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>
41 * How many user pages to map in one call to get_user_pages(). This determines
42 * the size of a structure in the slab cache
47 * This code generally works in units of "dio_blocks". A dio_block is
48 * somewhere between the hard sector size and the filesystem block size. it
49 * is determined on a per-invocation basis. When talking to the filesystem
50 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
51 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
52 * to bio_block quantities by shifting left by blkfactor.
54 * If blkfactor is zero then the user's request was aligned to the filesystem's
58 /* dio_state only used in the submission path */
61 struct bio *bio; /* bio under assembly */
62 unsigned blkbits; /* doesn't change */
63 unsigned blkfactor; /* When we're using an alignment which
64 is finer than the filesystem's soft
65 blocksize, this specifies how much
66 finer. blkfactor=2 means 1/4-block
67 alignment. Does not change */
68 unsigned start_zero_done; /* flag: sub-blocksize zeroing has
69 been performed at the start of a
71 int pages_in_io; /* approximate total IO pages */
72 size_t size; /* total request size (doesn't change)*/
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 unsigned first_block_in_page; /* doesn't change, Used only once */
79 int boundary; /* prev block is at a boundary */
80 get_block_t *get_block; /* block mapping function */
81 dio_submit_t *submit_io; /* IO submition function */
83 loff_t logical_offset_in_bio; /* current first logical block in bio */
84 sector_t final_block_in_bio; /* current final block in bio + 1 */
85 sector_t next_block_for_io; /* next block to be put under IO,
86 in dio_blocks units */
89 * Deferred addition of a page to the dio. These variables are
90 * private to dio_send_cur_page(), submit_page_section() and
93 struct page *cur_page; /* The page */
94 unsigned cur_page_offset; /* Offset into it, in bytes */
95 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
96 sector_t cur_page_block; /* Where it starts */
97 loff_t cur_page_fs_offset; /* Offset in file */
100 * Page fetching state. These variables belong to dio_refill_pages().
102 int curr_page; /* changes */
103 int total_pages; /* doesn't change */
104 unsigned long curr_user_address;/* changes */
107 * Page queue. These variables belong to dio_refill_pages() and
110 unsigned head; /* next page to process */
111 unsigned tail; /* last valid page + 1 */
114 /* dio_state communicated between submission path and end_io */
116 int flags; /* doesn't change */
119 loff_t i_size; /* i_size when submitted */
120 dio_iodone_t *end_io; /* IO completion function */
122 void *private; /* copy from map_bh.b_private */
124 /* BIO completion state */
125 spinlock_t bio_lock; /* protects BIO fields below */
126 int page_errors; /* errno from get_user_pages() */
127 int is_async; /* is IO async ? */
128 int io_error; /* IO error in completion path */
129 unsigned long refcount; /* direct_io_worker() and bios */
130 struct bio *bio_list; /* singly linked via bi_private */
131 struct task_struct *waiter; /* waiting task (NULL if none) */
133 /* AIO related stuff */
134 struct kiocb *iocb; /* kiocb */
135 ssize_t result; /* IO result */
138 * pages[] (and any fields placed after it) are not zeroed out at
139 * allocation time. Don't add new fields after pages[] unless you
140 * wish that they not be zeroed.
142 struct page *pages[DIO_PAGES]; /* page buffer */
143 } ____cacheline_aligned_in_smp;
145 static struct kmem_cache *dio_cache __read_mostly;
147 static void __inode_dio_wait(struct inode *inode)
149 wait_queue_head_t *wq = bit_waitqueue(&inode->i_state, __I_DIO_WAKEUP);
150 DEFINE_WAIT_BIT(q, &inode->i_state, __I_DIO_WAKEUP);
153 prepare_to_wait(wq, &q.wait, TASK_UNINTERRUPTIBLE);
154 if (atomic_read(&inode->i_dio_count))
156 } while (atomic_read(&inode->i_dio_count));
157 finish_wait(wq, &q.wait);
161 * inode_dio_wait - wait for outstanding DIO requests to finish
162 * @inode: inode to wait for
164 * Waits for all pending direct I/O requests to finish so that we can
165 * proceed with a truncate or equivalent operation.
167 * Must be called under a lock that serializes taking new references
168 * to i_dio_count, usually by inode->i_mutex.
170 void inode_dio_wait(struct inode *inode)
172 if (atomic_read(&inode->i_dio_count))
173 __inode_dio_wait(inode);
175 EXPORT_SYMBOL_GPL(inode_dio_wait);
178 * inode_dio_done - signal finish of a direct I/O requests
179 * @inode: inode the direct I/O happens on
181 * This is called once we've finished processing a direct I/O request,
182 * and is used to wake up callers waiting for direct I/O to be quiesced.
184 void inode_dio_done(struct inode *inode)
186 if (atomic_dec_and_test(&inode->i_dio_count))
187 wake_up_bit(&inode->i_state, __I_DIO_WAKEUP);
189 EXPORT_SYMBOL_GPL(inode_dio_done);
192 * How many pages are in the queue?
194 static inline unsigned dio_pages_present(struct dio_submit *sdio)
196 return sdio->tail - sdio->head;
200 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
202 static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
207 nr_pages = min(sdio->total_pages - sdio->curr_page, DIO_PAGES);
208 ret = get_user_pages_fast(
209 sdio->curr_user_address, /* Where from? */
210 nr_pages, /* How many pages? */
211 dio->rw == READ, /* Write to memory? */
212 &dio->pages[0]); /* Put results here */
214 if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) {
215 struct page *page = ZERO_PAGE(0);
217 * A memory fault, but the filesystem has some outstanding
218 * mapped blocks. We need to use those blocks up to avoid
219 * leaking stale data in the file.
221 if (dio->page_errors == 0)
222 dio->page_errors = ret;
223 page_cache_get(page);
224 dio->pages[0] = page;
232 sdio->curr_user_address += ret * PAGE_SIZE;
233 sdio->curr_page += ret;
243 * Get another userspace page. Returns an ERR_PTR on error. Pages are
244 * buffered inside the dio so that we can call get_user_pages() against a
245 * decent number of pages, less frequently. To provide nicer use of the
248 static inline struct page *dio_get_page(struct dio *dio,
249 struct dio_submit *sdio)
251 if (dio_pages_present(sdio) == 0) {
254 ret = dio_refill_pages(dio, sdio);
257 BUG_ON(dio_pages_present(sdio) == 0);
259 return dio->pages[sdio->head++];
263 * dio_complete() - called when all DIO BIO I/O has been completed
264 * @offset: the byte offset in the file of the completed operation
266 * This releases locks as dictated by the locking type, lets interested parties
267 * know that a DIO operation has completed, and calculates the resulting return
268 * code for the operation.
270 * It lets the filesystem know if it registered an interest earlier via
271 * get_block. Pass the private field of the map buffer_head so that
272 * filesystems can use it to hold additional state between get_block calls and
275 static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret, bool is_async)
277 ssize_t transferred = 0;
280 * AIO submission can race with bio completion to get here while
281 * expecting to have the last io completed by bio completion.
282 * In that case -EIOCBQUEUED is in fact not an error we want
283 * to preserve through this call.
285 if (ret == -EIOCBQUEUED)
289 transferred = dio->result;
291 /* Check for short read case */
292 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
293 transferred = dio->i_size - offset;
297 ret = dio->page_errors;
303 if (dio->end_io && dio->result) {
304 dio->end_io(dio->iocb, offset, transferred,
305 dio->private, ret, is_async);
308 aio_complete(dio->iocb, ret, 0);
309 inode_dio_done(dio->inode);
315 static int dio_bio_complete(struct dio *dio, struct bio *bio);
317 * Asynchronous IO callback.
319 static void dio_bio_end_aio(struct bio *bio, int error)
321 struct dio *dio = bio->bi_private;
322 unsigned long remaining;
325 /* cleanup the bio */
326 dio_bio_complete(dio, bio);
328 spin_lock_irqsave(&dio->bio_lock, flags);
329 remaining = --dio->refcount;
330 if (remaining == 1 && dio->waiter)
331 wake_up_process(dio->waiter);
332 spin_unlock_irqrestore(&dio->bio_lock, flags);
334 if (remaining == 0) {
335 dio_complete(dio, dio->iocb->ki_pos, 0, true);
336 kmem_cache_free(dio_cache, dio);
341 * The BIO completion handler simply queues the BIO up for the process-context
344 * During I/O bi_private points at the dio. After I/O, bi_private is used to
345 * implement a singly-linked list of completed BIOs, at dio->bio_list.
347 static void dio_bio_end_io(struct bio *bio, int error)
349 struct dio *dio = bio->bi_private;
352 spin_lock_irqsave(&dio->bio_lock, flags);
353 bio->bi_private = dio->bio_list;
355 if (--dio->refcount == 1 && dio->waiter)
356 wake_up_process(dio->waiter);
357 spin_unlock_irqrestore(&dio->bio_lock, flags);
361 * dio_end_io - handle the end io action for the given bio
362 * @bio: The direct io bio thats being completed
363 * @error: Error if there was one
365 * This is meant to be called by any filesystem that uses their own dio_submit_t
366 * so that the DIO specific endio actions are dealt with after the filesystem
367 * has done it's completion work.
369 void dio_end_io(struct bio *bio, int error)
371 struct dio *dio = bio->bi_private;
374 dio_bio_end_aio(bio, error);
376 dio_bio_end_io(bio, error);
378 EXPORT_SYMBOL_GPL(dio_end_io);
381 dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
382 struct block_device *bdev,
383 sector_t first_sector, int nr_vecs)
388 * bio_alloc() is guaranteed to return a bio when called with
389 * __GFP_WAIT and we request a valid number of vectors.
391 bio = bio_alloc(GFP_KERNEL, nr_vecs);
394 bio->bi_sector = first_sector;
396 bio->bi_end_io = dio_bio_end_aio;
398 bio->bi_end_io = dio_bio_end_io;
401 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
405 * In the AIO read case we speculatively dirty the pages before starting IO.
406 * During IO completion, any of these pages which happen to have been written
407 * back will be redirtied by bio_check_pages_dirty().
409 * bios hold a dio reference between submit_bio and ->end_io.
411 static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
413 struct bio *bio = sdio->bio;
416 bio->bi_private = dio;
418 spin_lock_irqsave(&dio->bio_lock, flags);
420 spin_unlock_irqrestore(&dio->bio_lock, flags);
422 if (dio->is_async && dio->rw == READ)
423 bio_set_pages_dirty(bio);
426 sdio->submit_io(dio->rw, bio, dio->inode,
427 sdio->logical_offset_in_bio);
429 submit_bio(dio->rw, bio);
433 sdio->logical_offset_in_bio = 0;
437 * Release any resources in case of a failure
439 static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
441 while (dio_pages_present(sdio))
442 page_cache_release(dio_get_page(dio, sdio));
446 * Wait for the next BIO to complete. Remove it and return it. NULL is
447 * returned once all BIOs have been completed. This must only be called once
448 * all bios have been issued so that dio->refcount can only decrease. This
449 * requires that that the caller hold a reference on the dio.
451 static struct bio *dio_await_one(struct dio *dio)
454 struct bio *bio = NULL;
456 spin_lock_irqsave(&dio->bio_lock, flags);
459 * Wait as long as the list is empty and there are bios in flight. bio
460 * completion drops the count, maybe adds to the list, and wakes while
461 * holding the bio_lock so we don't need set_current_state()'s barrier
462 * and can call it after testing our condition.
464 while (dio->refcount > 1 && dio->bio_list == NULL) {
465 __set_current_state(TASK_UNINTERRUPTIBLE);
466 dio->waiter = current;
467 spin_unlock_irqrestore(&dio->bio_lock, flags);
469 /* wake up sets us TASK_RUNNING */
470 spin_lock_irqsave(&dio->bio_lock, flags);
475 dio->bio_list = bio->bi_private;
477 spin_unlock_irqrestore(&dio->bio_lock, flags);
482 * Process one completed BIO. No locks are held.
484 static int dio_bio_complete(struct dio *dio, struct bio *bio)
486 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
487 struct bio_vec *bvec = bio->bi_io_vec;
491 dio->io_error = -EIO;
493 if (dio->is_async && dio->rw == READ) {
494 bio_check_pages_dirty(bio); /* transfers ownership */
496 for (page_no = 0; page_no < bio->bi_vcnt; page_no++) {
497 struct page *page = bvec[page_no].bv_page;
499 if (dio->rw == READ && !PageCompound(page))
500 set_page_dirty_lock(page);
501 page_cache_release(page);
505 return uptodate ? 0 : -EIO;
509 * Wait on and process all in-flight BIOs. This must only be called once
510 * all bios have been issued so that the refcount can only decrease.
511 * This just waits for all bios to make it through dio_bio_complete. IO
512 * errors are propagated through dio->io_error and should be propagated via
515 static void dio_await_completion(struct dio *dio)
519 bio = dio_await_one(dio);
521 dio_bio_complete(dio, bio);
526 * A really large O_DIRECT read or write can generate a lot of BIOs. So
527 * to keep the memory consumption sane we periodically reap any completed BIOs
528 * during the BIO generation phase.
530 * This also helps to limit the peak amount of pinned userspace memory.
532 static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
536 if (sdio->reap_counter++ >= 64) {
537 while (dio->bio_list) {
542 spin_lock_irqsave(&dio->bio_lock, flags);
544 dio->bio_list = bio->bi_private;
545 spin_unlock_irqrestore(&dio->bio_lock, flags);
546 ret2 = dio_bio_complete(dio, bio);
550 sdio->reap_counter = 0;
556 * Call into the fs to map some more disk blocks. We record the current number
557 * of available blocks at sdio->blocks_available. These are in units of the
558 * fs blocksize, (1 << inode->i_blkbits).
560 * The fs is allowed to map lots of blocks at once. If it wants to do that,
561 * it uses the passed inode-relative block number as the file offset, as usual.
563 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
564 * has remaining to do. The fs should not map more than this number of blocks.
566 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
567 * indicate how much contiguous disk space has been made available at
570 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
571 * This isn't very efficient...
573 * In the case of filesystem holes: the fs may return an arbitrarily-large
574 * hole by returning an appropriate value in b_size and by clearing
575 * buffer_mapped(). However the direct-io code will only process holes one
576 * block at a time - it will repeatedly call get_block() as it walks the hole.
578 static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
579 struct buffer_head *map_bh)
582 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
583 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
584 unsigned long fs_count; /* Number of filesystem-sized blocks */
588 * If there was a memory error and we've overwritten all the
589 * mapped blocks then we can now return that memory error
591 ret = dio->page_errors;
593 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
594 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
595 fs_endblk = (sdio->final_block_in_request - 1) >>
597 fs_count = fs_endblk - fs_startblk + 1;
600 map_bh->b_size = fs_count << dio->inode->i_blkbits;
603 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we
604 * forbid block creations: only overwrites are permitted.
605 * We will return early to the caller once we see an
606 * unmapped buffer head returned, and the caller will fall
607 * back to buffered I/O.
609 * Otherwise the decision is left to the get_blocks method,
610 * which may decide to handle it or also return an unmapped
613 create = dio->rw & WRITE;
614 if (dio->flags & DIO_SKIP_HOLES) {
615 if (sdio->block_in_file < (i_size_read(dio->inode) >>
620 ret = (*sdio->get_block)(dio->inode, fs_startblk,
623 /* Store for completion */
624 dio->private = map_bh->b_private;
630 * There is no bio. Make one now.
632 static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
633 sector_t start_sector, struct buffer_head *map_bh)
638 ret = dio_bio_reap(dio, sdio);
641 sector = start_sector << (sdio->blkbits - 9);
642 nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(map_bh->b_bdev));
643 nr_pages = min(nr_pages, BIO_MAX_PAGES);
644 BUG_ON(nr_pages <= 0);
645 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
652 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
653 * that was successful then update final_block_in_bio and take a ref against
654 * the just-added page.
656 * Return zero on success. Non-zero means the caller needs to start a new BIO.
658 static inline int dio_bio_add_page(struct dio_submit *sdio)
662 ret = bio_add_page(sdio->bio, sdio->cur_page,
663 sdio->cur_page_len, sdio->cur_page_offset);
664 if (ret == sdio->cur_page_len) {
666 * Decrement count only, if we are done with this page
668 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
670 page_cache_get(sdio->cur_page);
671 sdio->final_block_in_bio = sdio->cur_page_block +
672 (sdio->cur_page_len >> sdio->blkbits);
681 * Put cur_page under IO. The section of cur_page which is described by
682 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
683 * starts on-disk at cur_page_block.
685 * We take a ref against the page here (on behalf of its presence in the bio).
687 * The caller of this function is responsible for removing cur_page from the
688 * dio, and for dropping the refcount which came from that presence.
690 static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
691 struct buffer_head *map_bh)
696 loff_t cur_offset = sdio->cur_page_fs_offset;
697 loff_t bio_next_offset = sdio->logical_offset_in_bio +
701 * See whether this new request is contiguous with the old.
703 * Btrfs cannot handle having logically non-contiguous requests
704 * submitted. For example if you have
706 * Logical: [0-4095][HOLE][8192-12287]
707 * Physical: [0-4095] [4096-8191]
709 * We cannot submit those pages together as one BIO. So if our
710 * current logical offset in the file does not equal what would
711 * be the next logical offset in the bio, submit the bio we
714 if (sdio->final_block_in_bio != sdio->cur_page_block ||
715 cur_offset != bio_next_offset)
716 dio_bio_submit(dio, sdio);
718 * Submit now if the underlying fs is about to perform a
721 else if (sdio->boundary)
722 dio_bio_submit(dio, sdio);
725 if (sdio->bio == NULL) {
726 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
731 if (dio_bio_add_page(sdio) != 0) {
732 dio_bio_submit(dio, sdio);
733 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
735 ret = dio_bio_add_page(sdio);
744 * An autonomous function to put a chunk of a page under deferred IO.
746 * The caller doesn't actually know (or care) whether this piece of page is in
747 * a BIO, or is under IO or whatever. We just take care of all possible
748 * situations here. The separation between the logic of do_direct_IO() and
749 * that of submit_page_section() is important for clarity. Please don't break.
751 * The chunk of page starts on-disk at blocknr.
753 * We perform deferred IO, by recording the last-submitted page inside our
754 * private part of the dio structure. If possible, we just expand the IO
755 * across that page here.
757 * If that doesn't work out then we put the old page into the bio and add this
758 * page to the dio instead.
761 submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
762 unsigned offset, unsigned len, sector_t blocknr,
763 struct buffer_head *map_bh)
767 if (dio->rw & WRITE) {
769 * Read accounting is performed in submit_bio()
771 task_io_account_write(len);
775 * Can we just grow the current page's presence in the dio?
777 if (sdio->cur_page == page &&
778 sdio->cur_page_offset + sdio->cur_page_len == offset &&
779 sdio->cur_page_block +
780 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
781 sdio->cur_page_len += len;
784 * If sdio->boundary then we want to schedule the IO now to
785 * avoid metadata seeks.
787 if (sdio->boundary) {
788 ret = dio_send_cur_page(dio, sdio, map_bh);
789 page_cache_release(sdio->cur_page);
790 sdio->cur_page = NULL;
796 * If there's a deferred page already there then send it.
798 if (sdio->cur_page) {
799 ret = dio_send_cur_page(dio, sdio, map_bh);
800 page_cache_release(sdio->cur_page);
801 sdio->cur_page = NULL;
806 page_cache_get(page); /* It is in dio */
807 sdio->cur_page = page;
808 sdio->cur_page_offset = offset;
809 sdio->cur_page_len = len;
810 sdio->cur_page_block = blocknr;
811 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
817 * Clean any dirty buffers in the blockdev mapping which alias newly-created
818 * file blocks. Only called for S_ISREG files - blockdevs do not set
821 static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
826 nblocks = map_bh->b_size >> dio->inode->i_blkbits;
828 for (i = 0; i < nblocks; i++) {
829 unmap_underlying_metadata(map_bh->b_bdev,
830 map_bh->b_blocknr + i);
835 * If we are not writing the entire block and get_block() allocated
836 * the block for us, we need to fill-in the unused portion of the
837 * block with zeros. This happens only if user-buffer, fileoffset or
838 * io length is not filesystem block-size multiple.
840 * `end' is zero if we're doing the start of the IO, 1 at the end of the
843 static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
844 int end, struct buffer_head *map_bh)
846 unsigned dio_blocks_per_fs_block;
847 unsigned this_chunk_blocks; /* In dio_blocks */
848 unsigned this_chunk_bytes;
851 sdio->start_zero_done = 1;
852 if (!sdio->blkfactor || !buffer_new(map_bh))
855 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
856 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
858 if (!this_chunk_blocks)
862 * We need to zero out part of an fs block. It is either at the
863 * beginning or the end of the fs block.
866 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
868 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
871 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
872 sdio->next_block_for_io, map_bh))
875 sdio->next_block_for_io += this_chunk_blocks;
879 * Walk the user pages, and the file, mapping blocks to disk and generating
880 * a sequence of (page,offset,len,block) mappings. These mappings are injected
881 * into submit_page_section(), which takes care of the next stage of submission
883 * Direct IO against a blockdev is different from a file. Because we can
884 * happily perform page-sized but 512-byte aligned IOs. It is important that
885 * blockdev IO be able to have fine alignment and large sizes.
887 * So what we do is to permit the ->get_block function to populate bh.b_size
888 * with the size of IO which is permitted at this offset and this i_blkbits.
890 * For best results, the blockdev should be set up with 512-byte i_blkbits and
891 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
892 * fine alignment but still allows this function to work in PAGE_SIZE units.
894 static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
895 struct buffer_head *map_bh)
897 const unsigned blkbits = sdio->blkbits;
898 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
900 unsigned block_in_page;
903 /* The I/O can start at any block offset within the first page */
904 block_in_page = sdio->first_block_in_page;
906 while (sdio->block_in_file < sdio->final_block_in_request) {
907 page = dio_get_page(dio, sdio);
913 while (block_in_page < blocks_per_page) {
914 unsigned offset_in_page = block_in_page << blkbits;
915 unsigned this_chunk_bytes; /* # of bytes mapped */
916 unsigned this_chunk_blocks; /* # of blocks */
919 if (sdio->blocks_available == 0) {
921 * Need to go and map some more disk
923 unsigned long blkmask;
924 unsigned long dio_remainder;
926 ret = get_more_blocks(dio, sdio, map_bh);
928 page_cache_release(page);
931 if (!buffer_mapped(map_bh))
934 sdio->blocks_available =
935 map_bh->b_size >> sdio->blkbits;
936 sdio->next_block_for_io =
937 map_bh->b_blocknr << sdio->blkfactor;
938 if (buffer_new(map_bh))
939 clean_blockdev_aliases(dio, map_bh);
941 if (!sdio->blkfactor)
944 blkmask = (1 << sdio->blkfactor) - 1;
945 dio_remainder = (sdio->block_in_file & blkmask);
948 * If we are at the start of IO and that IO
949 * starts partway into a fs-block,
950 * dio_remainder will be non-zero. If the IO
951 * is a read then we can simply advance the IO
952 * cursor to the first block which is to be
953 * read. But if the IO is a write and the
954 * block was newly allocated we cannot do that;
955 * the start of the fs block must be zeroed out
958 if (!buffer_new(map_bh))
959 sdio->next_block_for_io += dio_remainder;
960 sdio->blocks_available -= dio_remainder;
964 if (!buffer_mapped(map_bh)) {
965 loff_t i_size_aligned;
967 /* AKPM: eargh, -ENOTBLK is a hack */
968 if (dio->rw & WRITE) {
969 page_cache_release(page);
974 * Be sure to account for a partial block as the
975 * last block in the file
977 i_size_aligned = ALIGN(i_size_read(dio->inode),
979 if (sdio->block_in_file >=
980 i_size_aligned >> blkbits) {
982 page_cache_release(page);
985 zero_user(page, block_in_page << blkbits,
987 sdio->block_in_file++;
993 * If we're performing IO which has an alignment which
994 * is finer than the underlying fs, go check to see if
995 * we must zero out the start of this block.
997 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
998 dio_zero_block(dio, sdio, 0, map_bh);
1001 * Work out, in this_chunk_blocks, how much disk we
1002 * can add to this page
1004 this_chunk_blocks = sdio->blocks_available;
1005 u = (PAGE_SIZE - offset_in_page) >> blkbits;
1006 if (this_chunk_blocks > u)
1007 this_chunk_blocks = u;
1008 u = sdio->final_block_in_request - sdio->block_in_file;
1009 if (this_chunk_blocks > u)
1010 this_chunk_blocks = u;
1011 this_chunk_bytes = this_chunk_blocks << blkbits;
1012 BUG_ON(this_chunk_bytes == 0);
1014 sdio->boundary = buffer_boundary(map_bh);
1015 ret = submit_page_section(dio, sdio, page,
1018 sdio->next_block_for_io,
1021 page_cache_release(page);
1024 sdio->next_block_for_io += this_chunk_blocks;
1026 sdio->block_in_file += this_chunk_blocks;
1027 block_in_page += this_chunk_blocks;
1028 sdio->blocks_available -= this_chunk_blocks;
1030 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1031 if (sdio->block_in_file == sdio->final_block_in_request)
1035 /* Drop the ref which was taken in get_user_pages() */
1036 page_cache_release(page);
1043 static inline int drop_refcount(struct dio *dio)
1046 unsigned long flags;
1049 * Sync will always be dropping the final ref and completing the
1050 * operation. AIO can if it was a broken operation described above or
1051 * in fact if all the bios race to complete before we get here. In
1052 * that case dio_complete() translates the EIOCBQUEUED into the proper
1053 * return code that the caller will hand to aio_complete().
1055 * This is managed by the bio_lock instead of being an atomic_t so that
1056 * completion paths can drop their ref and use the remaining count to
1057 * decide to wake the submission path atomically.
1059 spin_lock_irqsave(&dio->bio_lock, flags);
1060 ret2 = --dio->refcount;
1061 spin_unlock_irqrestore(&dio->bio_lock, flags);
1066 * This is a library function for use by filesystem drivers.
1068 * The locking rules are governed by the flags parameter:
1069 * - if the flags value contains DIO_LOCKING we use a fancy locking
1070 * scheme for dumb filesystems.
1071 * For writes this function is called under i_mutex and returns with
1072 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1073 * taken and dropped again before returning.
1074 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1075 * internal locking but rather rely on the filesystem to synchronize
1076 * direct I/O reads/writes versus each other and truncate.
1078 * To help with locking against truncate we incremented the i_dio_count
1079 * counter before starting direct I/O, and decrement it once we are done.
1080 * Truncate can wait for it to reach zero to provide exclusion. It is
1081 * expected that filesystem provide exclusion between new direct I/O
1082 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1083 * but other filesystems need to take care of this on their own.
1085 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1086 * is always inlined. Otherwise gcc is unable to split the structure into
1087 * individual fields and will generate much worse code. This is important
1088 * for the whole file.
1090 static inline ssize_t
1091 do_blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1092 struct block_device *bdev, const struct iovec *iov, loff_t offset,
1093 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1094 dio_submit_t submit_io, int flags)
1099 unsigned blkbits = inode->i_blkbits;
1100 unsigned blocksize_mask = (1 << blkbits) - 1;
1101 ssize_t retval = -EINVAL;
1102 loff_t end = offset;
1104 struct dio_submit sdio = { 0, };
1105 unsigned long user_addr;
1107 struct buffer_head map_bh = { 0, };
1113 * Avoid references to bdev if not absolutely needed to give
1114 * the early prefetch in the caller enough time.
1117 if (offset & blocksize_mask) {
1119 blkbits = blksize_bits(bdev_logical_block_size(bdev));
1120 blocksize_mask = (1 << blkbits) - 1;
1121 if (offset & blocksize_mask)
1125 /* Check the memory alignment. Blocks cannot straddle pages */
1126 for (seg = 0; seg < nr_segs; seg++) {
1127 addr = (unsigned long)iov[seg].iov_base;
1128 size = iov[seg].iov_len;
1130 if (unlikely((addr & blocksize_mask) ||
1131 (size & blocksize_mask))) {
1133 blkbits = blksize_bits(
1134 bdev_logical_block_size(bdev));
1135 blocksize_mask = (1 << blkbits) - 1;
1136 if ((addr & blocksize_mask) || (size & blocksize_mask))
1141 /* watch out for a 0 len io from a tricksy fs */
1142 if (rw == READ && end == offset)
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) {
1159 struct address_space *mapping =
1160 iocb->ki_filp->f_mapping;
1162 /* will be released by direct_io_worker */
1163 mutex_lock(&inode->i_mutex);
1165 retval = filemap_write_and_wait_range(mapping, offset,
1168 mutex_unlock(&inode->i_mutex);
1169 kmem_cache_free(dio_cache, dio);
1176 * Will be decremented at I/O completion time.
1178 atomic_inc(&inode->i_dio_count);
1181 * For file extending writes updating i_size before data
1182 * writeouts complete can expose uninitialized blocks. So
1183 * even for AIO, we need to wait for i/o to complete before
1184 * returning in this case.
1186 dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
1187 (end > i_size_read(inode)));
1193 sdio.blkbits = blkbits;
1194 sdio.blkfactor = inode->i_blkbits - blkbits;
1195 sdio.block_in_file = offset >> blkbits;
1197 sdio.get_block = get_block;
1198 dio->end_io = end_io;
1199 sdio.submit_io = submit_io;
1200 sdio.final_block_in_bio = -1;
1201 sdio.next_block_for_io = -1;
1204 dio->i_size = i_size_read(inode);
1206 spin_lock_init(&dio->bio_lock);
1210 * In case of non-aligned buffers, we may need 2 more
1211 * pages since we need to zero out first and last block.
1213 if (unlikely(sdio.blkfactor))
1214 sdio.pages_in_io = 2;
1216 for (seg = 0; seg < nr_segs; seg++) {
1217 user_addr = (unsigned long)iov[seg].iov_base;
1219 ((user_addr + iov[seg].iov_len + PAGE_SIZE-1) /
1220 PAGE_SIZE - user_addr / PAGE_SIZE);
1223 for (seg = 0; seg < nr_segs; seg++) {
1224 user_addr = (unsigned long)iov[seg].iov_base;
1225 sdio.size += bytes = iov[seg].iov_len;
1227 /* Index into the first page of the first block */
1228 sdio.first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
1229 sdio.final_block_in_request = sdio.block_in_file +
1231 /* Page fetching state */
1236 sdio.total_pages = 0;
1237 if (user_addr & (PAGE_SIZE-1)) {
1239 bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
1241 sdio.total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
1242 sdio.curr_user_address = user_addr;
1244 retval = do_direct_IO(dio, &sdio, &map_bh);
1246 dio->result += iov[seg].iov_len -
1247 ((sdio.final_block_in_request - sdio.block_in_file) <<
1251 dio_cleanup(dio, &sdio);
1254 } /* end iovec loop */
1256 if (retval == -ENOTBLK) {
1258 * The remaining part of the request will be
1259 * be handled by buffered I/O when we return
1264 * There may be some unwritten disk at the end of a part-written
1265 * fs-block-sized block. Go zero that now.
1267 dio_zero_block(dio, &sdio, 1, &map_bh);
1269 if (sdio.cur_page) {
1272 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1275 page_cache_release(sdio.cur_page);
1276 sdio.cur_page = NULL;
1279 dio_bio_submit(dio, &sdio);
1282 * It is possible that, we return short IO due to end of file.
1283 * In that case, we need to release all the pages we got hold on.
1285 dio_cleanup(dio, &sdio);
1288 * All block lookups have been performed. For READ requests
1289 * we can let i_mutex go now that its achieved its purpose
1290 * of protecting us from looking up uninitialized blocks.
1292 if (rw == READ && (dio->flags & DIO_LOCKING))
1293 mutex_unlock(&dio->inode->i_mutex);
1296 * The only time we want to leave bios in flight is when a successful
1297 * partial aio read or full aio write have been setup. In that case
1298 * bio completion will call aio_complete. The only time it's safe to
1299 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1300 * This had *better* be the only place that raises -EIOCBQUEUED.
1302 BUG_ON(retval == -EIOCBQUEUED);
1303 if (dio->is_async && retval == 0 && dio->result &&
1304 ((rw & READ) || (dio->result == sdio.size)))
1305 retval = -EIOCBQUEUED;
1307 if (retval != -EIOCBQUEUED)
1308 dio_await_completion(dio);
1310 if (drop_refcount(dio) == 0) {
1311 retval = dio_complete(dio, offset, retval, false);
1312 kmem_cache_free(dio_cache, dio);
1314 BUG_ON(retval != -EIOCBQUEUED);
1321 __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1322 struct block_device *bdev, const struct iovec *iov, loff_t offset,
1323 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1324 dio_submit_t submit_io, int flags)
1327 * The block device state is needed in the end to finally
1328 * submit everything. Since it's likely to be cache cold
1329 * prefetch it here as first thing to hide some of the
1332 * Attempt to prefetch the pieces we likely need later.
1334 prefetch(&bdev->bd_disk->part_tbl);
1335 prefetch(bdev->bd_queue);
1336 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1338 return do_blockdev_direct_IO(rw, iocb, inode, bdev, iov, offset,
1339 nr_segs, get_block, end_io,
1343 EXPORT_SYMBOL(__blockdev_direct_IO);
1345 static __init int dio_init(void)
1347 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1350 module_init(dio_init)