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 * This code generally works in units of "dio_blocks". A dio_block is
49 * somewhere between the hard sector size and the filesystem block size. it
50 * is determined on a per-invocation basis. When talking to the filesystem
51 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
52 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
53 * to bio_block quantities by shifting left by blkfactor.
55 * If blkfactor is zero then the user's request was aligned to the filesystem's
59 /* dio_state only used in the submission path */
62 struct bio *bio; /* bio under assembly */
63 unsigned blkbits; /* doesn't change */
64 unsigned blkfactor; /* When we're using an alignment which
65 is finer than the filesystem's soft
66 blocksize, this specifies how much
67 finer. blkfactor=2 means 1/4-block
68 alignment. Does not change */
69 unsigned start_zero_done; /* flag: sub-blocksize zeroing has
70 been performed at the start of a
72 int pages_in_io; /* approximate total IO pages */
73 sector_t block_in_file; /* Current offset into the underlying
74 file in dio_block units. */
75 unsigned blocks_available; /* At block_in_file. changes */
76 int reap_counter; /* rate limit reaping */
77 sector_t final_block_in_request;/* doesn't change */
78 int boundary; /* prev block is at a boundary */
79 get_block_t *get_block; /* block mapping function */
80 dio_submit_t *submit_io; /* IO submition function */
82 loff_t logical_offset_in_bio; /* current first logical block in bio */
83 sector_t final_block_in_bio; /* current final block in bio + 1 */
84 sector_t next_block_for_io; /* next block to be put under IO,
85 in dio_blocks units */
88 * Deferred addition of a page to the dio. These variables are
89 * private to dio_send_cur_page(), submit_page_section() and
92 struct page *cur_page; /* The page */
93 unsigned cur_page_offset; /* Offset into it, in bytes */
94 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
95 sector_t cur_page_block; /* Where it starts */
96 loff_t cur_page_fs_offset; /* Offset in file */
98 struct iov_iter *iter;
100 * Page queue. These variables belong to dio_refill_pages() and
103 unsigned head; /* next page to process */
104 unsigned tail; /* last valid page + 1 */
108 /* dio_state communicated between submission path and end_io */
110 int flags; /* doesn't change */
114 struct block_device *bio_bdev;
116 loff_t i_size; /* i_size when submitted */
117 dio_iodone_t *end_io; /* IO completion function */
119 void *private; /* copy from map_bh.b_private */
121 /* BIO completion state */
122 spinlock_t bio_lock; /* protects BIO fields below */
123 int page_errors; /* errno from get_user_pages() */
124 int is_async; /* is IO async ? */
125 bool defer_completion; /* defer AIO completion to workqueue? */
126 bool should_dirty; /* if pages should be dirtied */
127 int io_error; /* IO error in completion path */
128 unsigned long refcount; /* direct_io_worker() and bios */
129 struct bio *bio_list; /* singly linked via bi_private */
130 struct task_struct *waiter; /* waiting task (NULL if none) */
132 /* AIO related stuff */
133 struct kiocb *iocb; /* kiocb */
134 ssize_t result; /* IO result */
137 * pages[] (and any fields placed after it) are not zeroed out at
138 * allocation time. Don't add new fields after pages[] unless you
139 * wish that they not be zeroed.
142 struct page *pages[DIO_PAGES]; /* page buffer */
143 struct work_struct complete_work;/* deferred AIO completion */
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)
164 ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES,
167 if (ret < 0 && sdio->blocks_available && (dio->op == REQ_OP_WRITE)) {
168 struct page *page = ZERO_PAGE(0);
170 * A memory fault, but the filesystem has some outstanding
171 * mapped blocks. We need to use those blocks up to avoid
172 * leaking stale data in the file.
174 if (dio->page_errors == 0)
175 dio->page_errors = ret;
177 dio->pages[0] = page;
181 sdio->to = PAGE_SIZE;
186 iov_iter_advance(sdio->iter, ret);
189 sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
190 sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1;
197 * Get another userspace page. Returns an ERR_PTR on error. Pages are
198 * buffered inside the dio so that we can call get_user_pages() against a
199 * decent number of pages, less frequently. To provide nicer use of the
202 static inline struct page *dio_get_page(struct dio *dio,
203 struct dio_submit *sdio)
205 if (dio_pages_present(sdio) == 0) {
208 ret = dio_refill_pages(dio, sdio);
211 BUG_ON(dio_pages_present(sdio) == 0);
213 return dio->pages[sdio->head];
217 * dio_complete() - called when all DIO BIO I/O has been completed
218 * @offset: the byte offset in the file of the completed operation
220 * This drops i_dio_count, lets interested parties know that a DIO operation
221 * has completed, and calculates the resulting return code for the operation.
223 * It lets the filesystem know if it registered an interest earlier via
224 * get_block. Pass the private field of the map buffer_head so that
225 * filesystems can use it to hold additional state between get_block calls and
228 static ssize_t dio_complete(struct dio *dio, ssize_t ret, bool is_async)
230 loff_t offset = dio->iocb->ki_pos;
231 ssize_t transferred = 0;
234 * AIO submission can race with bio completion to get here while
235 * expecting to have the last io completed by bio completion.
236 * In that case -EIOCBQUEUED is in fact not an error we want
237 * to preserve through this call.
239 if (ret == -EIOCBQUEUED)
243 transferred = dio->result;
245 /* Check for short read case */
246 if ((dio->op == REQ_OP_READ) &&
247 ((offset + transferred) > dio->i_size))
248 transferred = dio->i_size - offset;
249 /* ignore EFAULT if some IO has been done */
250 if (unlikely(ret == -EFAULT) && transferred)
255 ret = dio->page_errors;
265 err = dio->end_io(dio->iocb, offset, ret, dio->private);
270 if (!(dio->flags & DIO_SKIP_DIO_COUNT))
271 inode_dio_end(dio->inode);
275 * generic_write_sync expects ki_pos to have been updated
276 * already, but the submission path only does this for
279 dio->iocb->ki_pos += transferred;
281 if (dio->op == REQ_OP_WRITE)
282 ret = generic_write_sync(dio->iocb, transferred);
283 dio->iocb->ki_complete(dio->iocb, ret, 0);
286 kmem_cache_free(dio_cache, dio);
290 static void dio_aio_complete_work(struct work_struct *work)
292 struct dio *dio = container_of(work, struct dio, complete_work);
294 dio_complete(dio, 0, true);
297 static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio);
300 * Asynchronous IO callback.
302 static void dio_bio_end_aio(struct bio *bio)
304 struct dio *dio = bio->bi_private;
305 unsigned long remaining;
308 /* cleanup the bio */
309 dio_bio_complete(dio, bio);
311 spin_lock_irqsave(&dio->bio_lock, flags);
312 remaining = --dio->refcount;
313 if (remaining == 1 && dio->waiter)
314 wake_up_process(dio->waiter);
315 spin_unlock_irqrestore(&dio->bio_lock, flags);
317 if (remaining == 0) {
318 if (dio->result && dio->defer_completion) {
319 INIT_WORK(&dio->complete_work, dio_aio_complete_work);
320 queue_work(dio->inode->i_sb->s_dio_done_wq,
321 &dio->complete_work);
323 dio_complete(dio, 0, true);
329 * The BIO completion handler simply queues the BIO up for the process-context
332 * During I/O bi_private points at the dio. After I/O, bi_private is used to
333 * implement a singly-linked list of completed BIOs, at dio->bio_list.
335 static void dio_bio_end_io(struct bio *bio)
337 struct dio *dio = bio->bi_private;
340 spin_lock_irqsave(&dio->bio_lock, flags);
341 bio->bi_private = dio->bio_list;
343 if (--dio->refcount == 1 && dio->waiter)
344 wake_up_process(dio->waiter);
345 spin_unlock_irqrestore(&dio->bio_lock, flags);
349 * dio_end_io - handle the end io action for the given bio
350 * @bio: The direct io bio thats being completed
352 * This is meant to be called by any filesystem that uses their own dio_submit_t
353 * so that the DIO specific endio actions are dealt with after the filesystem
354 * has done it's completion work.
356 void dio_end_io(struct bio *bio)
358 struct dio *dio = bio->bi_private;
361 dio_bio_end_aio(bio);
365 EXPORT_SYMBOL_GPL(dio_end_io);
368 dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
369 struct block_device *bdev,
370 sector_t first_sector, int nr_vecs)
375 * bio_alloc() is guaranteed to return a bio when called with
376 * __GFP_RECLAIM and we request a valid number of vectors.
378 bio = bio_alloc(GFP_KERNEL, nr_vecs);
381 bio->bi_iter.bi_sector = first_sector;
382 bio_set_op_attrs(bio, dio->op, dio->op_flags);
384 bio->bi_end_io = dio_bio_end_aio;
386 bio->bi_end_io = dio_bio_end_io;
388 bio->bi_write_hint = dio->iocb->ki_hint;
391 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
395 * In the AIO read case we speculatively dirty the pages before starting IO.
396 * During IO completion, any of these pages which happen to have been written
397 * back will be redirtied by bio_check_pages_dirty().
399 * bios hold a dio reference between submit_bio and ->end_io.
401 static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
403 struct bio *bio = sdio->bio;
406 bio->bi_private = dio;
408 spin_lock_irqsave(&dio->bio_lock, flags);
410 spin_unlock_irqrestore(&dio->bio_lock, flags);
412 if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty)
413 bio_set_pages_dirty(bio);
415 dio->bio_bdev = bio->bi_bdev;
417 if (sdio->submit_io) {
418 sdio->submit_io(bio, dio->inode, sdio->logical_offset_in_bio);
419 dio->bio_cookie = BLK_QC_T_NONE;
421 dio->bio_cookie = submit_bio(bio);
425 sdio->logical_offset_in_bio = 0;
429 * Release any resources in case of a failure
431 static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
433 while (sdio->head < sdio->tail)
434 put_page(dio->pages[sdio->head++]);
438 * Wait for the next BIO to complete. Remove it and return it. NULL is
439 * returned once all BIOs have been completed. This must only be called once
440 * all bios have been issued so that dio->refcount can only decrease. This
441 * requires that that the caller hold a reference on the dio.
443 static struct bio *dio_await_one(struct dio *dio)
446 struct bio *bio = NULL;
448 spin_lock_irqsave(&dio->bio_lock, flags);
451 * Wait as long as the list is empty and there are bios in flight. bio
452 * completion drops the count, maybe adds to the list, and wakes while
453 * holding the bio_lock so we don't need set_current_state()'s barrier
454 * and can call it after testing our condition.
456 while (dio->refcount > 1 && dio->bio_list == NULL) {
457 __set_current_state(TASK_UNINTERRUPTIBLE);
458 dio->waiter = current;
459 spin_unlock_irqrestore(&dio->bio_lock, flags);
460 if (!(dio->iocb->ki_flags & IOCB_HIPRI) ||
461 !blk_mq_poll(bdev_get_queue(dio->bio_bdev), dio->bio_cookie))
463 /* wake up sets us TASK_RUNNING */
464 spin_lock_irqsave(&dio->bio_lock, flags);
469 dio->bio_list = bio->bi_private;
471 spin_unlock_irqrestore(&dio->bio_lock, flags);
476 * Process one completed BIO. No locks are held.
478 static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio)
480 struct bio_vec *bvec;
482 blk_status_t err = bio->bi_status;
485 if (err == BLK_STS_AGAIN && (bio->bi_opf & REQ_NOWAIT))
486 dio->io_error = -EAGAIN;
488 dio->io_error = -EIO;
491 if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty) {
492 bio_check_pages_dirty(bio); /* transfers ownership */
494 bio_for_each_segment_all(bvec, bio, i) {
495 struct page *page = bvec->bv_page;
497 if (dio->op == REQ_OP_READ && !PageCompound(page) &&
499 set_page_dirty_lock(page);
508 * Wait on and process all in-flight BIOs. This must only be called once
509 * all bios have been issued so that the refcount can only decrease.
510 * This just waits for all bios to make it through dio_bio_complete. IO
511 * errors are propagated through dio->io_error and should be propagated via
514 static void dio_await_completion(struct dio *dio)
518 bio = dio_await_one(dio);
520 dio_bio_complete(dio, bio);
525 * A really large O_DIRECT read or write can generate a lot of BIOs. So
526 * to keep the memory consumption sane we periodically reap any completed BIOs
527 * during the BIO generation phase.
529 * This also helps to limit the peak amount of pinned userspace memory.
531 static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
535 if (sdio->reap_counter++ >= 64) {
536 while (dio->bio_list) {
541 spin_lock_irqsave(&dio->bio_lock, flags);
543 dio->bio_list = bio->bi_private;
544 spin_unlock_irqrestore(&dio->bio_lock, flags);
545 ret2 = blk_status_to_errno(dio_bio_complete(dio, bio));
549 sdio->reap_counter = 0;
555 * Create workqueue for deferred direct IO completions. We allocate the
556 * workqueue when it's first needed. This avoids creating workqueue for
557 * filesystems that don't need it and also allows us to create the workqueue
558 * late enough so the we can include s_id in the name of the workqueue.
560 int sb_init_dio_done_wq(struct super_block *sb)
562 struct workqueue_struct *old;
563 struct workqueue_struct *wq = alloc_workqueue("dio/%s",
569 * This has to be atomic as more DIOs can race to create the workqueue
571 old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
572 /* Someone created workqueue before us? Free ours... */
574 destroy_workqueue(wq);
578 static int dio_set_defer_completion(struct dio *dio)
580 struct super_block *sb = dio->inode->i_sb;
582 if (dio->defer_completion)
584 dio->defer_completion = true;
585 if (!sb->s_dio_done_wq)
586 return sb_init_dio_done_wq(sb);
591 * Call into the fs to map some more disk blocks. We record the current number
592 * of available blocks at sdio->blocks_available. These are in units of the
593 * fs blocksize, i_blocksize(inode).
595 * The fs is allowed to map lots of blocks at once. If it wants to do that,
596 * it uses the passed inode-relative block number as the file offset, as usual.
598 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
599 * has remaining to do. The fs should not map more than this number of blocks.
601 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
602 * indicate how much contiguous disk space has been made available at
605 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
606 * This isn't very efficient...
608 * In the case of filesystem holes: the fs may return an arbitrarily-large
609 * hole by returning an appropriate value in b_size and by clearing
610 * buffer_mapped(). However the direct-io code will only process holes one
611 * block at a time - it will repeatedly call get_block() as it walks the hole.
613 static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
614 struct buffer_head *map_bh)
617 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
618 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
619 unsigned long fs_count; /* Number of filesystem-sized blocks */
621 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
624 * If there was a memory error and we've overwritten all the
625 * mapped blocks then we can now return that memory error
627 ret = dio->page_errors;
629 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
630 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
631 fs_endblk = (sdio->final_block_in_request - 1) >>
633 fs_count = fs_endblk - fs_startblk + 1;
636 map_bh->b_size = fs_count << i_blkbits;
639 * For writes that could fill holes inside i_size on a
640 * DIO_SKIP_HOLES filesystem we forbid block creations: only
641 * overwrites are permitted. We will return early to the caller
642 * once we see an unmapped buffer head returned, and the caller
643 * will fall back to buffered I/O.
645 * Otherwise the decision is left to the get_blocks method,
646 * which may decide to handle it or also return an unmapped
649 create = dio->op == REQ_OP_WRITE;
650 if (dio->flags & DIO_SKIP_HOLES) {
651 if (fs_startblk <= ((i_size_read(dio->inode) - 1) >>
656 ret = (*sdio->get_block)(dio->inode, fs_startblk,
659 /* Store for completion */
660 dio->private = map_bh->b_private;
662 if (ret == 0 && buffer_defer_completion(map_bh))
663 ret = dio_set_defer_completion(dio);
669 * There is no bio. Make one now.
671 static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
672 sector_t start_sector, struct buffer_head *map_bh)
677 ret = dio_bio_reap(dio, sdio);
680 sector = start_sector << (sdio->blkbits - 9);
681 nr_pages = min(sdio->pages_in_io, BIO_MAX_PAGES);
682 BUG_ON(nr_pages <= 0);
683 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
690 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
691 * that was successful then update final_block_in_bio and take a ref against
692 * the just-added page.
694 * Return zero on success. Non-zero means the caller needs to start a new BIO.
696 static inline int dio_bio_add_page(struct dio_submit *sdio)
700 ret = bio_add_page(sdio->bio, sdio->cur_page,
701 sdio->cur_page_len, sdio->cur_page_offset);
702 if (ret == sdio->cur_page_len) {
704 * Decrement count only, if we are done with this page
706 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
708 get_page(sdio->cur_page);
709 sdio->final_block_in_bio = sdio->cur_page_block +
710 (sdio->cur_page_len >> sdio->blkbits);
719 * Put cur_page under IO. The section of cur_page which is described by
720 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
721 * starts on-disk at cur_page_block.
723 * We take a ref against the page here (on behalf of its presence in the bio).
725 * The caller of this function is responsible for removing cur_page from the
726 * dio, and for dropping the refcount which came from that presence.
728 static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
729 struct buffer_head *map_bh)
734 loff_t cur_offset = sdio->cur_page_fs_offset;
735 loff_t bio_next_offset = sdio->logical_offset_in_bio +
736 sdio->bio->bi_iter.bi_size;
739 * See whether this new request is contiguous with the old.
741 * Btrfs cannot handle having logically non-contiguous requests
742 * submitted. For example if you have
744 * Logical: [0-4095][HOLE][8192-12287]
745 * Physical: [0-4095] [4096-8191]
747 * We cannot submit those pages together as one BIO. So if our
748 * current logical offset in the file does not equal what would
749 * be the next logical offset in the bio, submit the bio we
752 if (sdio->final_block_in_bio != sdio->cur_page_block ||
753 cur_offset != bio_next_offset)
754 dio_bio_submit(dio, sdio);
757 if (sdio->bio == NULL) {
758 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
763 if (dio_bio_add_page(sdio) != 0) {
764 dio_bio_submit(dio, sdio);
765 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
767 ret = dio_bio_add_page(sdio);
776 * An autonomous function to put a chunk of a page under deferred IO.
778 * The caller doesn't actually know (or care) whether this piece of page is in
779 * a BIO, or is under IO or whatever. We just take care of all possible
780 * situations here. The separation between the logic of do_direct_IO() and
781 * that of submit_page_section() is important for clarity. Please don't break.
783 * The chunk of page starts on-disk at blocknr.
785 * We perform deferred IO, by recording the last-submitted page inside our
786 * private part of the dio structure. If possible, we just expand the IO
787 * across that page here.
789 * If that doesn't work out then we put the old page into the bio and add this
790 * page to the dio instead.
793 submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
794 unsigned offset, unsigned len, sector_t blocknr,
795 struct buffer_head *map_bh)
799 if (dio->op == REQ_OP_WRITE) {
801 * Read accounting is performed in submit_bio()
803 task_io_account_write(len);
807 * Can we just grow the current page's presence in the dio?
809 if (sdio->cur_page == page &&
810 sdio->cur_page_offset + sdio->cur_page_len == offset &&
811 sdio->cur_page_block +
812 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
813 sdio->cur_page_len += len;
818 * If there's a deferred page already there then send it.
820 if (sdio->cur_page) {
821 ret = dio_send_cur_page(dio, sdio, map_bh);
822 put_page(sdio->cur_page);
823 sdio->cur_page = NULL;
828 get_page(page); /* It is in dio */
829 sdio->cur_page = page;
830 sdio->cur_page_offset = offset;
831 sdio->cur_page_len = len;
832 sdio->cur_page_block = blocknr;
833 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
836 * If sdio->boundary then we want to schedule the IO now to
837 * avoid metadata seeks.
839 if (sdio->boundary) {
840 ret = dio_send_cur_page(dio, sdio, map_bh);
841 dio_bio_submit(dio, sdio);
842 put_page(sdio->cur_page);
843 sdio->cur_page = NULL;
849 * If we are not writing the entire block and get_block() allocated
850 * the block for us, we need to fill-in the unused portion of the
851 * block with zeros. This happens only if user-buffer, fileoffset or
852 * io length is not filesystem block-size multiple.
854 * `end' is zero if we're doing the start of the IO, 1 at the end of the
857 static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
858 int end, struct buffer_head *map_bh)
860 unsigned dio_blocks_per_fs_block;
861 unsigned this_chunk_blocks; /* In dio_blocks */
862 unsigned this_chunk_bytes;
865 sdio->start_zero_done = 1;
866 if (!sdio->blkfactor || !buffer_new(map_bh))
869 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
870 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
872 if (!this_chunk_blocks)
876 * We need to zero out part of an fs block. It is either at the
877 * beginning or the end of the fs block.
880 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
882 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
885 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
886 sdio->next_block_for_io, map_bh))
889 sdio->next_block_for_io += this_chunk_blocks;
893 * Walk the user pages, and the file, mapping blocks to disk and generating
894 * a sequence of (page,offset,len,block) mappings. These mappings are injected
895 * into submit_page_section(), which takes care of the next stage of submission
897 * Direct IO against a blockdev is different from a file. Because we can
898 * happily perform page-sized but 512-byte aligned IOs. It is important that
899 * blockdev IO be able to have fine alignment and large sizes.
901 * So what we do is to permit the ->get_block function to populate bh.b_size
902 * with the size of IO which is permitted at this offset and this i_blkbits.
904 * For best results, the blockdev should be set up with 512-byte i_blkbits and
905 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
906 * fine alignment but still allows this function to work in PAGE_SIZE units.
908 static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
909 struct buffer_head *map_bh)
911 const unsigned blkbits = sdio->blkbits;
912 const unsigned i_blkbits = blkbits + sdio->blkfactor;
915 while (sdio->block_in_file < sdio->final_block_in_request) {
919 page = dio_get_page(dio, sdio);
924 from = sdio->head ? 0 : sdio->from;
925 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
929 unsigned this_chunk_bytes; /* # of bytes mapped */
930 unsigned this_chunk_blocks; /* # of blocks */
933 if (sdio->blocks_available == 0) {
935 * Need to go and map some more disk
937 unsigned long blkmask;
938 unsigned long dio_remainder;
940 ret = get_more_blocks(dio, sdio, map_bh);
945 if (!buffer_mapped(map_bh))
948 sdio->blocks_available =
949 map_bh->b_size >> blkbits;
950 sdio->next_block_for_io =
951 map_bh->b_blocknr << sdio->blkfactor;
952 if (buffer_new(map_bh)) {
956 map_bh->b_size >> i_blkbits);
959 if (!sdio->blkfactor)
962 blkmask = (1 << sdio->blkfactor) - 1;
963 dio_remainder = (sdio->block_in_file & blkmask);
966 * If we are at the start of IO and that IO
967 * starts partway into a fs-block,
968 * dio_remainder will be non-zero. If the IO
969 * is a read then we can simply advance the IO
970 * cursor to the first block which is to be
971 * read. But if the IO is a write and the
972 * block was newly allocated we cannot do that;
973 * the start of the fs block must be zeroed out
976 if (!buffer_new(map_bh))
977 sdio->next_block_for_io += dio_remainder;
978 sdio->blocks_available -= dio_remainder;
982 if (!buffer_mapped(map_bh)) {
983 loff_t i_size_aligned;
985 /* AKPM: eargh, -ENOTBLK is a hack */
986 if (dio->op == REQ_OP_WRITE) {
992 * Be sure to account for a partial block as the
993 * last block in the file
995 i_size_aligned = ALIGN(i_size_read(dio->inode),
997 if (sdio->block_in_file >=
998 i_size_aligned >> blkbits) {
1003 zero_user(page, from, 1 << blkbits);
1004 sdio->block_in_file++;
1005 from += 1 << blkbits;
1006 dio->result += 1 << blkbits;
1011 * If we're performing IO which has an alignment which
1012 * is finer than the underlying fs, go check to see if
1013 * we must zero out the start of this block.
1015 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
1016 dio_zero_block(dio, sdio, 0, map_bh);
1019 * Work out, in this_chunk_blocks, how much disk we
1020 * can add to this page
1022 this_chunk_blocks = sdio->blocks_available;
1023 u = (to - from) >> blkbits;
1024 if (this_chunk_blocks > u)
1025 this_chunk_blocks = u;
1026 u = sdio->final_block_in_request - sdio->block_in_file;
1027 if (this_chunk_blocks > u)
1028 this_chunk_blocks = u;
1029 this_chunk_bytes = this_chunk_blocks << blkbits;
1030 BUG_ON(this_chunk_bytes == 0);
1032 if (this_chunk_blocks == sdio->blocks_available)
1033 sdio->boundary = buffer_boundary(map_bh);
1034 ret = submit_page_section(dio, sdio, page,
1037 sdio->next_block_for_io,
1043 sdio->next_block_for_io += this_chunk_blocks;
1045 sdio->block_in_file += this_chunk_blocks;
1046 from += this_chunk_bytes;
1047 dio->result += this_chunk_bytes;
1048 sdio->blocks_available -= this_chunk_blocks;
1050 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1051 if (sdio->block_in_file == sdio->final_block_in_request)
1055 /* Drop the ref which was taken in get_user_pages() */
1062 static inline int drop_refcount(struct dio *dio)
1065 unsigned long flags;
1068 * Sync will always be dropping the final ref and completing the
1069 * operation. AIO can if it was a broken operation described above or
1070 * in fact if all the bios race to complete before we get here. In
1071 * that case dio_complete() translates the EIOCBQUEUED into the proper
1072 * return code that the caller will hand to ->complete().
1074 * This is managed by the bio_lock instead of being an atomic_t so that
1075 * completion paths can drop their ref and use the remaining count to
1076 * decide to wake the submission path atomically.
1078 spin_lock_irqsave(&dio->bio_lock, flags);
1079 ret2 = --dio->refcount;
1080 spin_unlock_irqrestore(&dio->bio_lock, flags);
1085 * This is a library function for use by filesystem drivers.
1087 * The locking rules are governed by the flags parameter:
1088 * - if the flags value contains DIO_LOCKING we use a fancy locking
1089 * scheme for dumb filesystems.
1090 * For writes this function is called under i_mutex and returns with
1091 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1092 * taken and dropped again before returning.
1093 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1094 * internal locking but rather rely on the filesystem to synchronize
1095 * direct I/O reads/writes versus each other and truncate.
1097 * To help with locking against truncate we incremented the i_dio_count
1098 * counter before starting direct I/O, and decrement it once we are done.
1099 * Truncate can wait for it to reach zero to provide exclusion. It is
1100 * expected that filesystem provide exclusion between new direct I/O
1101 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1102 * but other filesystems need to take care of this on their own.
1104 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1105 * is always inlined. Otherwise gcc is unable to split the structure into
1106 * individual fields and will generate much worse code. This is important
1107 * for the whole file.
1109 static inline ssize_t
1110 do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1111 struct block_device *bdev, struct iov_iter *iter,
1112 get_block_t get_block, dio_iodone_t end_io,
1113 dio_submit_t submit_io, int flags)
1115 unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits);
1116 unsigned blkbits = i_blkbits;
1117 unsigned blocksize_mask = (1 << blkbits) - 1;
1118 ssize_t retval = -EINVAL;
1119 size_t count = iov_iter_count(iter);
1120 loff_t offset = iocb->ki_pos;
1121 loff_t end = offset + count;
1123 struct dio_submit sdio = { 0, };
1124 struct buffer_head map_bh = { 0, };
1125 struct blk_plug plug;
1126 unsigned long align = offset | iov_iter_alignment(iter);
1129 * Avoid references to bdev if not absolutely needed to give
1130 * the early prefetch in the caller enough time.
1133 if (align & blocksize_mask) {
1135 blkbits = blksize_bits(bdev_logical_block_size(bdev));
1136 blocksize_mask = (1 << blkbits) - 1;
1137 if (align & blocksize_mask)
1141 /* watch out for a 0 len io from a tricksy fs */
1142 if (iov_iter_rw(iter) == READ && !iov_iter_count(iter))
1145 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
1150 * Believe it or not, zeroing out the page array caused a .5%
1151 * performance regression in a database benchmark. So, we take
1152 * care to only zero out what's needed.
1154 memset(dio, 0, offsetof(struct dio, pages));
1157 if (dio->flags & DIO_LOCKING) {
1158 if (iov_iter_rw(iter) == READ) {
1159 struct address_space *mapping =
1160 iocb->ki_filp->f_mapping;
1162 /* will be released by direct_io_worker */
1165 retval = filemap_write_and_wait_range(mapping, offset,
1168 inode_unlock(inode);
1169 kmem_cache_free(dio_cache, dio);
1175 /* Once we sampled i_size check for reads beyond EOF */
1176 dio->i_size = i_size_read(inode);
1177 if (iov_iter_rw(iter) == READ && offset >= dio->i_size) {
1178 if (dio->flags & DIO_LOCKING)
1179 inode_unlock(inode);
1180 kmem_cache_free(dio_cache, dio);
1186 * For file extending writes updating i_size before data writeouts
1187 * complete can expose uninitialized blocks in dumb filesystems.
1188 * In that case we need to wait for I/O completion even if asked
1189 * for an asynchronous write.
1191 if (is_sync_kiocb(iocb))
1192 dio->is_async = false;
1193 else if (!(dio->flags & DIO_ASYNC_EXTEND) &&
1194 iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
1195 dio->is_async = false;
1197 dio->is_async = true;
1200 if (iov_iter_rw(iter) == WRITE) {
1201 dio->op = REQ_OP_WRITE;
1202 dio->op_flags = REQ_SYNC | REQ_IDLE;
1203 if (iocb->ki_flags & IOCB_NOWAIT)
1204 dio->op_flags |= REQ_NOWAIT;
1206 dio->op = REQ_OP_READ;
1210 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1211 * so that we can call ->fsync.
1213 if (dio->is_async && iov_iter_rw(iter) == WRITE &&
1214 ((iocb->ki_filp->f_flags & O_DSYNC) ||
1215 IS_SYNC(iocb->ki_filp->f_mapping->host))) {
1216 retval = dio_set_defer_completion(dio);
1219 * We grab i_mutex only for reads so we don't have
1220 * to release it here
1222 kmem_cache_free(dio_cache, dio);
1228 * Will be decremented at I/O completion time.
1230 if (!(dio->flags & DIO_SKIP_DIO_COUNT))
1231 inode_dio_begin(inode);
1234 sdio.blkbits = blkbits;
1235 sdio.blkfactor = i_blkbits - blkbits;
1236 sdio.block_in_file = offset >> blkbits;
1238 sdio.get_block = get_block;
1239 dio->end_io = end_io;
1240 sdio.submit_io = submit_io;
1241 sdio.final_block_in_bio = -1;
1242 sdio.next_block_for_io = -1;
1246 spin_lock_init(&dio->bio_lock);
1249 dio->should_dirty = (iter->type == ITER_IOVEC);
1251 sdio.final_block_in_request =
1252 (offset + iov_iter_count(iter)) >> blkbits;
1255 * In case of non-aligned buffers, we may need 2 more
1256 * pages since we need to zero out first and last block.
1258 if (unlikely(sdio.blkfactor))
1259 sdio.pages_in_io = 2;
1261 sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
1263 blk_start_plug(&plug);
1265 retval = do_direct_IO(dio, &sdio, &map_bh);
1267 dio_cleanup(dio, &sdio);
1269 if (retval == -ENOTBLK) {
1271 * The remaining part of the request will be
1272 * be handled by buffered I/O when we return
1277 * There may be some unwritten disk at the end of a part-written
1278 * fs-block-sized block. Go zero that now.
1280 dio_zero_block(dio, &sdio, 1, &map_bh);
1282 if (sdio.cur_page) {
1285 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1288 put_page(sdio.cur_page);
1289 sdio.cur_page = NULL;
1292 dio_bio_submit(dio, &sdio);
1294 blk_finish_plug(&plug);
1297 * It is possible that, we return short IO due to end of file.
1298 * In that case, we need to release all the pages we got hold on.
1300 dio_cleanup(dio, &sdio);
1303 * All block lookups have been performed. For READ requests
1304 * we can let i_mutex go now that its achieved its purpose
1305 * of protecting us from looking up uninitialized blocks.
1307 if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING))
1308 inode_unlock(dio->inode);
1311 * The only time we want to leave bios in flight is when a successful
1312 * partial aio read or full aio write have been setup. In that case
1313 * bio completion will call aio_complete. The only time it's safe to
1314 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1315 * This had *better* be the only place that raises -EIOCBQUEUED.
1317 BUG_ON(retval == -EIOCBQUEUED);
1318 if (dio->is_async && retval == 0 && dio->result &&
1319 (iov_iter_rw(iter) == READ || dio->result == count))
1320 retval = -EIOCBQUEUED;
1322 dio_await_completion(dio);
1324 if (drop_refcount(dio) == 0) {
1325 retval = dio_complete(dio, retval, false);
1327 BUG_ON(retval != -EIOCBQUEUED);
1333 ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1334 struct block_device *bdev, struct iov_iter *iter,
1335 get_block_t get_block,
1336 dio_iodone_t end_io, dio_submit_t submit_io,
1340 * The block device state is needed in the end to finally
1341 * submit everything. Since it's likely to be cache cold
1342 * prefetch it here as first thing to hide some of the
1345 * Attempt to prefetch the pieces we likely need later.
1347 prefetch(&bdev->bd_disk->part_tbl);
1348 prefetch(bdev->bd_queue);
1349 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1351 return do_blockdev_direct_IO(iocb, inode, bdev, iter, get_block,
1352 end_io, submit_io, flags);
1355 EXPORT_SYMBOL(__blockdev_direct_IO);
1357 static __init int dio_init(void)
1359 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1362 module_init(dio_init)