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 */
113 struct block_device *bio_bdev;
115 loff_t i_size; /* i_size when submitted */
116 dio_iodone_t *end_io; /* IO completion function */
118 void *private; /* copy from map_bh.b_private */
120 /* BIO completion state */
121 spinlock_t bio_lock; /* protects BIO fields below */
122 int page_errors; /* errno from get_user_pages() */
123 int is_async; /* is IO async ? */
124 bool defer_completion; /* defer AIO completion to workqueue? */
125 bool should_dirty; /* if pages should be dirtied */
126 int io_error; /* IO error in completion path */
127 unsigned long refcount; /* direct_io_worker() and bios */
128 struct bio *bio_list; /* singly linked via bi_private */
129 struct task_struct *waiter; /* waiting task (NULL if none) */
131 /* AIO related stuff */
132 struct kiocb *iocb; /* kiocb */
133 ssize_t result; /* IO result */
136 * pages[] (and any fields placed after it) are not zeroed out at
137 * allocation time. Don't add new fields after pages[] unless you
138 * wish that they not be zeroed.
141 struct page *pages[DIO_PAGES]; /* page buffer */
142 struct work_struct complete_work;/* deferred AIO completion */
144 } ____cacheline_aligned_in_smp;
146 static struct kmem_cache *dio_cache __read_mostly;
149 * How many pages are in the queue?
151 static inline unsigned dio_pages_present(struct dio_submit *sdio)
153 return sdio->tail - sdio->head;
157 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
159 static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
163 ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES,
166 if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) {
167 struct page *page = ZERO_PAGE(0);
169 * A memory fault, but the filesystem has some outstanding
170 * mapped blocks. We need to use those blocks up to avoid
171 * leaking stale data in the file.
173 if (dio->page_errors == 0)
174 dio->page_errors = ret;
175 page_cache_get(page);
176 dio->pages[0] = page;
180 sdio->to = PAGE_SIZE;
185 iov_iter_advance(sdio->iter, ret);
188 sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
189 sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1;
196 * Get another userspace page. Returns an ERR_PTR on error. Pages are
197 * buffered inside the dio so that we can call get_user_pages() against a
198 * decent number of pages, less frequently. To provide nicer use of the
201 static inline struct page *dio_get_page(struct dio *dio,
202 struct dio_submit *sdio)
204 if (dio_pages_present(sdio) == 0) {
207 ret = dio_refill_pages(dio, sdio);
210 BUG_ON(dio_pages_present(sdio) == 0);
212 return dio->pages[sdio->head];
216 * dio_complete() - called when all DIO BIO I/O has been completed
217 * @offset: the byte offset in the file of the completed operation
219 * This drops i_dio_count, lets interested parties know that a DIO operation
220 * has completed, and calculates the resulting return code for the operation.
222 * It lets the filesystem know if it registered an interest earlier via
223 * get_block. Pass the private field of the map buffer_head so that
224 * filesystems can use it to hold additional state between get_block calls and
227 static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret,
230 ssize_t transferred = 0;
233 * AIO submission can race with bio completion to get here while
234 * expecting to have the last io completed by bio completion.
235 * In that case -EIOCBQUEUED is in fact not an error we want
236 * to preserve through this call.
238 if (ret == -EIOCBQUEUED)
242 transferred = dio->result;
244 /* Check for short read case */
245 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
246 transferred = dio->i_size - offset;
250 ret = dio->page_errors;
256 if (dio->end_io && dio->result)
257 dio->end_io(dio->iocb, offset, transferred, dio->private);
259 if (!(dio->flags & DIO_SKIP_DIO_COUNT))
260 inode_dio_end(dio->inode);
263 if (dio->rw & WRITE) {
266 err = generic_write_sync(dio->iocb->ki_filp, offset,
268 if (err < 0 && ret > 0)
272 dio->iocb->ki_complete(dio->iocb, ret, 0);
275 kmem_cache_free(dio_cache, dio);
279 static void dio_aio_complete_work(struct work_struct *work)
281 struct dio *dio = container_of(work, struct dio, complete_work);
283 dio_complete(dio, dio->iocb->ki_pos, 0, true);
286 static int dio_bio_complete(struct dio *dio, struct bio *bio);
289 * Asynchronous IO callback.
291 static void dio_bio_end_aio(struct bio *bio)
293 struct dio *dio = bio->bi_private;
294 unsigned long remaining;
297 /* cleanup the bio */
298 dio_bio_complete(dio, bio);
300 spin_lock_irqsave(&dio->bio_lock, flags);
301 remaining = --dio->refcount;
302 if (remaining == 1 && dio->waiter)
303 wake_up_process(dio->waiter);
304 spin_unlock_irqrestore(&dio->bio_lock, flags);
306 if (remaining == 0) {
307 if (dio->result && dio->defer_completion) {
308 INIT_WORK(&dio->complete_work, dio_aio_complete_work);
309 queue_work(dio->inode->i_sb->s_dio_done_wq,
310 &dio->complete_work);
312 dio_complete(dio, dio->iocb->ki_pos, 0, true);
318 * The BIO completion handler simply queues the BIO up for the process-context
321 * During I/O bi_private points at the dio. After I/O, bi_private is used to
322 * implement a singly-linked list of completed BIOs, at dio->bio_list.
324 static void dio_bio_end_io(struct bio *bio)
326 struct dio *dio = bio->bi_private;
329 spin_lock_irqsave(&dio->bio_lock, flags);
330 bio->bi_private = dio->bio_list;
332 if (--dio->refcount == 1 && dio->waiter)
333 wake_up_process(dio->waiter);
334 spin_unlock_irqrestore(&dio->bio_lock, flags);
338 * dio_end_io - handle the end io action for the given bio
339 * @bio: The direct io bio thats being completed
340 * @error: Error if there was one
342 * This is meant to be called by any filesystem that uses their own dio_submit_t
343 * so that the DIO specific endio actions are dealt with after the filesystem
344 * has done it's completion work.
346 void dio_end_io(struct bio *bio, int error)
348 struct dio *dio = bio->bi_private;
351 dio_bio_end_aio(bio);
355 EXPORT_SYMBOL_GPL(dio_end_io);
358 dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
359 struct block_device *bdev,
360 sector_t first_sector, int nr_vecs)
365 * bio_alloc() is guaranteed to return a bio when called with
366 * __GFP_RECLAIM and we request a valid number of vectors.
368 bio = bio_alloc(GFP_KERNEL, nr_vecs);
371 bio->bi_iter.bi_sector = first_sector;
373 bio->bi_end_io = dio_bio_end_aio;
375 bio->bi_end_io = dio_bio_end_io;
378 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
382 * In the AIO read case we speculatively dirty the pages before starting IO.
383 * During IO completion, any of these pages which happen to have been written
384 * back will be redirtied by bio_check_pages_dirty().
386 * bios hold a dio reference between submit_bio and ->end_io.
388 static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
390 struct bio *bio = sdio->bio;
393 bio->bi_private = dio;
395 spin_lock_irqsave(&dio->bio_lock, flags);
397 spin_unlock_irqrestore(&dio->bio_lock, flags);
399 if (dio->is_async && dio->rw == READ && dio->should_dirty)
400 bio_set_pages_dirty(bio);
402 dio->bio_bdev = bio->bi_bdev;
404 if (sdio->submit_io) {
405 sdio->submit_io(dio->rw, bio, dio->inode,
406 sdio->logical_offset_in_bio);
407 dio->bio_cookie = BLK_QC_T_NONE;
409 dio->bio_cookie = submit_bio(dio->rw, bio);
413 sdio->logical_offset_in_bio = 0;
417 * Release any resources in case of a failure
419 static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
421 while (sdio->head < sdio->tail)
422 page_cache_release(dio->pages[sdio->head++]);
426 * Wait for the next BIO to complete. Remove it and return it. NULL is
427 * returned once all BIOs have been completed. This must only be called once
428 * all bios have been issued so that dio->refcount can only decrease. This
429 * requires that that the caller hold a reference on the dio.
431 static struct bio *dio_await_one(struct dio *dio)
434 struct bio *bio = NULL;
436 spin_lock_irqsave(&dio->bio_lock, flags);
439 * Wait as long as the list is empty and there are bios in flight. bio
440 * completion drops the count, maybe adds to the list, and wakes while
441 * holding the bio_lock so we don't need set_current_state()'s barrier
442 * and can call it after testing our condition.
444 while (dio->refcount > 1 && dio->bio_list == NULL) {
445 __set_current_state(TASK_UNINTERRUPTIBLE);
446 dio->waiter = current;
447 spin_unlock_irqrestore(&dio->bio_lock, flags);
448 if (!blk_poll(bdev_get_queue(dio->bio_bdev), dio->bio_cookie))
450 /* wake up sets us TASK_RUNNING */
451 spin_lock_irqsave(&dio->bio_lock, flags);
456 dio->bio_list = bio->bi_private;
458 spin_unlock_irqrestore(&dio->bio_lock, flags);
463 * Process one completed BIO. No locks are held.
465 static int dio_bio_complete(struct dio *dio, struct bio *bio)
467 struct bio_vec *bvec;
472 dio->io_error = -EIO;
474 if (dio->is_async && dio->rw == READ && dio->should_dirty) {
476 bio_check_pages_dirty(bio); /* transfers ownership */
478 bio_for_each_segment_all(bvec, bio, i) {
479 struct page *page = bvec->bv_page;
481 if (dio->rw == READ && !PageCompound(page) &&
483 set_page_dirty_lock(page);
484 page_cache_release(page);
493 * Wait on and process all in-flight BIOs. This must only be called once
494 * all bios have been issued so that the refcount can only decrease.
495 * This just waits for all bios to make it through dio_bio_complete. IO
496 * errors are propagated through dio->io_error and should be propagated via
499 static void dio_await_completion(struct dio *dio)
503 bio = dio_await_one(dio);
505 dio_bio_complete(dio, bio);
510 * A really large O_DIRECT read or write can generate a lot of BIOs. So
511 * to keep the memory consumption sane we periodically reap any completed BIOs
512 * during the BIO generation phase.
514 * This also helps to limit the peak amount of pinned userspace memory.
516 static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
520 if (sdio->reap_counter++ >= 64) {
521 while (dio->bio_list) {
526 spin_lock_irqsave(&dio->bio_lock, flags);
528 dio->bio_list = bio->bi_private;
529 spin_unlock_irqrestore(&dio->bio_lock, flags);
530 ret2 = dio_bio_complete(dio, bio);
534 sdio->reap_counter = 0;
540 * Create workqueue for deferred direct IO completions. We allocate the
541 * workqueue when it's first needed. This avoids creating workqueue for
542 * filesystems that don't need it and also allows us to create the workqueue
543 * late enough so the we can include s_id in the name of the workqueue.
545 static int sb_init_dio_done_wq(struct super_block *sb)
547 struct workqueue_struct *old;
548 struct workqueue_struct *wq = alloc_workqueue("dio/%s",
554 * This has to be atomic as more DIOs can race to create the workqueue
556 old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
557 /* Someone created workqueue before us? Free ours... */
559 destroy_workqueue(wq);
563 static int dio_set_defer_completion(struct dio *dio)
565 struct super_block *sb = dio->inode->i_sb;
567 if (dio->defer_completion)
569 dio->defer_completion = true;
570 if (!sb->s_dio_done_wq)
571 return sb_init_dio_done_wq(sb);
576 * Call into the fs to map some more disk blocks. We record the current number
577 * of available blocks at sdio->blocks_available. These are in units of the
578 * fs blocksize, (1 << inode->i_blkbits).
580 * The fs is allowed to map lots of blocks at once. If it wants to do that,
581 * it uses the passed inode-relative block number as the file offset, as usual.
583 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
584 * has remaining to do. The fs should not map more than this number of blocks.
586 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
587 * indicate how much contiguous disk space has been made available at
590 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
591 * This isn't very efficient...
593 * In the case of filesystem holes: the fs may return an arbitrarily-large
594 * hole by returning an appropriate value in b_size and by clearing
595 * buffer_mapped(). However the direct-io code will only process holes one
596 * block at a time - it will repeatedly call get_block() as it walks the hole.
598 static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
599 struct buffer_head *map_bh)
602 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
603 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
604 unsigned long fs_count; /* Number of filesystem-sized blocks */
606 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
609 * If there was a memory error and we've overwritten all the
610 * mapped blocks then we can now return that memory error
612 ret = dio->page_errors;
614 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
615 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
616 fs_endblk = (sdio->final_block_in_request - 1) >>
618 fs_count = fs_endblk - fs_startblk + 1;
621 map_bh->b_size = fs_count << i_blkbits;
624 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we
625 * forbid block creations: only overwrites are permitted.
626 * We will return early to the caller once we see an
627 * unmapped buffer head returned, and the caller will fall
628 * back to buffered I/O.
630 * Otherwise the decision is left to the get_blocks method,
631 * which may decide to handle it or also return an unmapped
634 create = dio->rw & WRITE;
635 if (dio->flags & DIO_SKIP_HOLES) {
636 if (sdio->block_in_file < (i_size_read(dio->inode) >>
641 ret = (*sdio->get_block)(dio->inode, fs_startblk,
644 /* Store for completion */
645 dio->private = map_bh->b_private;
647 if (ret == 0 && buffer_defer_completion(map_bh))
648 ret = dio_set_defer_completion(dio);
654 * There is no bio. Make one now.
656 static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
657 sector_t start_sector, struct buffer_head *map_bh)
662 ret = dio_bio_reap(dio, sdio);
665 sector = start_sector << (sdio->blkbits - 9);
666 nr_pages = min(sdio->pages_in_io, BIO_MAX_PAGES);
667 BUG_ON(nr_pages <= 0);
668 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
675 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
676 * that was successful then update final_block_in_bio and take a ref against
677 * the just-added page.
679 * Return zero on success. Non-zero means the caller needs to start a new BIO.
681 static inline int dio_bio_add_page(struct dio_submit *sdio)
685 ret = bio_add_page(sdio->bio, sdio->cur_page,
686 sdio->cur_page_len, sdio->cur_page_offset);
687 if (ret == sdio->cur_page_len) {
689 * Decrement count only, if we are done with this page
691 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
693 page_cache_get(sdio->cur_page);
694 sdio->final_block_in_bio = sdio->cur_page_block +
695 (sdio->cur_page_len >> sdio->blkbits);
704 * Put cur_page under IO. The section of cur_page which is described by
705 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
706 * starts on-disk at cur_page_block.
708 * We take a ref against the page here (on behalf of its presence in the bio).
710 * The caller of this function is responsible for removing cur_page from the
711 * dio, and for dropping the refcount which came from that presence.
713 static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
714 struct buffer_head *map_bh)
719 loff_t cur_offset = sdio->cur_page_fs_offset;
720 loff_t bio_next_offset = sdio->logical_offset_in_bio +
721 sdio->bio->bi_iter.bi_size;
724 * See whether this new request is contiguous with the old.
726 * Btrfs cannot handle having logically non-contiguous requests
727 * submitted. For example if you have
729 * Logical: [0-4095][HOLE][8192-12287]
730 * Physical: [0-4095] [4096-8191]
732 * We cannot submit those pages together as one BIO. So if our
733 * current logical offset in the file does not equal what would
734 * be the next logical offset in the bio, submit the bio we
737 if (sdio->final_block_in_bio != sdio->cur_page_block ||
738 cur_offset != bio_next_offset)
739 dio_bio_submit(dio, sdio);
742 if (sdio->bio == NULL) {
743 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
748 if (dio_bio_add_page(sdio) != 0) {
749 dio_bio_submit(dio, sdio);
750 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
752 ret = dio_bio_add_page(sdio);
761 * An autonomous function to put a chunk of a page under deferred IO.
763 * The caller doesn't actually know (or care) whether this piece of page is in
764 * a BIO, or is under IO or whatever. We just take care of all possible
765 * situations here. The separation between the logic of do_direct_IO() and
766 * that of submit_page_section() is important for clarity. Please don't break.
768 * The chunk of page starts on-disk at blocknr.
770 * We perform deferred IO, by recording the last-submitted page inside our
771 * private part of the dio structure. If possible, we just expand the IO
772 * across that page here.
774 * If that doesn't work out then we put the old page into the bio and add this
775 * page to the dio instead.
778 submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
779 unsigned offset, unsigned len, sector_t blocknr,
780 struct buffer_head *map_bh)
784 if (dio->rw & WRITE) {
786 * Read accounting is performed in submit_bio()
788 task_io_account_write(len);
792 * Can we just grow the current page's presence in the dio?
794 if (sdio->cur_page == page &&
795 sdio->cur_page_offset + sdio->cur_page_len == offset &&
796 sdio->cur_page_block +
797 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
798 sdio->cur_page_len += len;
803 * If there's a deferred page already there then send it.
805 if (sdio->cur_page) {
806 ret = dio_send_cur_page(dio, sdio, map_bh);
807 page_cache_release(sdio->cur_page);
808 sdio->cur_page = NULL;
813 page_cache_get(page); /* It is in dio */
814 sdio->cur_page = page;
815 sdio->cur_page_offset = offset;
816 sdio->cur_page_len = len;
817 sdio->cur_page_block = blocknr;
818 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
821 * If sdio->boundary then we want to schedule the IO now to
822 * avoid metadata seeks.
824 if (sdio->boundary) {
825 ret = dio_send_cur_page(dio, sdio, map_bh);
826 dio_bio_submit(dio, sdio);
827 page_cache_release(sdio->cur_page);
828 sdio->cur_page = NULL;
834 * Clean any dirty buffers in the blockdev mapping which alias newly-created
835 * file blocks. Only called for S_ISREG files - blockdevs do not set
838 static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
843 nblocks = map_bh->b_size >> dio->inode->i_blkbits;
845 for (i = 0; i < nblocks; i++) {
846 unmap_underlying_metadata(map_bh->b_bdev,
847 map_bh->b_blocknr + i);
852 * If we are not writing the entire block and get_block() allocated
853 * the block for us, we need to fill-in the unused portion of the
854 * block with zeros. This happens only if user-buffer, fileoffset or
855 * io length is not filesystem block-size multiple.
857 * `end' is zero if we're doing the start of the IO, 1 at the end of the
860 static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
861 int end, struct buffer_head *map_bh)
863 unsigned dio_blocks_per_fs_block;
864 unsigned this_chunk_blocks; /* In dio_blocks */
865 unsigned this_chunk_bytes;
868 sdio->start_zero_done = 1;
869 if (!sdio->blkfactor || !buffer_new(map_bh))
872 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
873 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
875 if (!this_chunk_blocks)
879 * We need to zero out part of an fs block. It is either at the
880 * beginning or the end of the fs block.
883 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
885 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
888 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
889 sdio->next_block_for_io, map_bh))
892 sdio->next_block_for_io += this_chunk_blocks;
896 * Walk the user pages, and the file, mapping blocks to disk and generating
897 * a sequence of (page,offset,len,block) mappings. These mappings are injected
898 * into submit_page_section(), which takes care of the next stage of submission
900 * Direct IO against a blockdev is different from a file. Because we can
901 * happily perform page-sized but 512-byte aligned IOs. It is important that
902 * blockdev IO be able to have fine alignment and large sizes.
904 * So what we do is to permit the ->get_block function to populate bh.b_size
905 * with the size of IO which is permitted at this offset and this i_blkbits.
907 * For best results, the blockdev should be set up with 512-byte i_blkbits and
908 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
909 * fine alignment but still allows this function to work in PAGE_SIZE units.
911 static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
912 struct buffer_head *map_bh)
914 const unsigned blkbits = sdio->blkbits;
917 while (sdio->block_in_file < sdio->final_block_in_request) {
921 page = dio_get_page(dio, sdio);
926 from = sdio->head ? 0 : sdio->from;
927 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
931 unsigned this_chunk_bytes; /* # of bytes mapped */
932 unsigned this_chunk_blocks; /* # of blocks */
935 if (sdio->blocks_available == 0) {
937 * Need to go and map some more disk
939 unsigned long blkmask;
940 unsigned long dio_remainder;
942 ret = get_more_blocks(dio, sdio, map_bh);
944 page_cache_release(page);
947 if (!buffer_mapped(map_bh))
950 sdio->blocks_available =
951 map_bh->b_size >> sdio->blkbits;
952 sdio->next_block_for_io =
953 map_bh->b_blocknr << sdio->blkfactor;
954 if (buffer_new(map_bh))
955 clean_blockdev_aliases(dio, map_bh);
957 if (!sdio->blkfactor)
960 blkmask = (1 << sdio->blkfactor) - 1;
961 dio_remainder = (sdio->block_in_file & blkmask);
964 * If we are at the start of IO and that IO
965 * starts partway into a fs-block,
966 * dio_remainder will be non-zero. If the IO
967 * is a read then we can simply advance the IO
968 * cursor to the first block which is to be
969 * read. But if the IO is a write and the
970 * block was newly allocated we cannot do that;
971 * the start of the fs block must be zeroed out
974 if (!buffer_new(map_bh))
975 sdio->next_block_for_io += dio_remainder;
976 sdio->blocks_available -= dio_remainder;
980 if (!buffer_mapped(map_bh)) {
981 loff_t i_size_aligned;
983 /* AKPM: eargh, -ENOTBLK is a hack */
984 if (dio->rw & WRITE) {
985 page_cache_release(page);
990 * Be sure to account for a partial block as the
991 * last block in the file
993 i_size_aligned = ALIGN(i_size_read(dio->inode),
995 if (sdio->block_in_file >=
996 i_size_aligned >> blkbits) {
998 page_cache_release(page);
1001 zero_user(page, from, 1 << blkbits);
1002 sdio->block_in_file++;
1003 from += 1 << blkbits;
1004 dio->result += 1 << blkbits;
1009 * If we're performing IO which has an alignment which
1010 * is finer than the underlying fs, go check to see if
1011 * we must zero out the start of this block.
1013 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
1014 dio_zero_block(dio, sdio, 0, map_bh);
1017 * Work out, in this_chunk_blocks, how much disk we
1018 * can add to this page
1020 this_chunk_blocks = sdio->blocks_available;
1021 u = (to - from) >> blkbits;
1022 if (this_chunk_blocks > u)
1023 this_chunk_blocks = u;
1024 u = sdio->final_block_in_request - sdio->block_in_file;
1025 if (this_chunk_blocks > u)
1026 this_chunk_blocks = u;
1027 this_chunk_bytes = this_chunk_blocks << blkbits;
1028 BUG_ON(this_chunk_bytes == 0);
1030 if (this_chunk_blocks == sdio->blocks_available)
1031 sdio->boundary = buffer_boundary(map_bh);
1032 ret = submit_page_section(dio, sdio, page,
1035 sdio->next_block_for_io,
1038 page_cache_release(page);
1041 sdio->next_block_for_io += this_chunk_blocks;
1043 sdio->block_in_file += this_chunk_blocks;
1044 from += this_chunk_bytes;
1045 dio->result += this_chunk_bytes;
1046 sdio->blocks_available -= this_chunk_blocks;
1048 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1049 if (sdio->block_in_file == sdio->final_block_in_request)
1053 /* Drop the ref which was taken in get_user_pages() */
1054 page_cache_release(page);
1060 static inline int drop_refcount(struct dio *dio)
1063 unsigned long flags;
1066 * Sync will always be dropping the final ref and completing the
1067 * operation. AIO can if it was a broken operation described above or
1068 * in fact if all the bios race to complete before we get here. In
1069 * that case dio_complete() translates the EIOCBQUEUED into the proper
1070 * return code that the caller will hand to ->complete().
1072 * This is managed by the bio_lock instead of being an atomic_t so that
1073 * completion paths can drop their ref and use the remaining count to
1074 * decide to wake the submission path atomically.
1076 spin_lock_irqsave(&dio->bio_lock, flags);
1077 ret2 = --dio->refcount;
1078 spin_unlock_irqrestore(&dio->bio_lock, flags);
1083 * This is a library function for use by filesystem drivers.
1085 * The locking rules are governed by the flags parameter:
1086 * - if the flags value contains DIO_LOCKING we use a fancy locking
1087 * scheme for dumb filesystems.
1088 * For writes this function is called under i_mutex and returns with
1089 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1090 * taken and dropped again before returning.
1091 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1092 * internal locking but rather rely on the filesystem to synchronize
1093 * direct I/O reads/writes versus each other and truncate.
1095 * To help with locking against truncate we incremented the i_dio_count
1096 * counter before starting direct I/O, and decrement it once we are done.
1097 * Truncate can wait for it to reach zero to provide exclusion. It is
1098 * expected that filesystem provide exclusion between new direct I/O
1099 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1100 * but other filesystems need to take care of this on their own.
1102 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1103 * is always inlined. Otherwise gcc is unable to split the structure into
1104 * individual fields and will generate much worse code. This is important
1105 * for the whole file.
1107 static inline ssize_t
1108 do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1109 struct block_device *bdev, struct iov_iter *iter,
1110 loff_t offset, get_block_t get_block, dio_iodone_t end_io,
1111 dio_submit_t submit_io, int flags)
1113 unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits);
1114 unsigned blkbits = i_blkbits;
1115 unsigned blocksize_mask = (1 << blkbits) - 1;
1116 ssize_t retval = -EINVAL;
1117 size_t count = iov_iter_count(iter);
1118 loff_t end = offset + count;
1120 struct dio_submit sdio = { 0, };
1121 struct buffer_head map_bh = { 0, };
1122 struct blk_plug plug;
1123 unsigned long align = offset | iov_iter_alignment(iter);
1126 * Avoid references to bdev if not absolutely needed to give
1127 * the early prefetch in the caller enough time.
1130 if (align & blocksize_mask) {
1132 blkbits = blksize_bits(bdev_logical_block_size(bdev));
1133 blocksize_mask = (1 << blkbits) - 1;
1134 if (align & blocksize_mask)
1138 /* watch out for a 0 len io from a tricksy fs */
1139 if (iov_iter_rw(iter) == READ && !iov_iter_count(iter))
1142 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
1147 * Believe it or not, zeroing out the page array caused a .5%
1148 * performance regression in a database benchmark. So, we take
1149 * care to only zero out what's needed.
1151 memset(dio, 0, offsetof(struct dio, pages));
1154 if (dio->flags & DIO_LOCKING) {
1155 if (iov_iter_rw(iter) == READ) {
1156 struct address_space *mapping =
1157 iocb->ki_filp->f_mapping;
1159 /* will be released by direct_io_worker */
1160 mutex_lock(&inode->i_mutex);
1162 retval = filemap_write_and_wait_range(mapping, offset,
1165 mutex_unlock(&inode->i_mutex);
1166 kmem_cache_free(dio_cache, dio);
1172 /* Once we sampled i_size check for reads beyond EOF */
1173 dio->i_size = i_size_read(inode);
1174 if (iov_iter_rw(iter) == READ && offset >= dio->i_size) {
1175 if (dio->flags & DIO_LOCKING)
1176 mutex_unlock(&inode->i_mutex);
1177 kmem_cache_free(dio_cache, dio);
1183 * For file extending writes updating i_size before data writeouts
1184 * complete can expose uninitialized blocks in dumb filesystems.
1185 * In that case we need to wait for I/O completion even if asked
1186 * for an asynchronous write.
1188 if (is_sync_kiocb(iocb))
1189 dio->is_async = false;
1190 else if (!(dio->flags & DIO_ASYNC_EXTEND) &&
1191 iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
1192 dio->is_async = false;
1194 dio->is_async = true;
1197 dio->rw = iov_iter_rw(iter) == WRITE ? WRITE_ODIRECT : READ;
1200 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1201 * so that we can call ->fsync.
1203 if (dio->is_async && iov_iter_rw(iter) == WRITE &&
1204 ((iocb->ki_filp->f_flags & O_DSYNC) ||
1205 IS_SYNC(iocb->ki_filp->f_mapping->host))) {
1206 retval = dio_set_defer_completion(dio);
1209 * We grab i_mutex only for reads so we don't have
1210 * to release it here
1212 kmem_cache_free(dio_cache, dio);
1218 * Will be decremented at I/O completion time.
1220 if (!(dio->flags & DIO_SKIP_DIO_COUNT))
1221 inode_dio_begin(inode);
1224 sdio.blkbits = blkbits;
1225 sdio.blkfactor = i_blkbits - blkbits;
1226 sdio.block_in_file = offset >> blkbits;
1228 sdio.get_block = get_block;
1229 dio->end_io = end_io;
1230 sdio.submit_io = submit_io;
1231 sdio.final_block_in_bio = -1;
1232 sdio.next_block_for_io = -1;
1236 spin_lock_init(&dio->bio_lock);
1239 dio->should_dirty = (iter->type == ITER_IOVEC);
1241 sdio.final_block_in_request =
1242 (offset + iov_iter_count(iter)) >> blkbits;
1245 * In case of non-aligned buffers, we may need 2 more
1246 * pages since we need to zero out first and last block.
1248 if (unlikely(sdio.blkfactor))
1249 sdio.pages_in_io = 2;
1251 sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
1253 blk_start_plug(&plug);
1255 retval = do_direct_IO(dio, &sdio, &map_bh);
1257 dio_cleanup(dio, &sdio);
1259 if (retval == -ENOTBLK) {
1261 * The remaining part of the request will be
1262 * be handled by buffered I/O when we return
1267 * There may be some unwritten disk at the end of a part-written
1268 * fs-block-sized block. Go zero that now.
1270 dio_zero_block(dio, &sdio, 1, &map_bh);
1272 if (sdio.cur_page) {
1275 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1278 page_cache_release(sdio.cur_page);
1279 sdio.cur_page = NULL;
1282 dio_bio_submit(dio, &sdio);
1284 blk_finish_plug(&plug);
1287 * It is possible that, we return short IO due to end of file.
1288 * In that case, we need to release all the pages we got hold on.
1290 dio_cleanup(dio, &sdio);
1293 * All block lookups have been performed. For READ requests
1294 * we can let i_mutex go now that its achieved its purpose
1295 * of protecting us from looking up uninitialized blocks.
1297 if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING))
1298 mutex_unlock(&dio->inode->i_mutex);
1301 * The only time we want to leave bios in flight is when a successful
1302 * partial aio read or full aio write have been setup. In that case
1303 * bio completion will call aio_complete. The only time it's safe to
1304 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1305 * This had *better* be the only place that raises -EIOCBQUEUED.
1307 BUG_ON(retval == -EIOCBQUEUED);
1308 if (dio->is_async && retval == 0 && dio->result &&
1309 (iov_iter_rw(iter) == READ || dio->result == count))
1310 retval = -EIOCBQUEUED;
1312 dio_await_completion(dio);
1314 if (drop_refcount(dio) == 0) {
1315 retval = dio_complete(dio, offset, retval, false);
1317 BUG_ON(retval != -EIOCBQUEUED);
1323 ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1324 struct block_device *bdev, struct iov_iter *iter,
1325 loff_t offset, get_block_t get_block,
1326 dio_iodone_t end_io, dio_submit_t submit_io,
1330 * The block device state is needed in the end to finally
1331 * submit everything. Since it's likely to be cache cold
1332 * prefetch it here as first thing to hide some of the
1335 * Attempt to prefetch the pieces we likely need later.
1337 prefetch(&bdev->bd_disk->part_tbl);
1338 prefetch(bdev->bd_queue);
1339 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1341 return do_blockdev_direct_IO(iocb, inode, bdev, iter, offset, get_block,
1342 end_io, submit_io, flags);
1345 EXPORT_SYMBOL(__blockdev_direct_IO);
1347 static __init int dio_init(void)
1349 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1352 module_init(dio_init)