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 <asm/atomic.h>
41 * How many user pages to map in one call to get_user_pages(). This determines
42 * the size of a structure on the stack.
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
59 /* BIO submission state */
60 struct bio *bio; /* bio under assembly */
63 loff_t i_size; /* i_size when submitted */
64 int flags; /* doesn't change */
65 unsigned blkbits; /* doesn't change */
66 unsigned blkfactor; /* When we're using an alignment which
67 is finer than the filesystem's soft
68 blocksize, this specifies how much
69 finer. blkfactor=2 means 1/4-block
70 alignment. Does not change */
71 unsigned start_zero_done; /* flag: sub-blocksize zeroing has
72 been performed at the start of a
74 int pages_in_io; /* approximate total IO pages */
75 size_t size; /* total request size (doesn't change)*/
76 sector_t block_in_file; /* Current offset into the underlying
77 file in dio_block units. */
78 unsigned blocks_available; /* At block_in_file. changes */
79 sector_t final_block_in_request;/* doesn't change */
80 unsigned first_block_in_page; /* doesn't change, Used only once */
81 int boundary; /* prev block is at a boundary */
82 int reap_counter; /* rate limit reaping */
83 get_block_t *get_block; /* block mapping function */
84 dio_iodone_t *end_io; /* IO completion function */
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 */
88 struct buffer_head map_bh; /* last get_block() result */
91 * Deferred addition of a page to the dio. These variables are
92 * private to dio_send_cur_page(), submit_page_section() and
95 struct page *cur_page; /* The page */
96 unsigned cur_page_offset; /* Offset into it, in bytes */
97 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
98 sector_t cur_page_block; /* Where it starts */
100 /* BIO completion state */
101 spinlock_t bio_lock; /* protects BIO fields below */
102 unsigned long refcount; /* direct_io_worker() and bios */
103 struct bio *bio_list; /* singly linked via bi_private */
104 struct task_struct *waiter; /* waiting task (NULL if none) */
106 /* AIO related stuff */
107 struct kiocb *iocb; /* kiocb */
108 int is_async; /* is IO async ? */
109 int io_error; /* IO error in completion path */
110 ssize_t result; /* IO result */
113 * Page fetching state. These variables belong to dio_refill_pages().
115 int curr_page; /* changes */
116 int total_pages; /* doesn't change */
117 unsigned long curr_user_address;/* changes */
120 * Page queue. These variables belong to dio_refill_pages() and
123 unsigned head; /* next page to process */
124 unsigned tail; /* last valid page + 1 */
125 int page_errors; /* errno from get_user_pages() */
128 * pages[] (and any fields placed after it) are not zeroed out at
129 * allocation time. Don't add new fields after pages[] unless you
130 * wish that they not be zeroed.
132 struct page *pages[DIO_PAGES]; /* page buffer */
136 * How many pages are in the queue?
138 static inline unsigned dio_pages_present(struct dio *dio)
140 return dio->tail - dio->head;
144 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
146 static int dio_refill_pages(struct dio *dio)
151 nr_pages = min(dio->total_pages - dio->curr_page, DIO_PAGES);
152 ret = get_user_pages_fast(
153 dio->curr_user_address, /* Where from? */
154 nr_pages, /* How many pages? */
155 dio->rw == READ, /* Write to memory? */
156 &dio->pages[0]); /* Put results here */
158 if (ret < 0 && dio->blocks_available && (dio->rw & WRITE)) {
159 struct page *page = ZERO_PAGE(0);
161 * A memory fault, but the filesystem has some outstanding
162 * mapped blocks. We need to use those blocks up to avoid
163 * leaking stale data in the file.
165 if (dio->page_errors == 0)
166 dio->page_errors = ret;
167 page_cache_get(page);
168 dio->pages[0] = page;
176 dio->curr_user_address += ret * PAGE_SIZE;
177 dio->curr_page += ret;
187 * Get another userspace page. Returns an ERR_PTR on error. Pages are
188 * buffered inside the dio so that we can call get_user_pages() against a
189 * decent number of pages, less frequently. To provide nicer use of the
192 static struct page *dio_get_page(struct dio *dio)
194 if (dio_pages_present(dio) == 0) {
197 ret = dio_refill_pages(dio);
200 BUG_ON(dio_pages_present(dio) == 0);
202 return dio->pages[dio->head++];
206 * dio_complete() - called when all DIO BIO I/O has been completed
207 * @offset: the byte offset in the file of the completed operation
209 * This releases locks as dictated by the locking type, lets interested parties
210 * know that a DIO operation has completed, and calculates the resulting return
211 * code for the operation.
213 * It lets the filesystem know if it registered an interest earlier via
214 * get_block. Pass the private field of the map buffer_head so that
215 * filesystems can use it to hold additional state between get_block calls and
218 static int dio_complete(struct dio *dio, loff_t offset, int ret)
220 ssize_t transferred = 0;
223 * AIO submission can race with bio completion to get here while
224 * expecting to have the last io completed by bio completion.
225 * In that case -EIOCBQUEUED is in fact not an error we want
226 * to preserve through this call.
228 if (ret == -EIOCBQUEUED)
232 transferred = dio->result;
234 /* Check for short read case */
235 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
236 transferred = dio->i_size - offset;
239 if (dio->end_io && dio->result)
240 dio->end_io(dio->iocb, offset, transferred,
241 dio->map_bh.b_private);
243 if (dio->flags & DIO_LOCKING)
244 /* lockdep: non-owner release */
245 up_read_non_owner(&dio->inode->i_alloc_sem);
248 ret = dio->page_errors;
257 static int dio_bio_complete(struct dio *dio, struct bio *bio);
259 * Asynchronous IO callback.
261 static void dio_bio_end_aio(struct bio *bio, int error)
263 struct dio *dio = bio->bi_private;
264 unsigned long remaining;
267 /* cleanup the bio */
268 dio_bio_complete(dio, bio);
270 spin_lock_irqsave(&dio->bio_lock, flags);
271 remaining = --dio->refcount;
272 if (remaining == 1 && dio->waiter)
273 wake_up_process(dio->waiter);
274 spin_unlock_irqrestore(&dio->bio_lock, flags);
276 if (remaining == 0) {
277 int ret = dio_complete(dio, dio->iocb->ki_pos, 0);
278 aio_complete(dio->iocb, ret, 0);
284 * The BIO completion handler simply queues the BIO up for the process-context
287 * During I/O bi_private points at the dio. After I/O, bi_private is used to
288 * implement a singly-linked list of completed BIOs, at dio->bio_list.
290 static void dio_bio_end_io(struct bio *bio, int error)
292 struct dio *dio = bio->bi_private;
295 spin_lock_irqsave(&dio->bio_lock, flags);
296 bio->bi_private = dio->bio_list;
298 if (--dio->refcount == 1 && dio->waiter)
299 wake_up_process(dio->waiter);
300 spin_unlock_irqrestore(&dio->bio_lock, flags);
304 dio_bio_alloc(struct dio *dio, struct block_device *bdev,
305 sector_t first_sector, int nr_vecs)
309 bio = bio_alloc(GFP_KERNEL, nr_vecs);
312 bio->bi_sector = first_sector;
314 bio->bi_end_io = dio_bio_end_aio;
316 bio->bi_end_io = dio_bio_end_io;
323 * In the AIO read case we speculatively dirty the pages before starting IO.
324 * During IO completion, any of these pages which happen to have been written
325 * back will be redirtied by bio_check_pages_dirty().
327 * bios hold a dio reference between submit_bio and ->end_io.
329 static void dio_bio_submit(struct dio *dio)
331 struct bio *bio = dio->bio;
334 bio->bi_private = dio;
336 spin_lock_irqsave(&dio->bio_lock, flags);
338 spin_unlock_irqrestore(&dio->bio_lock, flags);
340 if (dio->is_async && dio->rw == READ)
341 bio_set_pages_dirty(bio);
343 submit_bio(dio->rw, bio);
350 * Release any resources in case of a failure
352 static void dio_cleanup(struct dio *dio)
354 while (dio_pages_present(dio))
355 page_cache_release(dio_get_page(dio));
359 * Wait for the next BIO to complete. Remove it and return it. NULL is
360 * returned once all BIOs have been completed. This must only be called once
361 * all bios have been issued so that dio->refcount can only decrease. This
362 * requires that that the caller hold a reference on the dio.
364 static struct bio *dio_await_one(struct dio *dio)
367 struct bio *bio = NULL;
369 spin_lock_irqsave(&dio->bio_lock, flags);
372 * Wait as long as the list is empty and there are bios in flight. bio
373 * completion drops the count, maybe adds to the list, and wakes while
374 * holding the bio_lock so we don't need set_current_state()'s barrier
375 * and can call it after testing our condition.
377 while (dio->refcount > 1 && dio->bio_list == NULL) {
378 __set_current_state(TASK_UNINTERRUPTIBLE);
379 dio->waiter = current;
380 spin_unlock_irqrestore(&dio->bio_lock, flags);
382 /* wake up sets us TASK_RUNNING */
383 spin_lock_irqsave(&dio->bio_lock, flags);
388 dio->bio_list = bio->bi_private;
390 spin_unlock_irqrestore(&dio->bio_lock, flags);
395 * Process one completed BIO. No locks are held.
397 static int dio_bio_complete(struct dio *dio, struct bio *bio)
399 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
400 struct bio_vec *bvec = bio->bi_io_vec;
404 dio->io_error = -EIO;
406 if (dio->is_async && dio->rw == READ) {
407 bio_check_pages_dirty(bio); /* transfers ownership */
409 for (page_no = 0; page_no < bio->bi_vcnt; page_no++) {
410 struct page *page = bvec[page_no].bv_page;
412 if (dio->rw == READ && !PageCompound(page))
413 set_page_dirty_lock(page);
414 page_cache_release(page);
418 return uptodate ? 0 : -EIO;
422 * Wait on and process all in-flight BIOs. This must only be called once
423 * all bios have been issued so that the refcount can only decrease.
424 * This just waits for all bios to make it through dio_bio_complete. IO
425 * errors are propagated through dio->io_error and should be propagated via
428 static void dio_await_completion(struct dio *dio)
432 bio = dio_await_one(dio);
434 dio_bio_complete(dio, bio);
439 * A really large O_DIRECT read or write can generate a lot of BIOs. So
440 * to keep the memory consumption sane we periodically reap any completed BIOs
441 * during the BIO generation phase.
443 * This also helps to limit the peak amount of pinned userspace memory.
445 static int dio_bio_reap(struct dio *dio)
449 if (dio->reap_counter++ >= 64) {
450 while (dio->bio_list) {
455 spin_lock_irqsave(&dio->bio_lock, flags);
457 dio->bio_list = bio->bi_private;
458 spin_unlock_irqrestore(&dio->bio_lock, flags);
459 ret2 = dio_bio_complete(dio, bio);
463 dio->reap_counter = 0;
469 * Call into the fs to map some more disk blocks. We record the current number
470 * of available blocks at dio->blocks_available. These are in units of the
471 * fs blocksize, (1 << inode->i_blkbits).
473 * The fs is allowed to map lots of blocks at once. If it wants to do that,
474 * it uses the passed inode-relative block number as the file offset, as usual.
476 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
477 * has remaining to do. The fs should not map more than this number of blocks.
479 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
480 * indicate how much contiguous disk space has been made available at
483 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
484 * This isn't very efficient...
486 * In the case of filesystem holes: the fs may return an arbitrarily-large
487 * hole by returning an appropriate value in b_size and by clearing
488 * buffer_mapped(). However the direct-io code will only process holes one
489 * block at a time - it will repeatedly call get_block() as it walks the hole.
491 static int get_more_blocks(struct dio *dio)
494 struct buffer_head *map_bh = &dio->map_bh;
495 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
496 unsigned long fs_count; /* Number of filesystem-sized blocks */
497 unsigned long dio_count;/* Number of dio_block-sized blocks */
498 unsigned long blkmask;
502 * If there was a memory error and we've overwritten all the
503 * mapped blocks then we can now return that memory error
505 ret = dio->page_errors;
507 BUG_ON(dio->block_in_file >= dio->final_block_in_request);
508 fs_startblk = dio->block_in_file >> dio->blkfactor;
509 dio_count = dio->final_block_in_request - dio->block_in_file;
510 fs_count = dio_count >> dio->blkfactor;
511 blkmask = (1 << dio->blkfactor) - 1;
512 if (dio_count & blkmask)
516 map_bh->b_size = fs_count << dio->inode->i_blkbits;
519 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we
520 * forbid block creations: only overwrites are permitted.
521 * We will return early to the caller once we see an
522 * unmapped buffer head returned, and the caller will fall
523 * back to buffered I/O.
525 * Otherwise the decision is left to the get_blocks method,
526 * which may decide to handle it or also return an unmapped
529 create = dio->rw & WRITE;
530 if (dio->flags & DIO_SKIP_HOLES) {
531 if (dio->block_in_file < (i_size_read(dio->inode) >>
536 ret = (*dio->get_block)(dio->inode, fs_startblk,
543 * There is no bio. Make one now.
545 static int dio_new_bio(struct dio *dio, sector_t start_sector)
550 ret = dio_bio_reap(dio);
553 sector = start_sector << (dio->blkbits - 9);
554 nr_pages = min(dio->pages_in_io, bio_get_nr_vecs(dio->map_bh.b_bdev));
555 BUG_ON(nr_pages <= 0);
556 ret = dio_bio_alloc(dio, dio->map_bh.b_bdev, sector, nr_pages);
563 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
564 * that was successful then update final_block_in_bio and take a ref against
565 * the just-added page.
567 * Return zero on success. Non-zero means the caller needs to start a new BIO.
569 static int dio_bio_add_page(struct dio *dio)
573 ret = bio_add_page(dio->bio, dio->cur_page,
574 dio->cur_page_len, dio->cur_page_offset);
575 if (ret == dio->cur_page_len) {
577 * Decrement count only, if we are done with this page
579 if ((dio->cur_page_len + dio->cur_page_offset) == PAGE_SIZE)
581 page_cache_get(dio->cur_page);
582 dio->final_block_in_bio = dio->cur_page_block +
583 (dio->cur_page_len >> dio->blkbits);
592 * Put cur_page under IO. The section of cur_page which is described by
593 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
594 * starts on-disk at cur_page_block.
596 * We take a ref against the page here (on behalf of its presence in the bio).
598 * The caller of this function is responsible for removing cur_page from the
599 * dio, and for dropping the refcount which came from that presence.
601 static int dio_send_cur_page(struct dio *dio)
607 * See whether this new request is contiguous with the old
609 if (dio->final_block_in_bio != dio->cur_page_block)
612 * Submit now if the underlying fs is about to perform a
619 if (dio->bio == NULL) {
620 ret = dio_new_bio(dio, dio->cur_page_block);
625 if (dio_bio_add_page(dio) != 0) {
627 ret = dio_new_bio(dio, dio->cur_page_block);
629 ret = dio_bio_add_page(dio);
638 * An autonomous function to put a chunk of a page under deferred IO.
640 * The caller doesn't actually know (or care) whether this piece of page is in
641 * a BIO, or is under IO or whatever. We just take care of all possible
642 * situations here. The separation between the logic of do_direct_IO() and
643 * that of submit_page_section() is important for clarity. Please don't break.
645 * The chunk of page starts on-disk at blocknr.
647 * We perform deferred IO, by recording the last-submitted page inside our
648 * private part of the dio structure. If possible, we just expand the IO
649 * across that page here.
651 * If that doesn't work out then we put the old page into the bio and add this
652 * page to the dio instead.
655 submit_page_section(struct dio *dio, struct page *page,
656 unsigned offset, unsigned len, sector_t blocknr)
660 if (dio->rw & WRITE) {
662 * Read accounting is performed in submit_bio()
664 task_io_account_write(len);
668 * Can we just grow the current page's presence in the dio?
670 if ( (dio->cur_page == page) &&
671 (dio->cur_page_offset + dio->cur_page_len == offset) &&
672 (dio->cur_page_block +
673 (dio->cur_page_len >> dio->blkbits) == blocknr)) {
674 dio->cur_page_len += len;
677 * If dio->boundary then we want to schedule the IO now to
678 * avoid metadata seeks.
681 ret = dio_send_cur_page(dio);
682 page_cache_release(dio->cur_page);
683 dio->cur_page = NULL;
689 * If there's a deferred page already there then send it.
692 ret = dio_send_cur_page(dio);
693 page_cache_release(dio->cur_page);
694 dio->cur_page = NULL;
699 page_cache_get(page); /* It is in dio */
700 dio->cur_page = page;
701 dio->cur_page_offset = offset;
702 dio->cur_page_len = len;
703 dio->cur_page_block = blocknr;
709 * Clean any dirty buffers in the blockdev mapping which alias newly-created
710 * file blocks. Only called for S_ISREG files - blockdevs do not set
713 static void clean_blockdev_aliases(struct dio *dio)
718 nblocks = dio->map_bh.b_size >> dio->inode->i_blkbits;
720 for (i = 0; i < nblocks; i++) {
721 unmap_underlying_metadata(dio->map_bh.b_bdev,
722 dio->map_bh.b_blocknr + i);
727 * If we are not writing the entire block and get_block() allocated
728 * the block for us, we need to fill-in the unused portion of the
729 * block with zeros. This happens only if user-buffer, fileoffset or
730 * io length is not filesystem block-size multiple.
732 * `end' is zero if we're doing the start of the IO, 1 at the end of the
735 static void dio_zero_block(struct dio *dio, int end)
737 unsigned dio_blocks_per_fs_block;
738 unsigned this_chunk_blocks; /* In dio_blocks */
739 unsigned this_chunk_bytes;
742 dio->start_zero_done = 1;
743 if (!dio->blkfactor || !buffer_new(&dio->map_bh))
746 dio_blocks_per_fs_block = 1 << dio->blkfactor;
747 this_chunk_blocks = dio->block_in_file & (dio_blocks_per_fs_block - 1);
749 if (!this_chunk_blocks)
753 * We need to zero out part of an fs block. It is either at the
754 * beginning or the end of the fs block.
757 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
759 this_chunk_bytes = this_chunk_blocks << dio->blkbits;
762 if (submit_page_section(dio, page, 0, this_chunk_bytes,
763 dio->next_block_for_io))
766 dio->next_block_for_io += this_chunk_blocks;
770 * Walk the user pages, and the file, mapping blocks to disk and generating
771 * a sequence of (page,offset,len,block) mappings. These mappings are injected
772 * into submit_page_section(), which takes care of the next stage of submission
774 * Direct IO against a blockdev is different from a file. Because we can
775 * happily perform page-sized but 512-byte aligned IOs. It is important that
776 * blockdev IO be able to have fine alignment and large sizes.
778 * So what we do is to permit the ->get_block function to populate bh.b_size
779 * with the size of IO which is permitted at this offset and this i_blkbits.
781 * For best results, the blockdev should be set up with 512-byte i_blkbits and
782 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
783 * fine alignment but still allows this function to work in PAGE_SIZE units.
785 static int do_direct_IO(struct dio *dio)
787 const unsigned blkbits = dio->blkbits;
788 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
790 unsigned block_in_page;
791 struct buffer_head *map_bh = &dio->map_bh;
794 /* The I/O can start at any block offset within the first page */
795 block_in_page = dio->first_block_in_page;
797 while (dio->block_in_file < dio->final_block_in_request) {
798 page = dio_get_page(dio);
804 while (block_in_page < blocks_per_page) {
805 unsigned offset_in_page = block_in_page << blkbits;
806 unsigned this_chunk_bytes; /* # of bytes mapped */
807 unsigned this_chunk_blocks; /* # of blocks */
810 if (dio->blocks_available == 0) {
812 * Need to go and map some more disk
814 unsigned long blkmask;
815 unsigned long dio_remainder;
817 ret = get_more_blocks(dio);
819 page_cache_release(page);
822 if (!buffer_mapped(map_bh))
825 dio->blocks_available =
826 map_bh->b_size >> dio->blkbits;
827 dio->next_block_for_io =
828 map_bh->b_blocknr << dio->blkfactor;
829 if (buffer_new(map_bh))
830 clean_blockdev_aliases(dio);
835 blkmask = (1 << dio->blkfactor) - 1;
836 dio_remainder = (dio->block_in_file & blkmask);
839 * If we are at the start of IO and that IO
840 * starts partway into a fs-block,
841 * dio_remainder will be non-zero. If the IO
842 * is a read then we can simply advance the IO
843 * cursor to the first block which is to be
844 * read. But if the IO is a write and the
845 * block was newly allocated we cannot do that;
846 * the start of the fs block must be zeroed out
849 if (!buffer_new(map_bh))
850 dio->next_block_for_io += dio_remainder;
851 dio->blocks_available -= dio_remainder;
855 if (!buffer_mapped(map_bh)) {
856 loff_t i_size_aligned;
858 /* AKPM: eargh, -ENOTBLK is a hack */
859 if (dio->rw & WRITE) {
860 page_cache_release(page);
865 * Be sure to account for a partial block as the
866 * last block in the file
868 i_size_aligned = ALIGN(i_size_read(dio->inode),
870 if (dio->block_in_file >=
871 i_size_aligned >> blkbits) {
873 page_cache_release(page);
876 zero_user(page, block_in_page << blkbits,
878 dio->block_in_file++;
884 * If we're performing IO which has an alignment which
885 * is finer than the underlying fs, go check to see if
886 * we must zero out the start of this block.
888 if (unlikely(dio->blkfactor && !dio->start_zero_done))
889 dio_zero_block(dio, 0);
892 * Work out, in this_chunk_blocks, how much disk we
893 * can add to this page
895 this_chunk_blocks = dio->blocks_available;
896 u = (PAGE_SIZE - offset_in_page) >> blkbits;
897 if (this_chunk_blocks > u)
898 this_chunk_blocks = u;
899 u = dio->final_block_in_request - dio->block_in_file;
900 if (this_chunk_blocks > u)
901 this_chunk_blocks = u;
902 this_chunk_bytes = this_chunk_blocks << blkbits;
903 BUG_ON(this_chunk_bytes == 0);
905 dio->boundary = buffer_boundary(map_bh);
906 ret = submit_page_section(dio, page, offset_in_page,
907 this_chunk_bytes, dio->next_block_for_io);
909 page_cache_release(page);
912 dio->next_block_for_io += this_chunk_blocks;
914 dio->block_in_file += this_chunk_blocks;
915 block_in_page += this_chunk_blocks;
916 dio->blocks_available -= this_chunk_blocks;
918 BUG_ON(dio->block_in_file > dio->final_block_in_request);
919 if (dio->block_in_file == dio->final_block_in_request)
923 /* Drop the ref which was taken in get_user_pages() */
924 page_cache_release(page);
932 * Releases both i_mutex and i_alloc_sem
935 direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
936 const struct iovec *iov, loff_t offset, unsigned long nr_segs,
937 unsigned blkbits, get_block_t get_block, dio_iodone_t end_io,
940 unsigned long user_addr;
949 dio->blkbits = blkbits;
950 dio->blkfactor = inode->i_blkbits - blkbits;
951 dio->block_in_file = offset >> blkbits;
953 dio->get_block = get_block;
954 dio->end_io = end_io;
955 dio->final_block_in_bio = -1;
956 dio->next_block_for_io = -1;
959 dio->i_size = i_size_read(inode);
961 spin_lock_init(&dio->bio_lock);
965 * In case of non-aligned buffers, we may need 2 more
966 * pages since we need to zero out first and last block.
968 if (unlikely(dio->blkfactor))
969 dio->pages_in_io = 2;
971 for (seg = 0; seg < nr_segs; seg++) {
972 user_addr = (unsigned long)iov[seg].iov_base;
974 ((user_addr+iov[seg].iov_len +PAGE_SIZE-1)/PAGE_SIZE
975 - user_addr/PAGE_SIZE);
978 for (seg = 0; seg < nr_segs; seg++) {
979 user_addr = (unsigned long)iov[seg].iov_base;
980 dio->size += bytes = iov[seg].iov_len;
982 /* Index into the first page of the first block */
983 dio->first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
984 dio->final_block_in_request = dio->block_in_file +
986 /* Page fetching state */
991 dio->total_pages = 0;
992 if (user_addr & (PAGE_SIZE-1)) {
994 bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
996 dio->total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
997 dio->curr_user_address = user_addr;
999 ret = do_direct_IO(dio);
1001 dio->result += iov[seg].iov_len -
1002 ((dio->final_block_in_request - dio->block_in_file) <<
1009 } /* end iovec loop */
1011 if (ret == -ENOTBLK && (rw & WRITE)) {
1013 * The remaining part of the request will be
1014 * be handled by buffered I/O when we return
1019 * There may be some unwritten disk at the end of a part-written
1020 * fs-block-sized block. Go zero that now.
1022 dio_zero_block(dio, 1);
1024 if (dio->cur_page) {
1025 ret2 = dio_send_cur_page(dio);
1028 page_cache_release(dio->cur_page);
1029 dio->cur_page = NULL;
1032 dio_bio_submit(dio);
1035 * It is possible that, we return short IO due to end of file.
1036 * In that case, we need to release all the pages we got hold on.
1041 * All block lookups have been performed. For READ requests
1042 * we can let i_mutex go now that its achieved its purpose
1043 * of protecting us from looking up uninitialized blocks.
1045 if (rw == READ && (dio->flags & DIO_LOCKING))
1046 mutex_unlock(&dio->inode->i_mutex);
1049 * The only time we want to leave bios in flight is when a successful
1050 * partial aio read or full aio write have been setup. In that case
1051 * bio completion will call aio_complete. The only time it's safe to
1052 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1053 * This had *better* be the only place that raises -EIOCBQUEUED.
1055 BUG_ON(ret == -EIOCBQUEUED);
1056 if (dio->is_async && ret == 0 && dio->result &&
1057 ((rw & READ) || (dio->result == dio->size)))
1060 if (ret != -EIOCBQUEUED) {
1061 /* All IO is now issued, send it on its way */
1062 blk_run_address_space(inode->i_mapping);
1063 dio_await_completion(dio);
1067 * Sync will always be dropping the final ref and completing the
1068 * operation. AIO can if it was a broken operation described above or
1069 * in fact if all the bios race to complete before we get here. In
1070 * that case dio_complete() translates the EIOCBQUEUED into the proper
1071 * return code that the caller will hand to aio_complete().
1073 * This is managed by the bio_lock instead of being an atomic_t so that
1074 * completion paths can drop their ref and use the remaining count to
1075 * decide to wake the submission path atomically.
1077 spin_lock_irqsave(&dio->bio_lock, flags);
1078 ret2 = --dio->refcount;
1079 spin_unlock_irqrestore(&dio->bio_lock, flags);
1082 ret = dio_complete(dio, offset, ret);
1085 BUG_ON(ret != -EIOCBQUEUED);
1091 * This is a library function for use by filesystem drivers.
1093 * The locking rules are governed by the flags parameter:
1094 * - if the flags value contains DIO_LOCKING we use a fancy locking
1095 * scheme for dumb filesystems.
1096 * For writes this function is called under i_mutex and returns with
1097 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1098 * taken and dropped again before returning.
1099 * For reads and writes i_alloc_sem is taken in shared mode and released
1100 * on I/O completion (which may happen asynchronously after returning to
1103 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1104 * internal locking but rather rely on the filesystem to synchronize
1105 * direct I/O reads/writes versus each other and truncate.
1106 * For reads and writes both i_mutex and i_alloc_sem are not held on
1107 * entry and are never taken.
1110 __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1111 struct block_device *bdev, const struct iovec *iov, loff_t offset,
1112 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1118 unsigned blkbits = inode->i_blkbits;
1119 unsigned bdev_blkbits = 0;
1120 unsigned blocksize_mask = (1 << blkbits) - 1;
1121 ssize_t retval = -EINVAL;
1122 loff_t end = offset;
1126 rw = WRITE_ODIRECT_PLUG;
1129 bdev_blkbits = blksize_bits(bdev_logical_block_size(bdev));
1131 if (offset & blocksize_mask) {
1133 blkbits = bdev_blkbits;
1134 blocksize_mask = (1 << blkbits) - 1;
1135 if (offset & blocksize_mask)
1139 /* Check the memory alignment. Blocks cannot straddle pages */
1140 for (seg = 0; seg < nr_segs; seg++) {
1141 addr = (unsigned long)iov[seg].iov_base;
1142 size = iov[seg].iov_len;
1144 if ((addr & blocksize_mask) || (size & blocksize_mask)) {
1146 blkbits = bdev_blkbits;
1147 blocksize_mask = (1 << blkbits) - 1;
1148 if ((addr & blocksize_mask) || (size & blocksize_mask))
1153 dio = kmalloc(sizeof(*dio), GFP_KERNEL);
1158 * Believe it or not, zeroing out the page array caused a .5%
1159 * performance regression in a database benchmark. So, we take
1160 * care to only zero out what's needed.
1162 memset(dio, 0, offsetof(struct dio, pages));
1165 if (dio->flags & DIO_LOCKING) {
1166 /* watch out for a 0 len io from a tricksy fs */
1167 if (rw == READ && end > offset) {
1168 struct address_space *mapping =
1169 iocb->ki_filp->f_mapping;
1171 /* will be released by direct_io_worker */
1172 mutex_lock(&inode->i_mutex);
1174 retval = filemap_write_and_wait_range(mapping, offset,
1177 mutex_unlock(&inode->i_mutex);
1184 * Will be released at I/O completion, possibly in a
1187 down_read_non_owner(&inode->i_alloc_sem);
1191 * For file extending writes updating i_size before data
1192 * writeouts complete can expose uninitialized blocks. So
1193 * even for AIO, we need to wait for i/o to complete before
1194 * returning in this case.
1196 dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
1197 (end > i_size_read(inode)));
1199 retval = direct_io_worker(rw, iocb, inode, iov, offset,
1200 nr_segs, blkbits, get_block, end_io, dio);
1203 * In case of error extending write may have instantiated a few
1204 * blocks outside i_size. Trim these off again for DIO_LOCKING.
1206 * NOTE: filesystems with their own locking have to handle this
1209 if (dio->flags & DIO_LOCKING) {
1210 if (unlikely((rw & WRITE) && retval < 0)) {
1211 loff_t isize = i_size_read(inode);
1213 vmtruncate(inode, isize);
1220 EXPORT_SYMBOL(__blockdev_direct_IO);