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>
40 #include <linux/aio.h>
43 * How many user pages to map in one call to get_user_pages(). This determines
44 * the size of a structure in the slab cache
49 * This code generally works in units of "dio_blocks". A dio_block is
50 * somewhere between the hard sector size and the filesystem block size. it
51 * is determined on a per-invocation basis. When talking to the filesystem
52 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
53 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
54 * to bio_block quantities by shifting left by blkfactor.
56 * If blkfactor is zero then the user's request was aligned to the filesystem's
60 /* dio_state only used in the submission path */
63 struct bio *bio; /* bio under assembly */
64 unsigned blkbits; /* doesn't change */
65 unsigned blkfactor; /* When we're using an alignment which
66 is finer than the filesystem's soft
67 blocksize, this specifies how much
68 finer. blkfactor=2 means 1/4-block
69 alignment. Does not change */
70 unsigned start_zero_done; /* flag: sub-blocksize zeroing has
71 been performed at the start of a
73 int pages_in_io; /* approximate total IO pages */
74 size_t size; /* total request size (doesn't change)*/
75 sector_t block_in_file; /* Current offset into the underlying
76 file in dio_block units. */
77 unsigned blocks_available; /* At block_in_file. changes */
78 int reap_counter; /* rate limit reaping */
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 get_block_t *get_block; /* block mapping function */
83 dio_submit_t *submit_io; /* IO submition function */
85 loff_t logical_offset_in_bio; /* current first logical block in bio */
86 sector_t final_block_in_bio; /* current final block in bio + 1 */
87 sector_t next_block_for_io; /* next block to be put under IO,
88 in dio_blocks units */
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 */
99 loff_t cur_page_fs_offset; /* Offset in file */
102 * Page fetching state. These variables belong to dio_refill_pages().
104 int curr_page; /* changes */
105 int total_pages; /* doesn't change */
106 unsigned long curr_user_address;/* changes */
109 * Page queue. These variables belong to dio_refill_pages() and
112 unsigned head; /* next page to process */
113 unsigned tail; /* last valid page + 1 */
116 /* dio_state communicated between submission path and end_io */
118 int flags; /* doesn't change */
121 loff_t i_size; /* i_size when submitted */
122 dio_iodone_t *end_io; /* IO completion function */
124 void *private; /* copy from map_bh.b_private */
126 /* BIO completion state */
127 spinlock_t bio_lock; /* protects BIO fields below */
128 int page_errors; /* errno from get_user_pages() */
129 int is_async; /* is IO async ? */
130 int should_dirty; /* should we mark read pages dirty? */
131 bool defer_completion; /* defer AIO completion to workqueue? */
132 int io_error; /* IO error in completion path */
133 unsigned long refcount; /* direct_io_worker() and bios */
134 struct bio *bio_list; /* singly linked via bi_private */
135 struct task_struct *waiter; /* waiting task (NULL if none) */
137 /* AIO related stuff */
138 struct kiocb *iocb; /* kiocb */
139 ssize_t result; /* IO result */
142 * pages[] (and any fields placed after it) are not zeroed out at
143 * allocation time. Don't add new fields after pages[] unless you
144 * wish that they not be zeroed.
147 struct page *pages[DIO_PAGES]; /* page buffer */
148 struct work_struct complete_work;/* deferred AIO completion */
150 } ____cacheline_aligned_in_smp;
152 static struct kmem_cache *dio_cache __read_mostly;
155 * How many pages are in the queue?
157 static inline unsigned dio_pages_present(struct dio_submit *sdio)
159 return sdio->tail - sdio->head;
163 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
165 static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
170 nr_pages = min(sdio->total_pages - sdio->curr_page, DIO_PAGES);
171 ret = get_user_pages_fast(
172 sdio->curr_user_address, /* Where from? */
173 nr_pages, /* How many pages? */
174 dio->rw == READ, /* Write to memory? */
175 &dio->pages[0]); /* Put results here */
177 if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) {
178 struct page *page = ZERO_PAGE(0);
180 * A memory fault, but the filesystem has some outstanding
181 * mapped blocks. We need to use those blocks up to avoid
182 * leaking stale data in the file.
184 if (dio->page_errors == 0)
185 dio->page_errors = ret;
186 page_cache_get(page);
187 dio->pages[0] = page;
195 sdio->curr_user_address += ret * PAGE_SIZE;
196 sdio->curr_page += ret;
206 * Get another userspace page. Returns an ERR_PTR on error. Pages are
207 * buffered inside the dio so that we can call get_user_pages() against a
208 * decent number of pages, less frequently. To provide nicer use of the
211 static inline struct page *dio_get_page(struct dio *dio,
212 struct dio_submit *sdio)
214 if (dio_pages_present(sdio) == 0) {
217 ret = dio_refill_pages(dio, sdio);
220 BUG_ON(dio_pages_present(sdio) == 0);
222 return dio->pages[sdio->head++];
226 * dio_complete() - called when all DIO BIO I/O has been completed
227 * @offset: the byte offset in the file of the completed operation
229 * This drops i_dio_count, lets interested parties know that a DIO operation
230 * has completed, and calculates the resulting return code for the operation.
232 * It lets the filesystem know if it registered an interest earlier via
233 * get_block. Pass the private field of the map buffer_head so that
234 * filesystems can use it to hold additional state between get_block calls and
237 static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret,
240 ssize_t transferred = 0;
243 * AIO submission can race with bio completion to get here while
244 * expecting to have the last io completed by bio completion.
245 * In that case -EIOCBQUEUED is in fact not an error we want
246 * to preserve through this call.
248 if (ret == -EIOCBQUEUED)
252 transferred = dio->result;
254 /* Check for short read case */
255 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
256 transferred = dio->i_size - offset;
260 ret = dio->page_errors;
266 if (dio->end_io && dio->result)
267 dio->end_io(dio->iocb, offset, transferred, dio->private);
269 inode_dio_done(dio->inode);
271 if (dio->rw & WRITE) {
274 err = generic_write_sync(dio->iocb->ki_filp, offset,
276 if (err < 0 && ret > 0)
280 aio_complete(dio->iocb, ret, 0);
283 kmem_cache_free(dio_cache, dio);
287 static void dio_aio_complete_work(struct work_struct *work)
289 struct dio *dio = container_of(work, struct dio, complete_work);
291 dio_complete(dio, dio->iocb->ki_pos, 0, true);
294 static int dio_bio_complete(struct dio *dio, struct bio *bio);
297 * Asynchronous IO callback.
299 static void dio_bio_end_aio(struct bio *bio, int error)
301 struct dio *dio = bio->bi_private;
302 unsigned long remaining;
305 /* cleanup the bio */
306 dio_bio_complete(dio, bio);
308 spin_lock_irqsave(&dio->bio_lock, flags);
309 remaining = --dio->refcount;
310 if (remaining == 1 && dio->waiter)
311 wake_up_process(dio->waiter);
312 spin_unlock_irqrestore(&dio->bio_lock, flags);
314 if (remaining == 0) {
315 if (dio->result && dio->defer_completion) {
316 INIT_WORK(&dio->complete_work, dio_aio_complete_work);
317 queue_work(dio->inode->i_sb->s_dio_done_wq,
318 &dio->complete_work);
320 dio_complete(dio, dio->iocb->ki_pos, 0, true);
326 * The BIO completion handler simply queues the BIO up for the process-context
329 * During I/O bi_private points at the dio. After I/O, bi_private is used to
330 * implement a singly-linked list of completed BIOs, at dio->bio_list.
332 static void dio_bio_end_io(struct bio *bio, int error)
334 struct dio *dio = bio->bi_private;
337 spin_lock_irqsave(&dio->bio_lock, flags);
338 bio->bi_private = dio->bio_list;
340 if (--dio->refcount == 1 && dio->waiter)
341 wake_up_process(dio->waiter);
342 spin_unlock_irqrestore(&dio->bio_lock, flags);
346 * dio_end_io - handle the end io action for the given bio
347 * @bio: The direct io bio thats being completed
348 * @error: Error if there was one
350 * This is meant to be called by any filesystem that uses their own dio_submit_t
351 * so that the DIO specific endio actions are dealt with after the filesystem
352 * has done it's completion work.
354 void dio_end_io(struct bio *bio, int error)
356 struct dio *dio = bio->bi_private;
359 dio_bio_end_aio(bio, error);
361 dio_bio_end_io(bio, error);
363 EXPORT_SYMBOL_GPL(dio_end_io);
366 dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
367 struct block_device *bdev,
368 sector_t first_sector, int nr_vecs)
373 * bio_alloc() is guaranteed to return a bio when called with
374 * __GFP_WAIT and we request a valid number of vectors.
376 bio = bio_alloc(GFP_KERNEL, nr_vecs);
379 bio->bi_sector = first_sector;
381 bio->bi_end_io = dio_bio_end_aio;
383 bio->bi_end_io = dio_bio_end_io;
386 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
390 * In the AIO read case we speculatively dirty the pages before starting IO.
391 * During IO completion, any of these pages which happen to have been written
392 * back will be redirtied by bio_check_pages_dirty().
394 * bios hold a dio reference between submit_bio and ->end_io.
396 static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
398 struct bio *bio = sdio->bio;
401 bio->bi_private = dio;
403 spin_lock_irqsave(&dio->bio_lock, flags);
405 spin_unlock_irqrestore(&dio->bio_lock, flags);
407 if (dio->is_async && dio->rw == READ && dio->should_dirty)
408 bio_set_pages_dirty(bio);
411 sdio->submit_io(dio->rw, bio, dio->inode,
412 sdio->logical_offset_in_bio);
414 submit_bio(dio->rw, bio);
418 sdio->logical_offset_in_bio = 0;
422 * Release any resources in case of a failure
424 static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
426 while (dio_pages_present(sdio))
427 page_cache_release(dio_get_page(dio, sdio));
431 * Wait for the next BIO to complete. Remove it and return it. NULL is
432 * returned once all BIOs have been completed. This must only be called once
433 * all bios have been issued so that dio->refcount can only decrease. This
434 * requires that that the caller hold a reference on the dio.
436 static struct bio *dio_await_one(struct dio *dio)
439 struct bio *bio = NULL;
441 spin_lock_irqsave(&dio->bio_lock, flags);
444 * Wait as long as the list is empty and there are bios in flight. bio
445 * completion drops the count, maybe adds to the list, and wakes while
446 * holding the bio_lock so we don't need set_current_state()'s barrier
447 * and can call it after testing our condition.
449 while (dio->refcount > 1 && dio->bio_list == NULL) {
450 __set_current_state(TASK_UNINTERRUPTIBLE);
451 dio->waiter = current;
452 spin_unlock_irqrestore(&dio->bio_lock, flags);
454 /* wake up sets us TASK_RUNNING */
455 spin_lock_irqsave(&dio->bio_lock, flags);
460 dio->bio_list = bio->bi_private;
462 spin_unlock_irqrestore(&dio->bio_lock, flags);
467 * Process one completed BIO. No locks are held.
469 static int dio_bio_complete(struct dio *dio, struct bio *bio)
471 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
472 struct bio_vec *bvec;
476 dio->io_error = -EIO;
478 if (dio->is_async && dio->rw == READ && dio->should_dirty) {
479 bio_check_pages_dirty(bio); /* transfers ownership */
481 bio_for_each_segment_all(bvec, bio, i) {
482 struct page *page = bvec->bv_page;
484 if (dio->rw == READ && !PageCompound(page) &&
486 set_page_dirty_lock(page);
487 page_cache_release(page);
491 return uptodate ? 0 : -EIO;
495 * Wait on and process all in-flight BIOs. This must only be called once
496 * all bios have been issued so that the refcount can only decrease.
497 * This just waits for all bios to make it through dio_bio_complete. IO
498 * errors are propagated through dio->io_error and should be propagated via
501 static void dio_await_completion(struct dio *dio)
505 bio = dio_await_one(dio);
507 dio_bio_complete(dio, bio);
512 * A really large O_DIRECT read or write can generate a lot of BIOs. So
513 * to keep the memory consumption sane we periodically reap any completed BIOs
514 * during the BIO generation phase.
516 * This also helps to limit the peak amount of pinned userspace memory.
518 static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
522 if (sdio->reap_counter++ >= 64) {
523 while (dio->bio_list) {
528 spin_lock_irqsave(&dio->bio_lock, flags);
530 dio->bio_list = bio->bi_private;
531 spin_unlock_irqrestore(&dio->bio_lock, flags);
532 ret2 = dio_bio_complete(dio, bio);
536 sdio->reap_counter = 0;
542 * Create workqueue for deferred direct IO completions. We allocate the
543 * workqueue when it's first needed. This avoids creating workqueue for
544 * filesystems that don't need it and also allows us to create the workqueue
545 * late enough so the we can include s_id in the name of the workqueue.
547 static int sb_init_dio_done_wq(struct super_block *sb)
549 struct workqueue_struct *old;
550 struct workqueue_struct *wq = alloc_workqueue("dio/%s",
556 * This has to be atomic as more DIOs can race to create the workqueue
558 old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
559 /* Someone created workqueue before us? Free ours... */
561 destroy_workqueue(wq);
565 static int dio_set_defer_completion(struct dio *dio)
567 struct super_block *sb = dio->inode->i_sb;
569 if (dio->defer_completion)
571 dio->defer_completion = true;
572 if (!sb->s_dio_done_wq)
573 return sb_init_dio_done_wq(sb);
578 * Call into the fs to map some more disk blocks. We record the current number
579 * of available blocks at sdio->blocks_available. These are in units of the
580 * fs blocksize, (1 << inode->i_blkbits).
582 * The fs is allowed to map lots of blocks at once. If it wants to do that,
583 * it uses the passed inode-relative block number as the file offset, as usual.
585 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
586 * has remaining to do. The fs should not map more than this number of blocks.
588 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
589 * indicate how much contiguous disk space has been made available at
592 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
593 * This isn't very efficient...
595 * In the case of filesystem holes: the fs may return an arbitrarily-large
596 * hole by returning an appropriate value in b_size and by clearing
597 * buffer_mapped(). However the direct-io code will only process holes one
598 * block at a time - it will repeatedly call get_block() as it walks the hole.
600 static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
601 struct buffer_head *map_bh)
604 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
605 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
606 unsigned long fs_count; /* Number of filesystem-sized blocks */
608 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
611 * If there was a memory error and we've overwritten all the
612 * mapped blocks then we can now return that memory error
614 ret = dio->page_errors;
616 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
617 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
618 fs_endblk = (sdio->final_block_in_request - 1) >>
620 fs_count = fs_endblk - fs_startblk + 1;
623 map_bh->b_size = fs_count << i_blkbits;
626 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we
627 * forbid block creations: only overwrites are permitted.
628 * We will return early to the caller once we see an
629 * unmapped buffer head returned, and the caller will fall
630 * back to buffered I/O.
632 * Otherwise the decision is left to the get_blocks method,
633 * which may decide to handle it or also return an unmapped
636 create = dio->rw & WRITE;
637 if (dio->flags & DIO_SKIP_HOLES) {
638 if (sdio->block_in_file < (i_size_read(dio->inode) >>
643 ret = (*sdio->get_block)(dio->inode, fs_startblk,
646 /* Store for completion */
647 dio->private = map_bh->b_private;
649 if (ret == 0 && buffer_defer_completion(map_bh))
650 ret = dio_set_defer_completion(dio);
656 * There is no bio. Make one now.
658 static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
659 sector_t start_sector, struct buffer_head *map_bh)
664 ret = dio_bio_reap(dio, sdio);
667 sector = start_sector << (sdio->blkbits - 9);
668 nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(map_bh->b_bdev));
669 nr_pages = min(nr_pages, BIO_MAX_PAGES);
670 BUG_ON(nr_pages <= 0);
671 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
678 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
679 * that was successful then update final_block_in_bio and take a ref against
680 * the just-added page.
682 * Return zero on success. Non-zero means the caller needs to start a new BIO.
684 static inline int dio_bio_add_page(struct dio_submit *sdio)
688 ret = bio_add_page(sdio->bio, sdio->cur_page,
689 sdio->cur_page_len, sdio->cur_page_offset);
690 if (ret == sdio->cur_page_len) {
692 * Decrement count only, if we are done with this page
694 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
696 page_cache_get(sdio->cur_page);
697 sdio->final_block_in_bio = sdio->cur_page_block +
698 (sdio->cur_page_len >> sdio->blkbits);
707 * Put cur_page under IO. The section of cur_page which is described by
708 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
709 * starts on-disk at cur_page_block.
711 * We take a ref against the page here (on behalf of its presence in the bio).
713 * The caller of this function is responsible for removing cur_page from the
714 * dio, and for dropping the refcount which came from that presence.
716 static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
717 struct buffer_head *map_bh)
722 loff_t cur_offset = sdio->cur_page_fs_offset;
723 loff_t bio_next_offset = sdio->logical_offset_in_bio +
727 * See whether this new request is contiguous with the old.
729 * Btrfs cannot handle having logically non-contiguous requests
730 * submitted. For example if you have
732 * Logical: [0-4095][HOLE][8192-12287]
733 * Physical: [0-4095] [4096-8191]
735 * We cannot submit those pages together as one BIO. So if our
736 * current logical offset in the file does not equal what would
737 * be the next logical offset in the bio, submit the bio we
740 if (sdio->final_block_in_bio != sdio->cur_page_block ||
741 cur_offset != bio_next_offset)
742 dio_bio_submit(dio, sdio);
745 if (sdio->bio == NULL) {
746 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
751 if (dio_bio_add_page(sdio) != 0) {
752 dio_bio_submit(dio, sdio);
753 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
755 ret = dio_bio_add_page(sdio);
764 * An autonomous function to put a chunk of a page under deferred IO.
766 * The caller doesn't actually know (or care) whether this piece of page is in
767 * a BIO, or is under IO or whatever. We just take care of all possible
768 * situations here. The separation between the logic of do_direct_IO() and
769 * that of submit_page_section() is important for clarity. Please don't break.
771 * The chunk of page starts on-disk at blocknr.
773 * We perform deferred IO, by recording the last-submitted page inside our
774 * private part of the dio structure. If possible, we just expand the IO
775 * across that page here.
777 * If that doesn't work out then we put the old page into the bio and add this
778 * page to the dio instead.
781 submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
782 unsigned offset, unsigned len, sector_t blocknr,
783 struct buffer_head *map_bh)
787 if (dio->rw & WRITE) {
789 * Read accounting is performed in submit_bio()
791 task_io_account_write(len);
795 * Can we just grow the current page's presence in the dio?
797 if (sdio->cur_page == page &&
798 sdio->cur_page_offset + sdio->cur_page_len == offset &&
799 sdio->cur_page_block +
800 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
801 sdio->cur_page_len += len;
806 * If there's a deferred page already there then send it.
808 if (sdio->cur_page) {
809 ret = dio_send_cur_page(dio, sdio, map_bh);
810 page_cache_release(sdio->cur_page);
811 sdio->cur_page = NULL;
816 page_cache_get(page); /* It is in dio */
817 sdio->cur_page = page;
818 sdio->cur_page_offset = offset;
819 sdio->cur_page_len = len;
820 sdio->cur_page_block = blocknr;
821 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
824 * If sdio->boundary then we want to schedule the IO now to
825 * avoid metadata seeks.
827 if (sdio->boundary) {
828 ret = dio_send_cur_page(dio, sdio, map_bh);
829 dio_bio_submit(dio, sdio);
830 page_cache_release(sdio->cur_page);
831 sdio->cur_page = NULL;
837 * Clean any dirty buffers in the blockdev mapping which alias newly-created
838 * file blocks. Only called for S_ISREG files - blockdevs do not set
841 static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
846 nblocks = map_bh->b_size >> dio->inode->i_blkbits;
848 for (i = 0; i < nblocks; i++) {
849 unmap_underlying_metadata(map_bh->b_bdev,
850 map_bh->b_blocknr + i);
855 * If we are not writing the entire block and get_block() allocated
856 * the block for us, we need to fill-in the unused portion of the
857 * block with zeros. This happens only if user-buffer, fileoffset or
858 * io length is not filesystem block-size multiple.
860 * `end' is zero if we're doing the start of the IO, 1 at the end of the
863 static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
864 int end, struct buffer_head *map_bh)
866 unsigned dio_blocks_per_fs_block;
867 unsigned this_chunk_blocks; /* In dio_blocks */
868 unsigned this_chunk_bytes;
871 sdio->start_zero_done = 1;
872 if (!sdio->blkfactor || !buffer_new(map_bh))
875 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
876 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
878 if (!this_chunk_blocks)
882 * We need to zero out part of an fs block. It is either at the
883 * beginning or the end of the fs block.
886 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
888 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
891 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
892 sdio->next_block_for_io, map_bh))
895 sdio->next_block_for_io += this_chunk_blocks;
899 * Walk the user pages, and the file, mapping blocks to disk and generating
900 * a sequence of (page,offset,len,block) mappings. These mappings are injected
901 * into submit_page_section(), which takes care of the next stage of submission
903 * Direct IO against a blockdev is different from a file. Because we can
904 * happily perform page-sized but 512-byte aligned IOs. It is important that
905 * blockdev IO be able to have fine alignment and large sizes.
907 * So what we do is to permit the ->get_block function to populate bh.b_size
908 * with the size of IO which is permitted at this offset and this i_blkbits.
910 * For best results, the blockdev should be set up with 512-byte i_blkbits and
911 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
912 * fine alignment but still allows this function to work in PAGE_SIZE units.
914 static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
915 struct buffer_head *map_bh)
917 const unsigned blkbits = sdio->blkbits;
918 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
920 unsigned block_in_page;
923 /* The I/O can start at any block offset within the first page */
924 block_in_page = sdio->first_block_in_page;
926 while (sdio->block_in_file < sdio->final_block_in_request) {
927 page = dio_get_page(dio, sdio);
933 while (block_in_page < blocks_per_page) {
934 unsigned offset_in_page = block_in_page << blkbits;
935 unsigned this_chunk_bytes; /* # of bytes mapped */
936 unsigned this_chunk_blocks; /* # of blocks */
939 if (sdio->blocks_available == 0) {
941 * Need to go and map some more disk
943 unsigned long blkmask;
944 unsigned long dio_remainder;
946 ret = get_more_blocks(dio, sdio, map_bh);
948 page_cache_release(page);
951 if (!buffer_mapped(map_bh))
954 sdio->blocks_available =
955 map_bh->b_size >> sdio->blkbits;
956 sdio->next_block_for_io =
957 map_bh->b_blocknr << sdio->blkfactor;
958 if (buffer_new(map_bh))
959 clean_blockdev_aliases(dio, map_bh);
961 if (!sdio->blkfactor)
964 blkmask = (1 << sdio->blkfactor) - 1;
965 dio_remainder = (sdio->block_in_file & blkmask);
968 * If we are at the start of IO and that IO
969 * starts partway into a fs-block,
970 * dio_remainder will be non-zero. If the IO
971 * is a read then we can simply advance the IO
972 * cursor to the first block which is to be
973 * read. But if the IO is a write and the
974 * block was newly allocated we cannot do that;
975 * the start of the fs block must be zeroed out
978 if (!buffer_new(map_bh))
979 sdio->next_block_for_io += dio_remainder;
980 sdio->blocks_available -= dio_remainder;
984 if (!buffer_mapped(map_bh)) {
985 loff_t i_size_aligned;
987 /* AKPM: eargh, -ENOTBLK is a hack */
988 if (dio->rw & WRITE) {
989 page_cache_release(page);
994 * Be sure to account for a partial block as the
995 * last block in the file
997 i_size_aligned = ALIGN(i_size_read(dio->inode),
999 if (sdio->block_in_file >=
1000 i_size_aligned >> blkbits) {
1002 page_cache_release(page);
1005 zero_user(page, block_in_page << blkbits,
1007 sdio->block_in_file++;
1013 * If we're performing IO which has an alignment which
1014 * is finer than the underlying fs, go check to see if
1015 * we must zero out the start of this block.
1017 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
1018 dio_zero_block(dio, sdio, 0, map_bh);
1021 * Work out, in this_chunk_blocks, how much disk we
1022 * can add to this page
1024 this_chunk_blocks = sdio->blocks_available;
1025 u = (PAGE_SIZE - offset_in_page) >> blkbits;
1026 if (this_chunk_blocks > u)
1027 this_chunk_blocks = u;
1028 u = sdio->final_block_in_request - sdio->block_in_file;
1029 if (this_chunk_blocks > u)
1030 this_chunk_blocks = u;
1031 this_chunk_bytes = this_chunk_blocks << blkbits;
1032 BUG_ON(this_chunk_bytes == 0);
1034 if (this_chunk_blocks == sdio->blocks_available)
1035 sdio->boundary = buffer_boundary(map_bh);
1036 ret = submit_page_section(dio, sdio, page,
1039 sdio->next_block_for_io,
1042 page_cache_release(page);
1045 sdio->next_block_for_io += this_chunk_blocks;
1047 sdio->block_in_file += this_chunk_blocks;
1048 block_in_page += this_chunk_blocks;
1049 sdio->blocks_available -= this_chunk_blocks;
1051 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1052 if (sdio->block_in_file == sdio->final_block_in_request)
1056 /* Drop the ref which was taken in get_user_pages() */
1057 page_cache_release(page);
1064 static inline int drop_refcount(struct dio *dio)
1067 unsigned long flags;
1070 * Sync will always be dropping the final ref and completing the
1071 * operation. AIO can if it was a broken operation described above or
1072 * in fact if all the bios race to complete before we get here. In
1073 * that case dio_complete() translates the EIOCBQUEUED into the proper
1074 * return code that the caller will hand to aio_complete().
1076 * This is managed by the bio_lock instead of being an atomic_t so that
1077 * completion paths can drop their ref and use the remaining count to
1078 * decide to wake the submission path atomically.
1080 spin_lock_irqsave(&dio->bio_lock, flags);
1081 ret2 = --dio->refcount;
1082 spin_unlock_irqrestore(&dio->bio_lock, flags);
1086 static ssize_t direct_IO_iovec(const struct iovec *iov, unsigned long nr_segs,
1087 struct dio *dio, struct dio_submit *sdio,
1088 unsigned blkbits, struct buffer_head *map_bh)
1093 unsigned long user_addr;
1095 for (seg = 0; seg < nr_segs; seg++) {
1096 user_addr = (unsigned long)iov[seg].iov_base;
1097 sdio->pages_in_io +=
1098 ((user_addr + iov[seg].iov_len + PAGE_SIZE-1) /
1099 PAGE_SIZE - user_addr / PAGE_SIZE);
1102 dio->should_dirty = 1;
1104 for (seg = 0; seg < nr_segs; seg++) {
1105 user_addr = (unsigned long)iov[seg].iov_base;
1106 sdio->size += bytes = iov[seg].iov_len;
1108 /* Index into the first page of the first block */
1109 sdio->first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
1110 sdio->final_block_in_request = sdio->block_in_file +
1112 /* Page fetching state */
1115 sdio->curr_page = 0;
1117 sdio->total_pages = 0;
1118 if (user_addr & (PAGE_SIZE-1)) {
1119 sdio->total_pages++;
1120 bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
1122 sdio->total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
1123 sdio->curr_user_address = user_addr;
1125 retval = do_direct_IO(dio, sdio, map_bh);
1127 dio->result += iov[seg].iov_len -
1128 ((sdio->final_block_in_request - sdio->block_in_file) <<
1132 dio_cleanup(dio, sdio);
1135 } /* end iovec loop */
1140 static ssize_t direct_IO_bvec(struct bio_vec *bvec, unsigned long nr_segs,
1141 struct dio *dio, struct dio_submit *sdio,
1142 unsigned blkbits, struct buffer_head *map_bh)
1147 sdio->pages_in_io += nr_segs;
1149 for (seg = 0; seg < nr_segs; seg++) {
1150 sdio->size += bvec[seg].bv_len;
1152 /* Index into the first page of the first block */
1153 sdio->first_block_in_page = bvec[seg].bv_offset >> blkbits;
1154 sdio->final_block_in_request = sdio->block_in_file +
1155 (bvec[seg].bv_len >> blkbits);
1156 /* Page fetching state */
1157 sdio->curr_page = 0;
1158 page_cache_get(bvec[seg].bv_page);
1159 dio->pages[0] = bvec[seg].bv_page;
1163 sdio->total_pages = 1;
1164 sdio->curr_user_address = 0;
1166 retval = do_direct_IO(dio, sdio, map_bh);
1168 dio->result += bvec[seg].bv_len -
1169 ((sdio->final_block_in_request - sdio->block_in_file) <<
1173 dio_cleanup(dio, sdio);
1182 * This is a library function for use by filesystem drivers.
1184 * The locking rules are governed by the flags parameter:
1185 * - if the flags value contains DIO_LOCKING we use a fancy locking
1186 * scheme for dumb filesystems.
1187 * For writes this function is called under i_mutex and returns with
1188 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1189 * taken and dropped again before returning.
1190 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1191 * internal locking but rather rely on the filesystem to synchronize
1192 * direct I/O reads/writes versus each other and truncate.
1194 * To help with locking against truncate we incremented the i_dio_count
1195 * counter before starting direct I/O, and decrement it once we are done.
1196 * Truncate can wait for it to reach zero to provide exclusion. It is
1197 * expected that filesystem provide exclusion between new direct I/O
1198 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1199 * but other filesystems need to take care of this on their own.
1201 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1202 * is always inlined. Otherwise gcc is unable to split the structure into
1203 * individual fields and will generate much worse code. This is important
1204 * for the whole file.
1206 static inline ssize_t
1207 do_blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1208 struct block_device *bdev, struct iov_iter *iter, loff_t offset,
1209 get_block_t get_block, dio_iodone_t end_io, dio_submit_t submit_io,
1215 unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits);
1216 unsigned blkbits = i_blkbits;
1217 unsigned blocksize_mask = (1 << blkbits) - 1;
1218 ssize_t retval = -EINVAL;
1219 loff_t end = offset;
1221 struct dio_submit sdio = { 0, };
1222 struct buffer_head map_bh = { 0, };
1223 struct blk_plug plug;
1224 unsigned long nr_segs = iter->nr_segs;
1230 * Avoid references to bdev if not absolutely needed to give
1231 * the early prefetch in the caller enough time.
1234 if (offset & blocksize_mask) {
1236 blkbits = blksize_bits(bdev_logical_block_size(bdev));
1237 blocksize_mask = (1 << blkbits) - 1;
1238 if (offset & blocksize_mask)
1242 /* Check the memory alignment. Blocks cannot straddle pages */
1243 if (iov_iter_has_iovec(iter)) {
1244 const struct iovec *iov = iov_iter_iovec(iter);
1246 for (seg = 0; seg < nr_segs; seg++) {
1247 addr = (unsigned long)iov[seg].iov_base;
1248 size = iov[seg].iov_len;
1250 if (unlikely((addr & blocksize_mask) ||
1251 (size & blocksize_mask))) {
1253 blkbits = blksize_bits(
1254 bdev_logical_block_size(bdev));
1255 blocksize_mask = (1 << blkbits) - 1;
1256 if ((addr & blocksize_mask) ||
1257 (size & blocksize_mask))
1261 } else if (iov_iter_has_bvec(iter)) {
1263 * Is this necessary, or can we trust the in-kernel
1264 * caller? Can we replace this with
1265 * end += iov_iter_count(iter); ?
1267 struct bio_vec *bvec = iov_iter_bvec(iter);
1269 for (seg = 0; seg < nr_segs; seg++) {
1270 addr = bvec[seg].bv_offset;
1271 size = bvec[seg].bv_len;
1273 if (unlikely((addr & blocksize_mask) ||
1274 (size & blocksize_mask))) {
1276 blkbits = blksize_bits(
1277 bdev_logical_block_size(bdev));
1278 blocksize_mask = (1 << blkbits) - 1;
1279 if ((addr & blocksize_mask) ||
1280 (size & blocksize_mask))
1287 /* watch out for a 0 len io from a tricksy fs */
1288 if (rw == READ && end == offset)
1291 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
1296 * Believe it or not, zeroing out the page array caused a .5%
1297 * performance regression in a database benchmark. So, we take
1298 * care to only zero out what's needed.
1300 memset(dio, 0, offsetof(struct dio, pages));
1303 if (dio->flags & DIO_LOCKING) {
1305 struct address_space *mapping =
1306 iocb->ki_filp->f_mapping;
1308 /* will be released by direct_io_worker */
1309 mutex_lock(&inode->i_mutex);
1311 retval = filemap_write_and_wait_range(mapping, offset,
1314 mutex_unlock(&inode->i_mutex);
1315 kmem_cache_free(dio_cache, dio);
1322 * For file extending writes updating i_size before data
1323 * writeouts complete can expose uninitialized blocks. So
1324 * even for AIO, we need to wait for i/o to complete before
1325 * returning in this case.
1327 dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
1328 (end > i_size_read(inode)));
1333 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1334 * so that we can call ->fsync.
1336 if (dio->is_async && (rw & WRITE) &&
1337 ((iocb->ki_filp->f_flags & O_DSYNC) ||
1338 IS_SYNC(iocb->ki_filp->f_mapping->host))) {
1339 retval = dio_set_defer_completion(dio);
1342 * We grab i_mutex only for reads so we don't have
1343 * to release it here
1345 kmem_cache_free(dio_cache, dio);
1351 * Will be decremented at I/O completion time.
1353 atomic_inc(&inode->i_dio_count);
1356 sdio.blkbits = blkbits;
1357 sdio.blkfactor = i_blkbits - blkbits;
1358 sdio.block_in_file = offset >> blkbits;
1360 sdio.get_block = get_block;
1361 dio->end_io = end_io;
1362 sdio.submit_io = submit_io;
1363 sdio.final_block_in_bio = -1;
1364 sdio.next_block_for_io = -1;
1367 dio->i_size = i_size_read(inode);
1369 spin_lock_init(&dio->bio_lock);
1373 * In case of non-aligned buffers, we may need 2 more
1374 * pages since we need to zero out first and last block.
1376 if (unlikely(sdio.blkfactor))
1377 sdio.pages_in_io = 2;
1379 blk_start_plug(&plug);
1381 if (iov_iter_has_iovec(iter))
1382 retval = direct_IO_iovec(iov_iter_iovec(iter), nr_segs, dio,
1383 &sdio, blkbits, &map_bh);
1385 retval = direct_IO_bvec(iov_iter_bvec(iter), nr_segs, dio,
1386 &sdio, blkbits, &map_bh);
1388 if (retval == -ENOTBLK) {
1390 * The remaining part of the request will be
1391 * be handled by buffered I/O when we return
1396 * There may be some unwritten disk at the end of a part-written
1397 * fs-block-sized block. Go zero that now.
1399 dio_zero_block(dio, &sdio, 1, &map_bh);
1401 if (sdio.cur_page) {
1404 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1407 page_cache_release(sdio.cur_page);
1408 sdio.cur_page = NULL;
1411 dio_bio_submit(dio, &sdio);
1413 blk_finish_plug(&plug);
1416 * It is possible that, we return short IO due to end of file.
1417 * In that case, we need to release all the pages we got hold on.
1419 dio_cleanup(dio, &sdio);
1422 * All block lookups have been performed. For READ requests
1423 * we can let i_mutex go now that its achieved its purpose
1424 * of protecting us from looking up uninitialized blocks.
1426 if (rw == READ && (dio->flags & DIO_LOCKING))
1427 mutex_unlock(&dio->inode->i_mutex);
1430 * The only time we want to leave bios in flight is when a successful
1431 * partial aio read or full aio write have been setup. In that case
1432 * bio completion will call aio_complete. The only time it's safe to
1433 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1434 * This had *better* be the only place that raises -EIOCBQUEUED.
1436 BUG_ON(retval == -EIOCBQUEUED);
1437 if (dio->is_async && retval == 0 && dio->result &&
1438 ((rw == READ) || (dio->result == sdio.size)))
1439 retval = -EIOCBQUEUED;
1441 if (retval != -EIOCBQUEUED)
1442 dio_await_completion(dio);
1444 if (drop_refcount(dio) == 0) {
1445 retval = dio_complete(dio, offset, retval, false);
1447 BUG_ON(retval != -EIOCBQUEUED);
1454 __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1455 struct block_device *bdev, struct iov_iter *iter, loff_t offset,
1456 get_block_t get_block, dio_iodone_t end_io, dio_submit_t submit_io,
1460 * The block device state is needed in the end to finally
1461 * submit everything. Since it's likely to be cache cold
1462 * prefetch it here as first thing to hide some of the
1465 * Attempt to prefetch the pieces we likely need later.
1467 prefetch(&bdev->bd_disk->part_tbl);
1468 prefetch(bdev->bd_queue);
1469 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1471 return do_blockdev_direct_IO(rw, iocb, inode, bdev, iter, offset,
1472 get_block, end_io, submit_io, flags);
1475 EXPORT_SYMBOL(__blockdev_direct_IO);
1477 static __init int dio_init(void)
1479 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1482 module_init(dio_init)