4 * Copyright (C) 2002, Linus Torvalds.
8 * 04Jul2002 Andrew Morton
10 * 11Sep2002 janetinc@us.ibm.com
11 * added readv/writev support.
12 * 29Oct2002 Andrew Morton
13 * rewrote bio_add_page() support.
14 * 30Oct2002 pbadari@us.ibm.com
15 * added support for non-aligned IO.
16 * 06Nov2002 pbadari@us.ibm.com
17 * added asynchronous IO support.
18 * 21Jul2003 nathans@sgi.com
19 * added IO completion notifier.
22 #include <linux/kernel.h>
23 #include <linux/module.h>
24 #include <linux/types.h>
27 #include <linux/slab.h>
28 #include <linux/highmem.h>
29 #include <linux/pagemap.h>
30 #include <linux/task_io_accounting_ops.h>
31 #include <linux/bio.h>
32 #include <linux/wait.h>
33 #include <linux/err.h>
34 #include <linux/blkdev.h>
35 #include <linux/buffer_head.h>
36 #include <linux/rwsem.h>
37 #include <linux/uio.h>
38 #include <linux/atomic.h>
39 #include <linux/prefetch.h>
40 #include <linux/aio.h>
43 * How many user pages to map in one call to get_user_pages(). This determines
44 * the size of a structure in the slab cache
49 * This code generally works in units of "dio_blocks". A dio_block is
50 * somewhere between the hard sector size and the filesystem block size. it
51 * is determined on a per-invocation basis. When talking to the filesystem
52 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
53 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
54 * to bio_block quantities by shifting left by blkfactor.
56 * If blkfactor is zero then the user's request was aligned to the filesystem's
60 /* dio_state only used in the submission path */
63 struct bio *bio; /* bio under assembly */
64 unsigned blkbits; /* doesn't change */
65 unsigned blkfactor; /* When we're using an alignment which
66 is finer than the filesystem's soft
67 blocksize, this specifies how much
68 finer. blkfactor=2 means 1/4-block
69 alignment. Does not change */
70 unsigned start_zero_done; /* flag: sub-blocksize zeroing has
71 been performed at the start of a
73 int pages_in_io; /* approximate total IO pages */
74 size_t size; /* total request size (doesn't change)*/
75 sector_t block_in_file; /* Current offset into the underlying
76 file in dio_block units. */
77 unsigned blocks_available; /* At block_in_file. changes */
78 int reap_counter; /* rate limit reaping */
79 sector_t final_block_in_request;/* doesn't change */
80 unsigned first_block_in_page; /* doesn't change, Used only once */
81 int boundary; /* prev block is at a boundary */
82 get_block_t *get_block; /* block mapping function */
83 dio_submit_t *submit_io; /* IO submition function */
85 loff_t logical_offset_in_bio; /* current first logical block in bio */
86 sector_t final_block_in_bio; /* current final block in bio + 1 */
87 sector_t next_block_for_io; /* next block to be put under IO,
88 in dio_blocks units */
91 * Deferred addition of a page to the dio. These variables are
92 * private to dio_send_cur_page(), submit_page_section() and
95 struct page *cur_page; /* The page */
96 unsigned cur_page_offset; /* Offset into it, in bytes */
97 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
98 sector_t cur_page_block; /* Where it starts */
99 loff_t cur_page_fs_offset; /* Offset in file */
102 * Page fetching state. These variables belong to dio_refill_pages().
104 int curr_page; /* changes */
105 int total_pages; /* doesn't change */
106 unsigned long curr_user_address;/* changes */
109 * Page queue. These variables belong to dio_refill_pages() and
112 unsigned head; /* next page to process */
113 unsigned tail; /* last valid page + 1 */
116 /* dio_state communicated between submission path and end_io */
118 int flags; /* doesn't change */
121 loff_t i_size; /* i_size when submitted */
122 dio_iodone_t *end_io; /* IO completion function */
124 void *private; /* copy from map_bh.b_private */
126 /* BIO completion state */
127 spinlock_t bio_lock; /* protects BIO fields below */
128 int page_errors; /* errno from get_user_pages() */
129 int is_async; /* is IO async ? */
130 int io_error; /* IO error in completion path */
131 unsigned long refcount; /* direct_io_worker() and bios */
132 struct bio *bio_list; /* singly linked via bi_private */
133 struct task_struct *waiter; /* waiting task (NULL if none) */
135 /* AIO related stuff */
136 struct kiocb *iocb; /* kiocb */
137 ssize_t result; /* IO result */
140 * pages[] (and any fields placed after it) are not zeroed out at
141 * allocation time. Don't add new fields after pages[] unless you
142 * wish that they not be zeroed.
144 struct page *pages[DIO_PAGES]; /* page buffer */
145 } ____cacheline_aligned_in_smp;
147 static struct kmem_cache *dio_cache __read_mostly;
150 * How many pages are in the queue?
152 static inline unsigned dio_pages_present(struct dio_submit *sdio)
154 return sdio->tail - sdio->head;
158 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
160 static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
165 nr_pages = min(sdio->total_pages - sdio->curr_page, DIO_PAGES);
166 ret = get_user_pages_fast(
167 sdio->curr_user_address, /* Where from? */
168 nr_pages, /* How many pages? */
169 dio->rw == READ, /* Write to memory? */
170 &dio->pages[0]); /* Put results here */
172 if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) {
173 struct page *page = ZERO_PAGE(0);
175 * A memory fault, but the filesystem has some outstanding
176 * mapped blocks. We need to use those blocks up to avoid
177 * leaking stale data in the file.
179 if (dio->page_errors == 0)
180 dio->page_errors = ret;
181 page_cache_get(page);
182 dio->pages[0] = page;
190 sdio->curr_user_address += ret * PAGE_SIZE;
191 sdio->curr_page += ret;
201 * Get another userspace page. Returns an ERR_PTR on error. Pages are
202 * buffered inside the dio so that we can call get_user_pages() against a
203 * decent number of pages, less frequently. To provide nicer use of the
206 static inline struct page *dio_get_page(struct dio *dio,
207 struct dio_submit *sdio)
209 if (dio_pages_present(sdio) == 0) {
212 ret = dio_refill_pages(dio, sdio);
215 BUG_ON(dio_pages_present(sdio) == 0);
217 return dio->pages[sdio->head++];
221 * dio_complete() - called when all DIO BIO I/O has been completed
222 * @offset: the byte offset in the file of the completed operation
224 * This releases locks as dictated by the locking type, lets interested parties
225 * know that a DIO operation has completed, and calculates the resulting return
226 * code for the operation.
228 * It lets the filesystem know if it registered an interest earlier via
229 * get_block. Pass the private field of the map buffer_head so that
230 * filesystems can use it to hold additional state between get_block calls and
233 static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret, bool is_async)
235 ssize_t transferred = 0;
238 * AIO submission can race with bio completion to get here while
239 * expecting to have the last io completed by bio completion.
240 * In that case -EIOCBQUEUED is in fact not an error we want
241 * to preserve through this call.
243 if (ret == -EIOCBQUEUED)
247 transferred = dio->result;
249 /* Check for short read case */
250 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
251 transferred = dio->i_size - offset;
255 ret = dio->page_errors;
261 if (dio->end_io && dio->result) {
262 dio->end_io(dio->iocb, offset, transferred,
263 dio->private, ret, is_async);
265 inode_dio_done(dio->inode);
267 aio_complete(dio->iocb, ret, 0);
273 static int dio_bio_complete(struct dio *dio, struct bio *bio);
275 * Asynchronous IO callback.
277 static void dio_bio_end_aio(struct bio *bio, int error)
279 struct dio *dio = bio->bi_private;
280 unsigned long remaining;
283 /* cleanup the bio */
284 dio_bio_complete(dio, bio);
286 spin_lock_irqsave(&dio->bio_lock, flags);
287 remaining = --dio->refcount;
288 if (remaining == 1 && dio->waiter)
289 wake_up_process(dio->waiter);
290 spin_unlock_irqrestore(&dio->bio_lock, flags);
292 if (remaining == 0) {
293 dio_complete(dio, dio->iocb->ki_pos, 0, true);
294 kmem_cache_free(dio_cache, dio);
299 * The BIO completion handler simply queues the BIO up for the process-context
302 * During I/O bi_private points at the dio. After I/O, bi_private is used to
303 * implement a singly-linked list of completed BIOs, at dio->bio_list.
305 static void dio_bio_end_io(struct bio *bio, int error)
307 struct dio *dio = bio->bi_private;
310 spin_lock_irqsave(&dio->bio_lock, flags);
311 bio->bi_private = dio->bio_list;
313 if (--dio->refcount == 1 && dio->waiter)
314 wake_up_process(dio->waiter);
315 spin_unlock_irqrestore(&dio->bio_lock, flags);
319 * dio_end_io - handle the end io action for the given bio
320 * @bio: The direct io bio thats being completed
321 * @error: Error if there was one
323 * This is meant to be called by any filesystem that uses their own dio_submit_t
324 * so that the DIO specific endio actions are dealt with after the filesystem
325 * has done it's completion work.
327 void dio_end_io(struct bio *bio, int error, struct batch_complete *batch)
329 struct dio *dio = bio->bi_private;
332 dio_bio_end_aio(bio, error, batch);
334 dio_bio_end_io(bio, error);
336 EXPORT_SYMBOL_GPL(dio_end_io);
339 dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
340 struct block_device *bdev,
341 sector_t first_sector, int nr_vecs)
346 * bio_alloc() is guaranteed to return a bio when called with
347 * __GFP_WAIT and we request a valid number of vectors.
349 bio = bio_alloc(GFP_KERNEL, nr_vecs);
352 bio->bi_sector = first_sector;
353 bio->bi_end_io = dio_end_io;
356 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
360 * In the AIO read case we speculatively dirty the pages before starting IO.
361 * During IO completion, any of these pages which happen to have been written
362 * back will be redirtied by bio_check_pages_dirty().
364 * bios hold a dio reference between submit_bio and ->end_io.
366 static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
368 struct bio *bio = sdio->bio;
371 bio->bi_private = dio;
373 spin_lock_irqsave(&dio->bio_lock, flags);
375 spin_unlock_irqrestore(&dio->bio_lock, flags);
377 if (dio->is_async && dio->rw == READ)
378 bio_set_pages_dirty(bio);
381 sdio->submit_io(dio->rw, bio, dio->inode,
382 sdio->logical_offset_in_bio);
384 submit_bio(dio->rw, bio);
388 sdio->logical_offset_in_bio = 0;
392 * Release any resources in case of a failure
394 static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
396 while (dio_pages_present(sdio))
397 page_cache_release(dio_get_page(dio, sdio));
401 * Wait for the next BIO to complete. Remove it and return it. NULL is
402 * returned once all BIOs have been completed. This must only be called once
403 * all bios have been issued so that dio->refcount can only decrease. This
404 * requires that that the caller hold a reference on the dio.
406 static struct bio *dio_await_one(struct dio *dio)
409 struct bio *bio = NULL;
411 spin_lock_irqsave(&dio->bio_lock, flags);
414 * Wait as long as the list is empty and there are bios in flight. bio
415 * completion drops the count, maybe adds to the list, and wakes while
416 * holding the bio_lock so we don't need set_current_state()'s barrier
417 * and can call it after testing our condition.
419 while (dio->refcount > 1 && dio->bio_list == NULL) {
420 __set_current_state(TASK_UNINTERRUPTIBLE);
421 dio->waiter = current;
422 spin_unlock_irqrestore(&dio->bio_lock, flags);
424 /* wake up sets us TASK_RUNNING */
425 spin_lock_irqsave(&dio->bio_lock, flags);
430 dio->bio_list = bio->bi_private;
432 spin_unlock_irqrestore(&dio->bio_lock, flags);
437 * Process one completed BIO. No locks are held.
439 static int dio_bio_complete(struct dio *dio, struct bio *bio)
441 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
442 struct bio_vec *bvec;
446 dio->io_error = -EIO;
448 if (dio->is_async && dio->rw == READ) {
449 bio_check_pages_dirty(bio); /* transfers ownership */
451 bio_for_each_segment_all(bvec, bio, i) {
452 struct page *page = bvec->bv_page;
454 if (dio->rw == READ && !PageCompound(page))
455 set_page_dirty_lock(page);
456 page_cache_release(page);
460 return uptodate ? 0 : -EIO;
464 * Wait on and process all in-flight BIOs. This must only be called once
465 * all bios have been issued so that the refcount can only decrease.
466 * This just waits for all bios to make it through dio_bio_complete. IO
467 * errors are propagated through dio->io_error and should be propagated via
470 static void dio_await_completion(struct dio *dio)
474 bio = dio_await_one(dio);
476 dio_bio_complete(dio, bio);
481 * A really large O_DIRECT read or write can generate a lot of BIOs. So
482 * to keep the memory consumption sane we periodically reap any completed BIOs
483 * during the BIO generation phase.
485 * This also helps to limit the peak amount of pinned userspace memory.
487 static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
491 if (sdio->reap_counter++ >= 64) {
492 while (dio->bio_list) {
497 spin_lock_irqsave(&dio->bio_lock, flags);
499 dio->bio_list = bio->bi_private;
500 spin_unlock_irqrestore(&dio->bio_lock, flags);
501 ret2 = dio_bio_complete(dio, bio);
505 sdio->reap_counter = 0;
511 * Call into the fs to map some more disk blocks. We record the current number
512 * of available blocks at sdio->blocks_available. These are in units of the
513 * fs blocksize, (1 << inode->i_blkbits).
515 * The fs is allowed to map lots of blocks at once. If it wants to do that,
516 * it uses the passed inode-relative block number as the file offset, as usual.
518 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
519 * has remaining to do. The fs should not map more than this number of blocks.
521 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
522 * indicate how much contiguous disk space has been made available at
525 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
526 * This isn't very efficient...
528 * In the case of filesystem holes: the fs may return an arbitrarily-large
529 * hole by returning an appropriate value in b_size and by clearing
530 * buffer_mapped(). However the direct-io code will only process holes one
531 * block at a time - it will repeatedly call get_block() as it walks the hole.
533 static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
534 struct buffer_head *map_bh)
537 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
538 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
539 unsigned long fs_count; /* Number of filesystem-sized blocks */
541 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
544 * If there was a memory error and we've overwritten all the
545 * mapped blocks then we can now return that memory error
547 ret = dio->page_errors;
549 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
550 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
551 fs_endblk = (sdio->final_block_in_request - 1) >>
553 fs_count = fs_endblk - fs_startblk + 1;
556 map_bh->b_size = fs_count << i_blkbits;
559 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we
560 * forbid block creations: only overwrites are permitted.
561 * We will return early to the caller once we see an
562 * unmapped buffer head returned, and the caller will fall
563 * back to buffered I/O.
565 * Otherwise the decision is left to the get_blocks method,
566 * which may decide to handle it or also return an unmapped
569 create = dio->rw & WRITE;
570 if (dio->flags & DIO_SKIP_HOLES) {
571 if (sdio->block_in_file < (i_size_read(dio->inode) >>
576 ret = (*sdio->get_block)(dio->inode, fs_startblk,
579 /* Store for completion */
580 dio->private = map_bh->b_private;
586 * There is no bio. Make one now.
588 static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
589 sector_t start_sector, struct buffer_head *map_bh)
594 ret = dio_bio_reap(dio, sdio);
597 sector = start_sector << (sdio->blkbits - 9);
598 nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(map_bh->b_bdev));
599 nr_pages = min(nr_pages, BIO_MAX_PAGES);
600 BUG_ON(nr_pages <= 0);
601 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
608 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
609 * that was successful then update final_block_in_bio and take a ref against
610 * the just-added page.
612 * Return zero on success. Non-zero means the caller needs to start a new BIO.
614 static inline int dio_bio_add_page(struct dio_submit *sdio)
618 ret = bio_add_page(sdio->bio, sdio->cur_page,
619 sdio->cur_page_len, sdio->cur_page_offset);
620 if (ret == sdio->cur_page_len) {
622 * Decrement count only, if we are done with this page
624 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
626 page_cache_get(sdio->cur_page);
627 sdio->final_block_in_bio = sdio->cur_page_block +
628 (sdio->cur_page_len >> sdio->blkbits);
637 * Put cur_page under IO. The section of cur_page which is described by
638 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
639 * starts on-disk at cur_page_block.
641 * We take a ref against the page here (on behalf of its presence in the bio).
643 * The caller of this function is responsible for removing cur_page from the
644 * dio, and for dropping the refcount which came from that presence.
646 static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
647 struct buffer_head *map_bh)
652 loff_t cur_offset = sdio->cur_page_fs_offset;
653 loff_t bio_next_offset = sdio->logical_offset_in_bio +
657 * See whether this new request is contiguous with the old.
659 * Btrfs cannot handle having logically non-contiguous requests
660 * submitted. For example if you have
662 * Logical: [0-4095][HOLE][8192-12287]
663 * Physical: [0-4095] [4096-8191]
665 * We cannot submit those pages together as one BIO. So if our
666 * current logical offset in the file does not equal what would
667 * be the next logical offset in the bio, submit the bio we
670 if (sdio->final_block_in_bio != sdio->cur_page_block ||
671 cur_offset != bio_next_offset)
672 dio_bio_submit(dio, sdio);
674 * Submit now if the underlying fs is about to perform a
677 else if (sdio->boundary)
678 dio_bio_submit(dio, sdio);
681 if (sdio->bio == NULL) {
682 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
687 if (dio_bio_add_page(sdio) != 0) {
688 dio_bio_submit(dio, sdio);
689 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
691 ret = dio_bio_add_page(sdio);
700 * An autonomous function to put a chunk of a page under deferred IO.
702 * The caller doesn't actually know (or care) whether this piece of page is in
703 * a BIO, or is under IO or whatever. We just take care of all possible
704 * situations here. The separation between the logic of do_direct_IO() and
705 * that of submit_page_section() is important for clarity. Please don't break.
707 * The chunk of page starts on-disk at blocknr.
709 * We perform deferred IO, by recording the last-submitted page inside our
710 * private part of the dio structure. If possible, we just expand the IO
711 * across that page here.
713 * If that doesn't work out then we put the old page into the bio and add this
714 * page to the dio instead.
717 submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
718 unsigned offset, unsigned len, sector_t blocknr,
719 struct buffer_head *map_bh)
723 if (dio->rw & WRITE) {
725 * Read accounting is performed in submit_bio()
727 task_io_account_write(len);
731 * Can we just grow the current page's presence in the dio?
733 if (sdio->cur_page == page &&
734 sdio->cur_page_offset + sdio->cur_page_len == offset &&
735 sdio->cur_page_block +
736 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
737 sdio->cur_page_len += len;
740 * If sdio->boundary then we want to schedule the IO now to
741 * avoid metadata seeks.
743 if (sdio->boundary) {
744 ret = dio_send_cur_page(dio, sdio, map_bh);
745 page_cache_release(sdio->cur_page);
746 sdio->cur_page = NULL;
752 * If there's a deferred page already there then send it.
754 if (sdio->cur_page) {
755 ret = dio_send_cur_page(dio, sdio, map_bh);
756 page_cache_release(sdio->cur_page);
757 sdio->cur_page = NULL;
762 page_cache_get(page); /* It is in dio */
763 sdio->cur_page = page;
764 sdio->cur_page_offset = offset;
765 sdio->cur_page_len = len;
766 sdio->cur_page_block = blocknr;
767 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
773 * Clean any dirty buffers in the blockdev mapping which alias newly-created
774 * file blocks. Only called for S_ISREG files - blockdevs do not set
777 static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
782 nblocks = map_bh->b_size >> dio->inode->i_blkbits;
784 for (i = 0; i < nblocks; i++) {
785 unmap_underlying_metadata(map_bh->b_bdev,
786 map_bh->b_blocknr + i);
791 * If we are not writing the entire block and get_block() allocated
792 * the block for us, we need to fill-in the unused portion of the
793 * block with zeros. This happens only if user-buffer, fileoffset or
794 * io length is not filesystem block-size multiple.
796 * `end' is zero if we're doing the start of the IO, 1 at the end of the
799 static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
800 int end, struct buffer_head *map_bh)
802 unsigned dio_blocks_per_fs_block;
803 unsigned this_chunk_blocks; /* In dio_blocks */
804 unsigned this_chunk_bytes;
807 sdio->start_zero_done = 1;
808 if (!sdio->blkfactor || !buffer_new(map_bh))
811 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
812 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
814 if (!this_chunk_blocks)
818 * We need to zero out part of an fs block. It is either at the
819 * beginning or the end of the fs block.
822 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
824 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
827 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
828 sdio->next_block_for_io, map_bh))
831 sdio->next_block_for_io += this_chunk_blocks;
835 * Walk the user pages, and the file, mapping blocks to disk and generating
836 * a sequence of (page,offset,len,block) mappings. These mappings are injected
837 * into submit_page_section(), which takes care of the next stage of submission
839 * Direct IO against a blockdev is different from a file. Because we can
840 * happily perform page-sized but 512-byte aligned IOs. It is important that
841 * blockdev IO be able to have fine alignment and large sizes.
843 * So what we do is to permit the ->get_block function to populate bh.b_size
844 * with the size of IO which is permitted at this offset and this i_blkbits.
846 * For best results, the blockdev should be set up with 512-byte i_blkbits and
847 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
848 * fine alignment but still allows this function to work in PAGE_SIZE units.
850 static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
851 struct buffer_head *map_bh)
853 const unsigned blkbits = sdio->blkbits;
854 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
856 unsigned block_in_page;
859 /* The I/O can start at any block offset within the first page */
860 block_in_page = sdio->first_block_in_page;
862 while (sdio->block_in_file < sdio->final_block_in_request) {
863 page = dio_get_page(dio, sdio);
869 while (block_in_page < blocks_per_page) {
870 unsigned offset_in_page = block_in_page << blkbits;
871 unsigned this_chunk_bytes; /* # of bytes mapped */
872 unsigned this_chunk_blocks; /* # of blocks */
875 if (sdio->blocks_available == 0) {
877 * Need to go and map some more disk
879 unsigned long blkmask;
880 unsigned long dio_remainder;
882 ret = get_more_blocks(dio, sdio, map_bh);
884 page_cache_release(page);
887 if (!buffer_mapped(map_bh))
890 sdio->blocks_available =
891 map_bh->b_size >> sdio->blkbits;
892 sdio->next_block_for_io =
893 map_bh->b_blocknr << sdio->blkfactor;
894 if (buffer_new(map_bh))
895 clean_blockdev_aliases(dio, map_bh);
897 if (!sdio->blkfactor)
900 blkmask = (1 << sdio->blkfactor) - 1;
901 dio_remainder = (sdio->block_in_file & blkmask);
904 * If we are at the start of IO and that IO
905 * starts partway into a fs-block,
906 * dio_remainder will be non-zero. If the IO
907 * is a read then we can simply advance the IO
908 * cursor to the first block which is to be
909 * read. But if the IO is a write and the
910 * block was newly allocated we cannot do that;
911 * the start of the fs block must be zeroed out
914 if (!buffer_new(map_bh))
915 sdio->next_block_for_io += dio_remainder;
916 sdio->blocks_available -= dio_remainder;
920 if (!buffer_mapped(map_bh)) {
921 loff_t i_size_aligned;
923 /* AKPM: eargh, -ENOTBLK is a hack */
924 if (dio->rw & WRITE) {
925 page_cache_release(page);
930 * Be sure to account for a partial block as the
931 * last block in the file
933 i_size_aligned = ALIGN(i_size_read(dio->inode),
935 if (sdio->block_in_file >=
936 i_size_aligned >> blkbits) {
938 page_cache_release(page);
941 zero_user(page, block_in_page << blkbits,
943 sdio->block_in_file++;
949 * If we're performing IO which has an alignment which
950 * is finer than the underlying fs, go check to see if
951 * we must zero out the start of this block.
953 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
954 dio_zero_block(dio, sdio, 0, map_bh);
957 * Work out, in this_chunk_blocks, how much disk we
958 * can add to this page
960 this_chunk_blocks = sdio->blocks_available;
961 u = (PAGE_SIZE - offset_in_page) >> blkbits;
962 if (this_chunk_blocks > u)
963 this_chunk_blocks = u;
964 u = sdio->final_block_in_request - sdio->block_in_file;
965 if (this_chunk_blocks > u)
966 this_chunk_blocks = u;
967 this_chunk_bytes = this_chunk_blocks << blkbits;
968 BUG_ON(this_chunk_bytes == 0);
970 if (this_chunk_blocks == sdio->blocks_available)
971 sdio->boundary = buffer_boundary(map_bh);
972 ret = submit_page_section(dio, sdio, page,
975 sdio->next_block_for_io,
978 page_cache_release(page);
981 sdio->next_block_for_io += this_chunk_blocks;
983 sdio->block_in_file += this_chunk_blocks;
984 block_in_page += this_chunk_blocks;
985 sdio->blocks_available -= this_chunk_blocks;
987 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
988 if (sdio->block_in_file == sdio->final_block_in_request)
992 /* Drop the ref which was taken in get_user_pages() */
993 page_cache_release(page);
1000 static inline int drop_refcount(struct dio *dio)
1003 unsigned long flags;
1006 * Sync will always be dropping the final ref and completing the
1007 * operation. AIO can if it was a broken operation described above or
1008 * in fact if all the bios race to complete before we get here. In
1009 * that case dio_complete() translates the EIOCBQUEUED into the proper
1010 * return code that the caller will hand to aio_complete().
1012 * This is managed by the bio_lock instead of being an atomic_t so that
1013 * completion paths can drop their ref and use the remaining count to
1014 * decide to wake the submission path atomically.
1016 spin_lock_irqsave(&dio->bio_lock, flags);
1017 ret2 = --dio->refcount;
1018 spin_unlock_irqrestore(&dio->bio_lock, flags);
1023 * This is a library function for use by filesystem drivers.
1025 * The locking rules are governed by the flags parameter:
1026 * - if the flags value contains DIO_LOCKING we use a fancy locking
1027 * scheme for dumb filesystems.
1028 * For writes this function is called under i_mutex and returns with
1029 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1030 * taken and dropped again before returning.
1031 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1032 * internal locking but rather rely on the filesystem to synchronize
1033 * direct I/O reads/writes versus each other and truncate.
1035 * To help with locking against truncate we incremented the i_dio_count
1036 * counter before starting direct I/O, and decrement it once we are done.
1037 * Truncate can wait for it to reach zero to provide exclusion. It is
1038 * expected that filesystem provide exclusion between new direct I/O
1039 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1040 * but other filesystems need to take care of this on their own.
1042 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1043 * is always inlined. Otherwise gcc is unable to split the structure into
1044 * individual fields and will generate much worse code. This is important
1045 * for the whole file.
1047 static inline ssize_t
1048 do_blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1049 struct block_device *bdev, const struct iovec *iov, loff_t offset,
1050 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1051 dio_submit_t submit_io, int flags)
1056 unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits);
1057 unsigned blkbits = i_blkbits;
1058 unsigned blocksize_mask = (1 << blkbits) - 1;
1059 ssize_t retval = -EINVAL;
1060 loff_t end = offset;
1062 struct dio_submit sdio = { 0, };
1063 unsigned long user_addr;
1065 struct buffer_head map_bh = { 0, };
1066 struct blk_plug plug;
1072 * Avoid references to bdev if not absolutely needed to give
1073 * the early prefetch in the caller enough time.
1076 if (offset & blocksize_mask) {
1078 blkbits = blksize_bits(bdev_logical_block_size(bdev));
1079 blocksize_mask = (1 << blkbits) - 1;
1080 if (offset & blocksize_mask)
1084 /* Check the memory alignment. Blocks cannot straddle pages */
1085 for (seg = 0; seg < nr_segs; seg++) {
1086 addr = (unsigned long)iov[seg].iov_base;
1087 size = iov[seg].iov_len;
1089 if (unlikely((addr & blocksize_mask) ||
1090 (size & blocksize_mask))) {
1092 blkbits = blksize_bits(
1093 bdev_logical_block_size(bdev));
1094 blocksize_mask = (1 << blkbits) - 1;
1095 if ((addr & blocksize_mask) || (size & blocksize_mask))
1100 /* watch out for a 0 len io from a tricksy fs */
1101 if (rw == READ && end == offset)
1104 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
1109 * Believe it or not, zeroing out the page array caused a .5%
1110 * performance regression in a database benchmark. So, we take
1111 * care to only zero out what's needed.
1113 memset(dio, 0, offsetof(struct dio, pages));
1116 if (dio->flags & DIO_LOCKING) {
1118 struct address_space *mapping =
1119 iocb->ki_filp->f_mapping;
1121 /* will be released by direct_io_worker */
1122 mutex_lock(&inode->i_mutex);
1124 retval = filemap_write_and_wait_range(mapping, offset,
1127 mutex_unlock(&inode->i_mutex);
1128 kmem_cache_free(dio_cache, dio);
1135 * Will be decremented at I/O completion time.
1137 atomic_inc(&inode->i_dio_count);
1140 * For file extending writes updating i_size before data
1141 * writeouts complete can expose uninitialized blocks. So
1142 * even for AIO, we need to wait for i/o to complete before
1143 * returning in this case.
1145 dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
1146 (end > i_size_read(inode)));
1152 sdio.blkbits = blkbits;
1153 sdio.blkfactor = i_blkbits - blkbits;
1154 sdio.block_in_file = offset >> blkbits;
1156 sdio.get_block = get_block;
1157 dio->end_io = end_io;
1158 sdio.submit_io = submit_io;
1159 sdio.final_block_in_bio = -1;
1160 sdio.next_block_for_io = -1;
1163 dio->i_size = i_size_read(inode);
1165 spin_lock_init(&dio->bio_lock);
1169 * In case of non-aligned buffers, we may need 2 more
1170 * pages since we need to zero out first and last block.
1172 if (unlikely(sdio.blkfactor))
1173 sdio.pages_in_io = 2;
1175 for (seg = 0; seg < nr_segs; seg++) {
1176 user_addr = (unsigned long)iov[seg].iov_base;
1178 ((user_addr + iov[seg].iov_len + PAGE_SIZE-1) /
1179 PAGE_SIZE - user_addr / PAGE_SIZE);
1182 blk_start_plug(&plug);
1184 for (seg = 0; seg < nr_segs; seg++) {
1185 user_addr = (unsigned long)iov[seg].iov_base;
1186 sdio.size += bytes = iov[seg].iov_len;
1188 /* Index into the first page of the first block */
1189 sdio.first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
1190 sdio.final_block_in_request = sdio.block_in_file +
1192 /* Page fetching state */
1197 sdio.total_pages = 0;
1198 if (user_addr & (PAGE_SIZE-1)) {
1200 bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
1202 sdio.total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
1203 sdio.curr_user_address = user_addr;
1205 retval = do_direct_IO(dio, &sdio, &map_bh);
1207 dio->result += iov[seg].iov_len -
1208 ((sdio.final_block_in_request - sdio.block_in_file) <<
1212 dio_cleanup(dio, &sdio);
1215 } /* end iovec loop */
1217 if (retval == -ENOTBLK) {
1219 * The remaining part of the request will be
1220 * be handled by buffered I/O when we return
1225 * There may be some unwritten disk at the end of a part-written
1226 * fs-block-sized block. Go zero that now.
1228 dio_zero_block(dio, &sdio, 1, &map_bh);
1230 if (sdio.cur_page) {
1233 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1236 page_cache_release(sdio.cur_page);
1237 sdio.cur_page = NULL;
1240 dio_bio_submit(dio, &sdio);
1242 blk_finish_plug(&plug);
1245 * It is possible that, we return short IO due to end of file.
1246 * In that case, we need to release all the pages we got hold on.
1248 dio_cleanup(dio, &sdio);
1251 * All block lookups have been performed. For READ requests
1252 * we can let i_mutex go now that its achieved its purpose
1253 * of protecting us from looking up uninitialized blocks.
1255 if (rw == READ && (dio->flags & DIO_LOCKING))
1256 mutex_unlock(&dio->inode->i_mutex);
1259 * The only time we want to leave bios in flight is when a successful
1260 * partial aio read or full aio write have been setup. In that case
1261 * bio completion will call aio_complete. The only time it's safe to
1262 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1263 * This had *better* be the only place that raises -EIOCBQUEUED.
1265 BUG_ON(retval == -EIOCBQUEUED);
1266 if (dio->is_async && retval == 0 && dio->result &&
1267 ((rw == READ) || (dio->result == sdio.size)))
1268 retval = -EIOCBQUEUED;
1270 if (retval != -EIOCBQUEUED)
1271 dio_await_completion(dio);
1273 if (drop_refcount(dio) == 0) {
1274 retval = dio_complete(dio, offset, retval, false);
1275 kmem_cache_free(dio_cache, dio);
1277 BUG_ON(retval != -EIOCBQUEUED);
1284 __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1285 struct block_device *bdev, const struct iovec *iov, loff_t offset,
1286 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1287 dio_submit_t submit_io, int flags)
1290 * The block device state is needed in the end to finally
1291 * submit everything. Since it's likely to be cache cold
1292 * prefetch it here as first thing to hide some of the
1295 * Attempt to prefetch the pieces we likely need later.
1297 prefetch(&bdev->bd_disk->part_tbl);
1298 prefetch(bdev->bd_queue);
1299 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1301 return do_blockdev_direct_IO(rw, iocb, inode, bdev, iov, offset,
1302 nr_segs, get_block, end_io,
1306 EXPORT_SYMBOL(__blockdev_direct_IO);
1308 static __init int dio_init(void)
1310 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1313 module_init(dio_init)