result
+= this_chunk_bytes
;
1123 sdio
->blocks_available
-= this_chunk_blocks
;
1125 BUG_ON(sdio
->block_in_file
> sdio
->final_block_in_request
);
1126 if (sdio
->block_in_file
== sdio
->final_block_in_request
)
1130 /* Drop the ref which was taken in get_user_pages() */
1137 static inline int drop_refcount(struct dio
*dio
)
1140 unsigned long flags
;
1143 * Sync will always be dropping the final ref and completing the
1144 * operation. AIO can if it was a broken operation described above or
1145 * in fact if all the bios race to complete before we get here. In
1146 * that case dio_complete() translates the EIOCBQUEUED into the proper
1147 * return code that the caller will hand to ->complete().
1149 * This is managed by the bio_lock instead of being an atomic_t so that
1150 * completion paths can drop their ref and use the remaining count to
1151 * decide to wake the submission path atomically.
1153 spin_lock_irqsave(&dio
->bio_lock
, flags
);
1154 ret2
= --dio
->refcount
;
1155 spin_unlock_irqrestore(&dio
->bio_lock
, flags
);
1160 * This is a library function for use by filesystem drivers.
1162 * The locking rules are governed by the flags parameter:
1163 * - if the flags value contains DIO_LOCKING we use a fancy locking
1164 * scheme for dumb filesystems.
1165 * For writes this function is called under i_mutex and returns with
1166 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1167 * taken and dropped again before returning.
1168 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1169 * internal locking but rather rely on the filesystem to synchronize
1170 * direct I/O reads/writes versus each other and truncate.
1172 * To help with locking against truncate we incremented the i_dio_count
1173 * counter before starting direct I/O, and decrement it once we are done.
1174 * Truncate can wait for it to reach zero to provide exclusion. It is
1175 * expected that filesystem provide exclusion between new direct I/O
1176 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1177 * but other filesystems need to take care of this on their own.
1179 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1180 * is always inlined. Otherwise gcc is unable to split the structure into
1181 * individual fields and will generate much worse code. This is important
1182 * for the whole file.
1184 static inline ssize_t
1185 do_blockdev_direct_IO(struct kiocb
*iocb
, struct inode
*inode
,
1186 struct block_device
*bdev
, struct iov_iter
*iter
,
1187 get_block_t get_block
, dio_iodone_t end_io
,
1188 dio_submit_t submit_io
, int flags
)
1190 unsigned i_blkbits
= ACCESS_ONCE(inode
->i_blkbits
);
1191 unsigned blkbits
= i_blkbits
;
1192 unsigned blocksize_mask
= (1 << blkbits
) - 1;
1193 ssize_t retval
= -EINVAL
;
1194 size_t count
= iov_iter_count(iter
);
1195 loff_t offset
= iocb
->ki_pos
;
1196 loff_t end
= offset
+ count
;
1198 struct dio_submit sdio
= { 0, };
1199 struct buffer_head map_bh
= { 0, };
1200 struct blk_plug plug
;
1201 unsigned long align
= offset
| iov_iter_alignment(iter
);
1204 * Avoid references to bdev if not absolutely needed to give
1205 * the early prefetch in the caller enough time.
1208 if (align
& blocksize_mask
) {
1210 blkbits
= blksize_bits(bdev_logical_block_size(bdev
));
1211 blocksize_mask
= (1 << blkbits
) - 1;
1212 if (align
& blocksize_mask
)
1216 /* watch out for a 0 len io from a tricksy fs */
1217 if (iov_iter_rw(iter
) == READ
&& !iov_iter_count(iter
))
1220 dio
= kmem_cache_alloc(dio_cache
, GFP_KERNEL
);
1225 * Believe it or not, zeroing out the page array caused a .5%
1226 * performance regression in a database benchmark. So, we take
1227 * care to only zero out what's needed.
1229 memset(dio
, 0, offsetof(struct dio
, pages
));
1232 if (dio
->flags
& DIO_LOCKING
) {
1233 if (iov_iter_rw(iter
) == READ
) {
1234 struct address_space
*mapping
=
1235 iocb
->ki_filp
->f_mapping
;
1237 /* will be released by direct_io_worker */
1240 retval
= filemap_write_and_wait_range(mapping
, offset
,
1243 inode_unlock(inode
);
1244 kmem_cache_free(dio_cache
, dio
);
1250 /* Once we sampled i_size check for reads beyond EOF */
1251 dio
->i_size
= i_size_read(inode
);
1252 if (iov_iter_rw(iter
) == READ
&& offset
>= dio
->i_size
) {
1253 if (dio
->flags
& DIO_LOCKING
)
1254 inode_unlock(inode
);
1255 kmem_cache_free(dio_cache
, dio
);
1261 * For file extending writes updating i_size before data writeouts
1262 * complete can expose uninitialized blocks in dumb filesystems.
1263 * In that case we need to wait for I/O completion even if asked
1264 * for an asynchronous write.
1266 if (is_sync_kiocb(iocb
))
1267 dio
->is_async
= false;
1268 else if (!(dio
->flags
& DIO_ASYNC_EXTEND
) &&
1269 iov_iter_rw(iter
) == WRITE
&& end
> i_size_read(inode
))
1270 dio
->is_async
= false;
1272 dio
->is_async
= true;
1275 if (iov_iter_rw(iter
) == WRITE
) {
1276 dio
->op
= REQ_OP_WRITE
;
1277 dio
->op_flags
= REQ_SYNC
| REQ_IDLE
;
1278 if (iocb
->ki_flags
& IOCB_NOWAIT
)
1279 dio
->op_flags
|= REQ_NOWAIT
;
1281 dio
->op
= REQ_OP_READ
;
1285 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1286 * so that we can call ->fsync.
1288 if (dio
->is_async
&& iov_iter_rw(iter
) == WRITE
) {
1290 if (iocb
->ki_flags
& IOCB_DSYNC
)
1291 retval
= dio_set_defer_completion(dio
);
1292 else if (!dio
->inode
->i_sb
->s_dio_done_wq
) {
1294 * In case of AIO write racing with buffered read we
1295 * need to defer completion. We can't decide this now,
1296 * however the workqueue needs to be initialized here.
1298 retval
= sb_init_dio_done_wq(dio
->inode
->i_sb
);
1302 * We grab i_mutex only for reads so we don't have
1303 * to release it here
1305 kmem_cache_free(dio_cache
, dio
);
1311 * Will be decremented at I/O completion time.
1313 if (!(dio
->flags
& DIO_SKIP_DIO_COUNT
))
1314 inode_dio_begin(inode
);
1317 sdio
.blkbits
= blkbits
;
1318 sdio
.blkfactor
= i_blkbits
- blkbits
;
1319 sdio
.block_in_file
= offset
>> blkbits
;
1321 sdio
.get_block
= get_block
;
1322 dio
->end_io
= end_io
;
1323 sdio
.submit_io
= submit_io
;
1324 sdio
.final_block_in_bio
= -1;
1325 sdio
.next_block_for_io
= -1;
1329 spin_lock_init(&dio
->bio_lock
);
1332 dio
->should_dirty
= (iter
->type
== ITER_IOVEC
);
1334 sdio
.final_block_in_request
=
1335 (offset
+ iov_iter_count(iter
)) >> blkbits
;
1338 * In case of non-aligned buffers, we may need 2 more
1339 * pages since we need to zero out first and last block.
1341 if (unlikely(sdio
.blkfactor
))
1342 sdio
.pages_in_io
= 2;
1344 sdio
.pages_in_io
+= iov_iter_npages(iter
, INT_MAX
);
1346 blk_start_plug(&plug
);
1348 retval
= do_direct_IO(dio
, &sdio
, &map_bh
);
1350 dio_cleanup(dio
, &sdio
);
1352 if (retval
== -ENOTBLK
) {
1354 * The remaining part of the request will be
1355 * be handled by buffered I/O when we return
1360 * There may be some unwritten disk at the end of a part-written
1361 * fs-block-sized block. Go zero that now.
1363 dio_zero_block(dio
, &sdio
, 1, &map_bh
);
1365 if (sdio
.cur_page
) {
1368 ret2
= dio_send_cur_page(dio
, &sdio
, &map_bh
);
1371 put_page(sdio
.cur_page
);
1372 sdio
.cur_page
= NULL
;
1375 dio_bio_submit(dio
, &sdio
);
1377 blk_finish_plug(&plug
);
1380 * It is possible that, we return short IO due to end of file.
1381 * In that case, we need to release all the pages we got hold on.
1383 dio_cleanup(dio
, &sdio
);
1386 * All block lookups have been performed. For READ requests
1387 * we can let i_mutex go now that its achieved its purpose
1388 * of protecting us from looking up uninitialized blocks.
1390 if (iov_iter_rw(iter
) == READ
&& (dio
->flags
& DIO_LOCKING
))
1391 inode_unlock(dio
->inode
);
1394 * The only time we want to leave bios in flight is when a successful
1395 * partial aio read or full aio write have been setup. In that case
1396 * bio completion will call aio_complete. The only time it's safe to
1397 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1398 * This had *better* be the only place that raises -EIOCBQUEUED.
1400 BUG_ON(retval
== -EIOCBQUEUED
);
1401 if (dio
->is_async
&& retval
== 0 && dio
->result
&&
1402 (iov_iter_rw(iter
) == READ
|| dio
->result
== count
))
1403 retval
= -EIOCBQUEUED
;
1405 dio_await_completion(dio
);
1407 if (drop_refcount(dio
) == 0) {
1408 retval
= dio_complete(dio
, retval
, DIO_COMPLETE_INVALIDATE
);
1410 BUG_ON(retval
!= -EIOCBQUEUED
);
1416 ssize_t
__blockdev_direct_IO(struct kiocb
*iocb
, struct inode
*inode
,
1417 struct block_device
*bdev
, struct iov_iter
*iter
,
1418 get_block_t get_block
,
1419 dio_iodone_t end_io
, dio_submit_t submit_io
,
1423 * The block device state is needed in the end to finally
1424 * submit everything. Since it's likely to be cache cold
1425 * prefetch it here as first thing to hide some of the
1428 * Attempt to prefetch the pieces we likely need later.
1430 prefetch(&bdev
->bd_disk
->part_tbl
);
1431 prefetch(bdev
->bd_queue
);
1432 prefetch((char *)bdev
->bd_queue
+ SMP_CACHE_BYTES
);
1434 return do_blockdev_direct_IO(iocb
, inode
, bdev
, iter
, get_block
,
1435 end_io
, submit_io
, flags
);
1438 EXPORT_SYMBOL(__blockdev_direct_IO
);
1440 static __init
int dio_init(void)
1442 dio_cache
= KMEM_CACHE(dio
, SLAB_PANIC
);
1445 module_init(dio_init
)