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 size_t size; /* total request size (doesn't change)*/
74 sector_t block_in_file; /* Current offset into the underlying
75 file in dio_block units. */
76 unsigned blocks_available; /* At block_in_file. changes */
77 int reap_counter; /* rate limit reaping */
78 sector_t final_block_in_request;/* doesn't change */
79 unsigned first_block_in_page; /* doesn't change, Used only once */
80 int boundary; /* prev block is at a boundary */
81 get_block_t *get_block; /* block mapping function */
82 dio_submit_t *submit_io; /* IO submition function */
84 loff_t logical_offset_in_bio; /* current first logical block in bio */
85 sector_t final_block_in_bio; /* current final block in bio + 1 */
86 sector_t next_block_for_io; /* next block to be put under IO,
87 in dio_blocks units */
90 * Deferred addition of a page to the dio. These variables are
91 * private to dio_send_cur_page(), submit_page_section() and
94 struct page *cur_page; /* The page */
95 unsigned cur_page_offset; /* Offset into it, in bytes */
96 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
97 sector_t cur_page_block; /* Where it starts */
98 loff_t cur_page_fs_offset; /* Offset in file */
101 * Page fetching state. These variables belong to dio_refill_pages().
103 int curr_page; /* changes */
104 int total_pages; /* doesn't change */
105 unsigned long curr_user_address;/* changes */
108 * Page queue. These variables belong to dio_refill_pages() and
111 unsigned head; /* next page to process */
112 unsigned tail; /* last valid page + 1 */
115 /* dio_state communicated between submission path and end_io */
117 int flags; /* doesn't change */
120 loff_t i_size; /* i_size when submitted */
121 dio_iodone_t *end_io; /* IO completion function */
123 void *private; /* copy from map_bh.b_private */
125 /* BIO completion state */
126 spinlock_t bio_lock; /* protects BIO fields below */
127 int page_errors; /* errno from get_user_pages() */
128 int is_async; /* is IO async ? */
129 int io_error; /* IO error in completion path */
130 unsigned long refcount; /* direct_io_worker() and bios */
131 struct bio *bio_list; /* singly linked via bi_private */
132 struct task_struct *waiter; /* waiting task (NULL if none) */
134 /* AIO related stuff */
135 struct kiocb *iocb; /* kiocb */
136 ssize_t result; /* IO result */
139 * pages[] (and any fields placed after it) are not zeroed out at
140 * allocation time. Don't add new fields after pages[] unless you
141 * wish that they not be zeroed.
143 struct page *pages[DIO_PAGES]; /* page buffer */
144 } ____cacheline_aligned_in_smp;
146 static struct kmem_cache *dio_cache __read_mostly;
148 static void __inode_dio_wait(struct inode *inode)
150 wait_queue_head_t *wq = bit_waitqueue(&inode->i_state, __I_DIO_WAKEUP);
151 DEFINE_WAIT_BIT(q, &inode->i_state, __I_DIO_WAKEUP);
154 prepare_to_wait(wq, &q.wait, TASK_UNINTERRUPTIBLE);
155 if (atomic_read(&inode->i_dio_count))
157 } while (atomic_read(&inode->i_dio_count));
158 finish_wait(wq, &q.wait);
162 * inode_dio_wait - wait for outstanding DIO requests to finish
163 * @inode: inode to wait for
165 * Waits for all pending direct I/O requests to finish so that we can
166 * proceed with a truncate or equivalent operation.
168 * Must be called under a lock that serializes taking new references
169 * to i_dio_count, usually by inode->i_mutex.
171 void inode_dio_wait(struct inode *inode)
173 if (atomic_read(&inode->i_dio_count))
174 __inode_dio_wait(inode);
176 EXPORT_SYMBOL(inode_dio_wait);
179 * inode_dio_done - signal finish of a direct I/O requests
180 * @inode: inode the direct I/O happens on
182 * This is called once we've finished processing a direct I/O request,
183 * and is used to wake up callers waiting for direct I/O to be quiesced.
185 void inode_dio_done(struct inode *inode)
187 if (atomic_dec_and_test(&inode->i_dio_count))
188 wake_up_bit(&inode->i_state, __I_DIO_WAKEUP);
190 EXPORT_SYMBOL(inode_dio_done);
193 * How many pages are in the queue?
195 static inline unsigned dio_pages_present(struct dio_submit *sdio)
197 return sdio->tail - sdio->head;
201 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
203 static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
208 nr_pages = min(sdio->total_pages - sdio->curr_page, DIO_PAGES);
209 ret = get_user_pages_fast(
210 sdio->curr_user_address, /* Where from? */
211 nr_pages, /* How many pages? */
212 dio->rw == READ, /* Write to memory? */
213 &dio->pages[0]); /* Put results here */
215 if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) {
216 struct page *page = ZERO_PAGE(0);
218 * A memory fault, but the filesystem has some outstanding
219 * mapped blocks. We need to use those blocks up to avoid
220 * leaking stale data in the file.
222 if (dio->page_errors == 0)
223 dio->page_errors = ret;
224 page_cache_get(page);
225 dio->pages[0] = page;
233 sdio->curr_user_address += ret * PAGE_SIZE;
234 sdio->curr_page += ret;
244 * Get another userspace page. Returns an ERR_PTR on error. Pages are
245 * buffered inside the dio so that we can call get_user_pages() against a
246 * decent number of pages, less frequently. To provide nicer use of the
249 static inline struct page *dio_get_page(struct dio *dio,
250 struct dio_submit *sdio)
252 if (dio_pages_present(sdio) == 0) {
255 ret = dio_refill_pages(dio, sdio);
258 BUG_ON(dio_pages_present(sdio) == 0);
260 return dio->pages[sdio->head++];
264 * dio_complete() - called when all DIO BIO I/O has been completed
265 * @offset: the byte offset in the file of the completed operation
267 * This releases locks as dictated by the locking type, lets interested parties
268 * know that a DIO operation has completed, and calculates the resulting return
269 * code for the operation.
271 * It lets the filesystem know if it registered an interest earlier via
272 * get_block. Pass the private field of the map buffer_head so that
273 * filesystems can use it to hold additional state between get_block calls and
276 static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret, bool is_async)
278 ssize_t transferred = 0;
281 * AIO submission can race with bio completion to get here while
282 * expecting to have the last io completed by bio completion.
283 * In that case -EIOCBQUEUED is in fact not an error we want
284 * to preserve through this call.
286 if (ret == -EIOCBQUEUED)
290 transferred = dio->result;
292 /* Check for short read case */
293 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
294 transferred = dio->i_size - offset;
298 ret = dio->page_errors;
304 if (dio->end_io && dio->result) {
305 dio->end_io(dio->iocb, offset, transferred,
306 dio->private, ret, is_async);
309 aio_complete(dio->iocb, ret, 0);
310 inode_dio_done(dio->inode);
316 static int dio_bio_complete(struct dio *dio, struct bio *bio);
318 * Asynchronous IO callback.
320 static void dio_bio_end_aio(struct bio *bio, int error)
322 struct dio *dio = bio->bi_private;
323 unsigned long remaining;
326 /* cleanup the bio */
327 dio_bio_complete(dio, bio);
329 spin_lock_irqsave(&dio->bio_lock, flags);
330 remaining = --dio->refcount;
331 if (remaining == 1 && dio->waiter)
332 wake_up_process(dio->waiter);
333 spin_unlock_irqrestore(&dio->bio_lock, flags);
335 if (remaining == 0) {
336 dio_complete(dio, dio->iocb->ki_pos, 0, true);
337 kmem_cache_free(dio_cache, dio);
342 * The BIO completion handler simply queues the BIO up for the process-context
345 * During I/O bi_private points at the dio. After I/O, bi_private is used to
346 * implement a singly-linked list of completed BIOs, at dio->bio_list.
348 static void dio_bio_end_io(struct bio *bio, int error)
350 struct dio *dio = bio->bi_private;
353 spin_lock_irqsave(&dio->bio_lock, flags);
354 bio->bi_private = dio->bio_list;
356 if (--dio->refcount == 1 && dio->waiter)
357 wake_up_process(dio->waiter);
358 spin_unlock_irqrestore(&dio->bio_lock, flags);
362 * dio_end_io - handle the end io action for the given bio
363 * @bio: The direct io bio thats being completed
364 * @error: Error if there was one
366 * This is meant to be called by any filesystem that uses their own dio_submit_t
367 * so that the DIO specific endio actions are dealt with after the filesystem
368 * has done it's completion work.
370 void dio_end_io(struct bio *bio, int error)
372 struct dio *dio = bio->bi_private;
375 dio_bio_end_aio(bio, error);
377 dio_bio_end_io(bio, error);
379 EXPORT_SYMBOL_GPL(dio_end_io);
382 dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
383 struct block_device *bdev,
384 sector_t first_sector, int nr_vecs)
389 * bio_alloc() is guaranteed to return a bio when called with
390 * __GFP_WAIT and we request a valid number of vectors.
392 bio = bio_alloc(GFP_KERNEL, nr_vecs);
395 bio->bi_sector = first_sector;
397 bio->bi_end_io = dio_bio_end_aio;
399 bio->bi_end_io = dio_bio_end_io;
402 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
406 * In the AIO read case we speculatively dirty the pages before starting IO.
407 * During IO completion, any of these pages which happen to have been written
408 * back will be redirtied by bio_check_pages_dirty().
410 * bios hold a dio reference between submit_bio and ->end_io.
412 static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
414 struct bio *bio = sdio->bio;
417 bio->bi_private = dio;
419 spin_lock_irqsave(&dio->bio_lock, flags);
421 spin_unlock_irqrestore(&dio->bio_lock, flags);
423 if (dio->is_async && dio->rw == READ)
424 bio_set_pages_dirty(bio);
427 sdio->submit_io(dio->rw, bio, dio->inode,
428 sdio->logical_offset_in_bio);
430 submit_bio(dio->rw, bio);
434 sdio->logical_offset_in_bio = 0;
438 * Release any resources in case of a failure
440 static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
442 while (dio_pages_present(sdio))
443 page_cache_release(dio_get_page(dio, sdio));
447 * Wait for the next BIO to complete. Remove it and return it. NULL is
448 * returned once all BIOs have been completed. This must only be called once
449 * all bios have been issued so that dio->refcount can only decrease. This
450 * requires that that the caller hold a reference on the dio.
452 static struct bio *dio_await_one(struct dio *dio)
455 struct bio *bio = NULL;
457 spin_lock_irqsave(&dio->bio_lock, flags);
460 * Wait as long as the list is empty and there are bios in flight. bio
461 * completion drops the count, maybe adds to the list, and wakes while
462 * holding the bio_lock so we don't need set_current_state()'s barrier
463 * and can call it after testing our condition.
465 while (dio->refcount > 1 && dio->bio_list == NULL) {
466 __set_current_state(TASK_UNINTERRUPTIBLE);
467 dio->waiter = current;
468 spin_unlock_irqrestore(&dio->bio_lock, flags);
470 /* wake up sets us TASK_RUNNING */
471 spin_lock_irqsave(&dio->bio_lock, flags);
476 dio->bio_list = bio->bi_private;
478 spin_unlock_irqrestore(&dio->bio_lock, flags);
483 * Process one completed BIO. No locks are held.
485 static int dio_bio_complete(struct dio *dio, struct bio *bio)
487 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
488 struct bio_vec *bvec = bio->bi_io_vec;
492 dio->io_error = -EIO;
494 if (dio->is_async && dio->rw == READ) {
495 bio_check_pages_dirty(bio); /* transfers ownership */
497 for (page_no = 0; page_no < bio->bi_vcnt; page_no++) {
498 struct page *page = bvec[page_no].bv_page;
500 if (dio->rw == READ && !PageCompound(page))
501 set_page_dirty_lock(page);
502 page_cache_release(page);
506 return uptodate ? 0 : -EIO;
510 * Wait on and process all in-flight BIOs. This must only be called once
511 * all bios have been issued so that the refcount can only decrease.
512 * This just waits for all bios to make it through dio_bio_complete. IO
513 * errors are propagated through dio->io_error and should be propagated via
516 static void dio_await_completion(struct dio *dio)
520 bio = dio_await_one(dio);
522 dio_bio_complete(dio, bio);
527 * A really large O_DIRECT read or write can generate a lot of BIOs. So
528 * to keep the memory consumption sane we periodically reap any completed BIOs
529 * during the BIO generation phase.
531 * This also helps to limit the peak amount of pinned userspace memory.
533 static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
537 if (sdio->reap_counter++ >= 64) {
538 while (dio->bio_list) {
543 spin_lock_irqsave(&dio->bio_lock, flags);
545 dio->bio_list = bio->bi_private;
546 spin_unlock_irqrestore(&dio->bio_lock, flags);
547 ret2 = dio_bio_complete(dio, bio);
551 sdio->reap_counter = 0;
557 * Call into the fs to map some more disk blocks. We record the current number
558 * of available blocks at sdio->blocks_available. These are in units of the
559 * fs blocksize, (1 << inode->i_blkbits).
561 * The fs is allowed to map lots of blocks at once. If it wants to do that,
562 * it uses the passed inode-relative block number as the file offset, as usual.
564 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
565 * has remaining to do. The fs should not map more than this number of blocks.
567 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
568 * indicate how much contiguous disk space has been made available at
571 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
572 * This isn't very efficient...
574 * In the case of filesystem holes: the fs may return an arbitrarily-large
575 * hole by returning an appropriate value in b_size and by clearing
576 * buffer_mapped(). However the direct-io code will only process holes one
577 * block at a time - it will repeatedly call get_block() as it walks the hole.
579 static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
580 struct buffer_head *map_bh)
583 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
584 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
585 unsigned long fs_count; /* Number of filesystem-sized blocks */
589 * If there was a memory error and we've overwritten all the
590 * mapped blocks then we can now return that memory error
592 ret = dio->page_errors;
594 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
595 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
596 fs_endblk = (sdio->final_block_in_request - 1) >>
598 fs_count = fs_endblk - fs_startblk + 1;
601 map_bh->b_size = fs_count << dio->inode->i_blkbits;
604 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we
605 * forbid block creations: only overwrites are permitted.
606 * We will return early to the caller once we see an
607 * unmapped buffer head returned, and the caller will fall
608 * back to buffered I/O.
610 * Otherwise the decision is left to the get_blocks method,
611 * which may decide to handle it or also return an unmapped
614 create = dio->rw & WRITE;
615 if (dio->flags & DIO_SKIP_HOLES) {
616 if (sdio->block_in_file < (i_size_read(dio->inode) >>
621 ret = (*sdio->get_block)(dio->inode, fs_startblk,
624 /* Store for completion */
625 dio->private = map_bh->b_private;
631 * There is no bio. Make one now.
633 static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
634 sector_t start_sector, struct buffer_head *map_bh)
639 ret = dio_bio_reap(dio, sdio);
642 sector = start_sector << (sdio->blkbits - 9);
643 nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(map_bh->b_bdev));
644 nr_pages = min(nr_pages, BIO_MAX_PAGES);
645 BUG_ON(nr_pages <= 0);
646 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
653 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
654 * that was successful then update final_block_in_bio and take a ref against
655 * the just-added page.
657 * Return zero on success. Non-zero means the caller needs to start a new BIO.
659 static inline int dio_bio_add_page(struct dio_submit *sdio)
663 ret = bio_add_page(sdio->bio, sdio->cur_page,
664 sdio->cur_page_len, sdio->cur_page_offset);
665 if (ret == sdio->cur_page_len) {
667 * Decrement count only, if we are done with this page
669 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
671 page_cache_get(sdio->cur_page);
672 sdio->final_block_in_bio = sdio->cur_page_block +
673 (sdio->cur_page_len >> sdio->blkbits);
682 * Put cur_page under IO. The section of cur_page which is described by
683 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
684 * starts on-disk at cur_page_block.
686 * We take a ref against the page here (on behalf of its presence in the bio).
688 * The caller of this function is responsible for removing cur_page from the
689 * dio, and for dropping the refcount which came from that presence.
691 static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
692 struct buffer_head *map_bh)
697 loff_t cur_offset = sdio->cur_page_fs_offset;
698 loff_t bio_next_offset = sdio->logical_offset_in_bio +
702 * See whether this new request is contiguous with the old.
704 * Btrfs cannot handle having logically non-contiguous requests
705 * submitted. For example if you have
707 * Logical: [0-4095][HOLE][8192-12287]
708 * Physical: [0-4095] [4096-8191]
710 * We cannot submit those pages together as one BIO. So if our
711 * current logical offset in the file does not equal what would
712 * be the next logical offset in the bio, submit the bio we
715 if (sdio->final_block_in_bio != sdio->cur_page_block ||
716 cur_offset != bio_next_offset)
717 dio_bio_submit(dio, sdio);
719 * Submit now if the underlying fs is about to perform a
722 else if (sdio->boundary)
723 dio_bio_submit(dio, sdio);
726 if (sdio->bio == NULL) {
727 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
732 if (dio_bio_add_page(sdio) != 0) {
733 dio_bio_submit(dio, sdio);
734 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
736 ret = dio_bio_add_page(sdio);
745 * An autonomous function to put a chunk of a page under deferred IO.
747 * The caller doesn't actually know (or care) whether this piece of page is in
748 * a BIO, or is under IO or whatever. We just take care of all possible
749 * situations here. The separation between the logic of do_direct_IO() and
750 * that of submit_page_section() is important for clarity. Please don't break.
752 * The chunk of page starts on-disk at blocknr.
754 * We perform deferred IO, by recording the last-submitted page inside our
755 * private part of the dio structure. If possible, we just expand the IO
756 * across that page here.
758 * If that doesn't work out then we put the old page into the bio and add this
759 * page to the dio instead.
762 submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
763 unsigned offset, unsigned len, sector_t blocknr,
764 struct buffer_head *map_bh)
768 if (dio->rw & WRITE) {
770 * Read accounting is performed in submit_bio()
772 task_io_account_write(len);
776 * Can we just grow the current page's presence in the dio?
778 if (sdio->cur_page == page &&
779 sdio->cur_page_offset + sdio->cur_page_len == offset &&
780 sdio->cur_page_block +
781 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
782 sdio->cur_page_len += len;
785 * If sdio->boundary then we want to schedule the IO now to
786 * avoid metadata seeks.
788 if (sdio->boundary) {
789 ret = dio_send_cur_page(dio, sdio, map_bh);
790 page_cache_release(sdio->cur_page);
791 sdio->cur_page = NULL;
797 * If there's a deferred page already there then send it.
799 if (sdio->cur_page) {
800 ret = dio_send_cur_page(dio, sdio, map_bh);
801 page_cache_release(sdio->cur_page);
802 sdio->cur_page = NULL;
807 page_cache_get(page); /* It is in dio */
808 sdio->cur_page = page;
809 sdio->cur_page_offset = offset;
810 sdio->cur_page_len = len;
811 sdio->cur_page_block = blocknr;
812 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
818 * Clean any dirty buffers in the blockdev mapping which alias newly-created
819 * file blocks. Only called for S_ISREG files - blockdevs do not set
822 static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
827 nblocks = map_bh->b_size >> dio->inode->i_blkbits;
829 for (i = 0; i < nblocks; i++) {
830 unmap_underlying_metadata(map_bh->b_bdev,
831 map_bh->b_blocknr + i);
836 * If we are not writing the entire block and get_block() allocated
837 * the block for us, we need to fill-in the unused portion of the
838 * block with zeros. This happens only if user-buffer, fileoffset or
839 * io length is not filesystem block-size multiple.
841 * `end' is zero if we're doing the start of the IO, 1 at the end of the
844 static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
845 int end, struct buffer_head *map_bh)
847 unsigned dio_blocks_per_fs_block;
848 unsigned this_chunk_blocks; /* In dio_blocks */
849 unsigned this_chunk_bytes;
852 sdio->start_zero_done = 1;
853 if (!sdio->blkfactor || !buffer_new(map_bh))
856 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
857 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
859 if (!this_chunk_blocks)
863 * We need to zero out part of an fs block. It is either at the
864 * beginning or the end of the fs block.
867 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
869 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
872 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
873 sdio->next_block_for_io, map_bh))
876 sdio->next_block_for_io += this_chunk_blocks;
880 * Walk the user pages, and the file, mapping blocks to disk and generating
881 * a sequence of (page,offset,len,block) mappings. These mappings are injected
882 * into submit_page_section(), which takes care of the next stage of submission
884 * Direct IO against a blockdev is different from a file. Because we can
885 * happily perform page-sized but 512-byte aligned IOs. It is important that
886 * blockdev IO be able to have fine alignment and large sizes.
888 * So what we do is to permit the ->get_block function to populate bh.b_size
889 * with the size of IO which is permitted at this offset and this i_blkbits.
891 * For best results, the blockdev should be set up with 512-byte i_blkbits and
892 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
893 * fine alignment but still allows this function to work in PAGE_SIZE units.
895 static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
896 struct buffer_head *map_bh)
898 const unsigned blkbits = sdio->blkbits;
899 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
901 unsigned block_in_page;
904 /* The I/O can start at any block offset within the first page */
905 block_in_page = sdio->first_block_in_page;
907 while (sdio->block_in_file < sdio->final_block_in_request) {
908 page = dio_get_page(dio, sdio);
914 while (block_in_page < blocks_per_page) {
915 unsigned offset_in_page = block_in_page << blkbits;
916 unsigned this_chunk_bytes; /* # of bytes mapped */
917 unsigned this_chunk_blocks; /* # of blocks */
920 if (sdio->blocks_available == 0) {
922 * Need to go and map some more disk
924 unsigned long blkmask;
925 unsigned long dio_remainder;
927 ret = get_more_blocks(dio, sdio, map_bh);
929 page_cache_release(page);
932 if (!buffer_mapped(map_bh))
935 sdio->blocks_available =
936 map_bh->b_size >> sdio->blkbits;
937 sdio->next_block_for_io =
938 map_bh->b_blocknr << sdio->blkfactor;
939 if (buffer_new(map_bh))
940 clean_blockdev_aliases(dio, map_bh);
942 if (!sdio->blkfactor)
945 blkmask = (1 << sdio->blkfactor) - 1;
946 dio_remainder = (sdio->block_in_file & blkmask);
949 * If we are at the start of IO and that IO
950 * starts partway into a fs-block,
951 * dio_remainder will be non-zero. If the IO
952 * is a read then we can simply advance the IO
953 * cursor to the first block which is to be
954 * read. But if the IO is a write and the
955 * block was newly allocated we cannot do that;
956 * the start of the fs block must be zeroed out
959 if (!buffer_new(map_bh))
960 sdio->next_block_for_io += dio_remainder;
961 sdio->blocks_available -= dio_remainder;
965 if (!buffer_mapped(map_bh)) {
966 loff_t i_size_aligned;
968 /* AKPM: eargh, -ENOTBLK is a hack */
969 if (dio->rw & WRITE) {
970 page_cache_release(page);
975 * Be sure to account for a partial block as the
976 * last block in the file
978 i_size_aligned = ALIGN(i_size_read(dio->inode),
980 if (sdio->block_in_file >=
981 i_size_aligned >> blkbits) {
983 page_cache_release(page);
986 zero_user(page, block_in_page << blkbits,
988 sdio->block_in_file++;
994 * If we're performing IO which has an alignment which
995 * is finer than the underlying fs, go check to see if
996 * we must zero out the start of this block.
998 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
999 dio_zero_block(dio, sdio, 0, map_bh);
1002 * Work out, in this_chunk_blocks, how much disk we
1003 * can add to this page
1005 this_chunk_blocks = sdio->blocks_available;
1006 u = (PAGE_SIZE - offset_in_page) >> blkbits;
1007 if (this_chunk_blocks > u)
1008 this_chunk_blocks = u;
1009 u = sdio->final_block_in_request - sdio->block_in_file;
1010 if (this_chunk_blocks > u)
1011 this_chunk_blocks = u;
1012 this_chunk_bytes = this_chunk_blocks << blkbits;
1013 BUG_ON(this_chunk_bytes == 0);
1015 sdio->boundary = buffer_boundary(map_bh);
1016 ret = submit_page_section(dio, sdio, page,
1019 sdio->next_block_for_io,
1022 page_cache_release(page);
1025 sdio->next_block_for_io += this_chunk_blocks;
1027 sdio->block_in_file += this_chunk_blocks;
1028 block_in_page += this_chunk_blocks;
1029 sdio->blocks_available -= this_chunk_blocks;
1031 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1032 if (sdio->block_in_file == sdio->final_block_in_request)
1036 /* Drop the ref which was taken in get_user_pages() */
1037 page_cache_release(page);
1044 static inline int drop_refcount(struct dio *dio)
1047 unsigned long flags;
1050 * Sync will always be dropping the final ref and completing the
1051 * operation. AIO can if it was a broken operation described above or
1052 * in fact if all the bios race to complete before we get here. In
1053 * that case dio_complete() translates the EIOCBQUEUED into the proper
1054 * return code that the caller will hand to aio_complete().
1056 * This is managed by the bio_lock instead of being an atomic_t so that
1057 * completion paths can drop their ref and use the remaining count to
1058 * decide to wake the submission path atomically.
1060 spin_lock_irqsave(&dio->bio_lock, flags);
1061 ret2 = --dio->refcount;
1062 spin_unlock_irqrestore(&dio->bio_lock, flags);
1067 * This is a library function for use by filesystem drivers.
1069 * The locking rules are governed by the flags parameter:
1070 * - if the flags value contains DIO_LOCKING we use a fancy locking
1071 * scheme for dumb filesystems.
1072 * For writes this function is called under i_mutex and returns with
1073 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1074 * taken and dropped again before returning.
1075 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1076 * internal locking but rather rely on the filesystem to synchronize
1077 * direct I/O reads/writes versus each other and truncate.
1079 * To help with locking against truncate we incremented the i_dio_count
1080 * counter before starting direct I/O, and decrement it once we are done.
1081 * Truncate can wait for it to reach zero to provide exclusion. It is
1082 * expected that filesystem provide exclusion between new direct I/O
1083 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1084 * but other filesystems need to take care of this on their own.
1086 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1087 * is always inlined. Otherwise gcc is unable to split the structure into
1088 * individual fields and will generate much worse code. This is important
1089 * for the whole file.
1091 static inline ssize_t
1092 do_blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1093 struct block_device *bdev, const struct iovec *iov, loff_t offset,
1094 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1095 dio_submit_t submit_io, int flags)
1100 unsigned blkbits = inode->i_blkbits;
1101 unsigned blocksize_mask = (1 << blkbits) - 1;
1102 ssize_t retval = -EINVAL;
1103 loff_t end = offset;
1105 struct dio_submit sdio = { 0, };
1106 unsigned long user_addr;
1108 struct buffer_head map_bh = { 0, };
1114 * Avoid references to bdev if not absolutely needed to give
1115 * the early prefetch in the caller enough time.
1118 if (offset & blocksize_mask) {
1120 blkbits = blksize_bits(bdev_logical_block_size(bdev));
1121 blocksize_mask = (1 << blkbits) - 1;
1122 if (offset & blocksize_mask)
1126 /* Check the memory alignment. Blocks cannot straddle pages */
1127 for (seg = 0; seg < nr_segs; seg++) {
1128 addr = (unsigned long)iov[seg].iov_base;
1129 size = iov[seg].iov_len;
1131 if (unlikely((addr & blocksize_mask) ||
1132 (size & blocksize_mask))) {
1134 blkbits = blksize_bits(
1135 bdev_logical_block_size(bdev));
1136 blocksize_mask = (1 << blkbits) - 1;
1137 if ((addr & blocksize_mask) || (size & blocksize_mask))
1142 /* watch out for a 0 len io from a tricksy fs */
1143 if (rw == READ && end == offset)
1146 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
1151 * Believe it or not, zeroing out the page array caused a .5%
1152 * performance regression in a database benchmark. So, we take
1153 * care to only zero out what's needed.
1155 memset(dio, 0, offsetof(struct dio, pages));
1158 if (dio->flags & DIO_LOCKING) {
1160 struct address_space *mapping =
1161 iocb->ki_filp->f_mapping;
1163 /* will be released by direct_io_worker */
1164 mutex_lock(&inode->i_mutex);
1166 retval = filemap_write_and_wait_range(mapping, offset,
1169 mutex_unlock(&inode->i_mutex);
1170 kmem_cache_free(dio_cache, dio);
1177 * Will be decremented at I/O completion time.
1179 atomic_inc(&inode->i_dio_count);
1182 * For file extending writes updating i_size before data
1183 * writeouts complete can expose uninitialized blocks. So
1184 * even for AIO, we need to wait for i/o to complete before
1185 * returning in this case.
1187 dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
1188 (end > i_size_read(inode)));
1194 sdio.blkbits = blkbits;
1195 sdio.blkfactor = inode->i_blkbits - blkbits;
1196 sdio.block_in_file = offset >> blkbits;
1198 sdio.get_block = get_block;
1199 dio->end_io = end_io;
1200 sdio.submit_io = submit_io;
1201 sdio.final_block_in_bio = -1;
1202 sdio.next_block_for_io = -1;
1205 dio->i_size = i_size_read(inode);
1207 spin_lock_init(&dio->bio_lock);
1211 * In case of non-aligned buffers, we may need 2 more
1212 * pages since we need to zero out first and last block.
1214 if (unlikely(sdio.blkfactor))
1215 sdio.pages_in_io = 2;
1217 for (seg = 0; seg < nr_segs; seg++) {
1218 user_addr = (unsigned long)iov[seg].iov_base;
1220 ((user_addr + iov[seg].iov_len + PAGE_SIZE-1) /
1221 PAGE_SIZE - user_addr / PAGE_SIZE);
1224 for (seg = 0; seg < nr_segs; seg++) {
1225 user_addr = (unsigned long)iov[seg].iov_base;
1226 sdio.size += bytes = iov[seg].iov_len;
1228 /* Index into the first page of the first block */
1229 sdio.first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
1230 sdio.final_block_in_request = sdio.block_in_file +
1232 /* Page fetching state */
1237 sdio.total_pages = 0;
1238 if (user_addr & (PAGE_SIZE-1)) {
1240 bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
1242 sdio.total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
1243 sdio.curr_user_address = user_addr;
1245 retval = do_direct_IO(dio, &sdio, &map_bh);
1247 dio->result += iov[seg].iov_len -
1248 ((sdio.final_block_in_request - sdio.block_in_file) <<
1252 dio_cleanup(dio, &sdio);
1255 } /* end iovec loop */
1257 if (retval == -ENOTBLK) {
1259 * The remaining part of the request will be
1260 * be handled by buffered I/O when we return
1265 * There may be some unwritten disk at the end of a part-written
1266 * fs-block-sized block. Go zero that now.
1268 dio_zero_block(dio, &sdio, 1, &map_bh);
1270 if (sdio.cur_page) {
1273 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1276 page_cache_release(sdio.cur_page);
1277 sdio.cur_page = NULL;
1280 dio_bio_submit(dio, &sdio);
1283 * It is possible that, we return short IO due to end of file.
1284 * In that case, we need to release all the pages we got hold on.
1286 dio_cleanup(dio, &sdio);
1289 * All block lookups have been performed. For READ requests
1290 * we can let i_mutex go now that its achieved its purpose
1291 * of protecting us from looking up uninitialized blocks.
1293 if (rw == READ && (dio->flags & DIO_LOCKING))
1294 mutex_unlock(&dio->inode->i_mutex);
1297 * The only time we want to leave bios in flight is when a successful
1298 * partial aio read or full aio write have been setup. In that case
1299 * bio completion will call aio_complete. The only time it's safe to
1300 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1301 * This had *better* be the only place that raises -EIOCBQUEUED.
1303 BUG_ON(retval == -EIOCBQUEUED);
1304 if (dio->is_async && retval == 0 && dio->result &&
1305 ((rw & READ) || (dio->result == sdio.size)))
1306 retval = -EIOCBQUEUED;
1308 if (retval != -EIOCBQUEUED)
1309 dio_await_completion(dio);
1311 if (drop_refcount(dio) == 0) {
1312 retval = dio_complete(dio, offset, retval, false);
1313 kmem_cache_free(dio_cache, dio);
1315 BUG_ON(retval != -EIOCBQUEUED);
1322 __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1323 struct block_device *bdev, const struct iovec *iov, loff_t offset,
1324 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1325 dio_submit_t submit_io, int flags)
1328 * The block device state is needed in the end to finally
1329 * submit everything. Since it's likely to be cache cold
1330 * prefetch it here as first thing to hide some of the
1333 * Attempt to prefetch the pieces we likely need later.
1335 prefetch(&bdev->bd_disk->part_tbl);
1336 prefetch(bdev->bd_queue);
1337 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1339 return do_blockdev_direct_IO(rw, iocb, inode, bdev, iov, offset,
1340 nr_segs, get_block, end_io,
1344 EXPORT_SYMBOL(__blockdev_direct_IO);
1346 static __init int dio_init(void)
1348 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1351 module_init(dio_init)