1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 2002, Linus Torvalds.
9 * 04Jul2002 Andrew Morton
11 * 11Sep2002 janetinc@us.ibm.com
12 * added readv/writev support.
13 * 29Oct2002 Andrew Morton
14 * rewrote bio_add_page() support.
15 * 30Oct2002 pbadari@us.ibm.com
16 * added support for non-aligned IO.
17 * 06Nov2002 pbadari@us.ibm.com
18 * added asynchronous IO support.
19 * 21Jul2003 nathans@sgi.com
20 * added IO completion notifier.
23 #include <linux/kernel.h>
24 #include <linux/module.h>
25 #include <linux/types.h>
28 #include <linux/slab.h>
29 #include <linux/highmem.h>
30 #include <linux/pagemap.h>
31 #include <linux/task_io_accounting_ops.h>
32 #include <linux/bio.h>
33 #include <linux/wait.h>
34 #include <linux/err.h>
35 #include <linux/blkdev.h>
36 #include <linux/buffer_head.h>
37 #include <linux/rwsem.h>
38 #include <linux/uio.h>
39 #include <linux/atomic.h>
40 #include <linux/prefetch.h>
45 * How many user pages to map in one call to iov_iter_extract_pages(). This
46 * determines the size of a structure in the slab cache
51 * Flags for dio_complete()
53 #define DIO_COMPLETE_ASYNC 0x01 /* This is async IO */
54 #define DIO_COMPLETE_INVALIDATE 0x02 /* Can invalidate pages */
57 * This code generally works in units of "dio_blocks". A dio_block is
58 * somewhere between the hard sector size and the filesystem block size. it
59 * is determined on a per-invocation basis. When talking to the filesystem
60 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
61 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
62 * to bio_block quantities by shifting left by blkfactor.
64 * If blkfactor is zero then the user's request was aligned to the filesystem's
68 /* dio_state only used in the submission path */
71 struct bio *bio; /* bio under assembly */
72 unsigned blkbits; /* doesn't change */
73 unsigned blkfactor; /* When we're using an alignment which
74 is finer than the filesystem's soft
75 blocksize, this specifies how much
76 finer. blkfactor=2 means 1/4-block
77 alignment. Does not change */
78 unsigned start_zero_done; /* flag: sub-blocksize zeroing has
79 been performed at the start of a
81 int pages_in_io; /* approximate total IO pages */
82 sector_t block_in_file; /* Current offset into the underlying
83 file in dio_block units. */
84 unsigned blocks_available; /* At block_in_file. changes */
85 int reap_counter; /* rate limit reaping */
86 sector_t final_block_in_request;/* doesn't change */
87 int boundary; /* prev block is at a boundary */
88 get_block_t *get_block; /* block mapping function */
90 loff_t logical_offset_in_bio; /* current first logical block in bio */
91 sector_t final_block_in_bio; /* current final block in bio + 1 */
92 sector_t next_block_for_io; /* next block to be put under IO,
93 in dio_blocks units */
96 * Deferred addition of a page to the dio. These variables are
97 * private to dio_send_cur_page(), submit_page_section() and
100 struct page *cur_page; /* The page */
101 unsigned cur_page_offset; /* Offset into it, in bytes */
102 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
103 sector_t cur_page_block; /* Where it starts */
104 loff_t cur_page_fs_offset; /* Offset in file */
106 struct iov_iter *iter;
108 * Page queue. These variables belong to dio_refill_pages() and
111 unsigned head; /* next page to process */
112 unsigned tail; /* last valid page + 1 */
116 /* dio_state communicated between submission path and end_io */
118 int flags; /* doesn't change */
119 blk_opf_t opf; /* request operation type and flags */
120 struct gendisk *bio_disk;
122 loff_t i_size; /* i_size when submitted */
123 dio_iodone_t *end_io; /* IO completion function */
124 bool is_pinned; /* T if we have pins on the pages */
126 void *private; /* copy from map_bh.b_private */
128 /* BIO completion state */
129 spinlock_t bio_lock; /* protects BIO fields below */
130 int page_errors; /* err from iov_iter_extract_pages() */
131 int is_async; /* is IO async ? */
132 bool defer_completion; /* defer AIO completion to workqueue? */
133 bool should_dirty; /* if pages should be dirtied */
134 int io_error; /* IO error in completion path */
135 unsigned long refcount; /* direct_io_worker() and bios */
136 struct bio *bio_list; /* singly linked via bi_private */
137 struct task_struct *waiter; /* waiting task (NULL if none) */
139 /* AIO related stuff */
140 struct kiocb *iocb; /* kiocb */
141 ssize_t result; /* IO result */
144 * pages[] (and any fields placed after it) are not zeroed out at
145 * allocation time. Don't add new fields after pages[] unless you
146 * wish that they not be zeroed.
149 struct page *pages[DIO_PAGES]; /* page buffer */
150 struct work_struct complete_work;/* deferred AIO completion */
152 } ____cacheline_aligned_in_smp;
154 static struct kmem_cache *dio_cache __read_mostly;
157 * How many pages are in the queue?
159 static inline unsigned dio_pages_present(struct dio_submit *sdio)
161 return sdio->tail - sdio->head;
165 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
167 static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
169 struct page **pages = dio->pages;
170 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
173 ret = iov_iter_extract_pages(sdio->iter, &pages, LONG_MAX,
174 DIO_PAGES, 0, &sdio->from);
176 if (ret < 0 && sdio->blocks_available && dio_op == REQ_OP_WRITE) {
178 * A memory fault, but the filesystem has some outstanding
179 * mapped blocks. We need to use those blocks up to avoid
180 * leaking stale data in the file.
182 if (dio->page_errors == 0)
183 dio->page_errors = ret;
184 dio->pages[0] = ZERO_PAGE(0);
188 sdio->to = PAGE_SIZE;
195 sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
196 sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1;
203 * Get another userspace page. Returns an ERR_PTR on error. Pages are
204 * buffered inside the dio so that we can call iov_iter_extract_pages()
205 * against a decent number of pages, less frequently. To provide nicer use of
208 static inline struct page *dio_get_page(struct dio *dio,
209 struct dio_submit *sdio)
211 if (dio_pages_present(sdio) == 0) {
214 ret = dio_refill_pages(dio, sdio);
217 BUG_ON(dio_pages_present(sdio) == 0);
219 return dio->pages[sdio->head];
222 static void dio_pin_page(struct dio *dio, struct page *page)
225 folio_add_pin(page_folio(page));
228 static void dio_unpin_page(struct dio *dio, struct page *page)
231 unpin_user_page(page);
235 * dio_complete() - called when all DIO BIO I/O has been completed
237 * This drops i_dio_count, lets interested parties know that a DIO operation
238 * has completed, and calculates the resulting return code for the operation.
240 * It lets the filesystem know if it registered an interest earlier via
241 * get_block. Pass the private field of the map buffer_head so that
242 * filesystems can use it to hold additional state between get_block calls and
245 static ssize_t dio_complete(struct dio *dio, ssize_t ret, unsigned int flags)
247 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
248 loff_t offset = dio->iocb->ki_pos;
249 ssize_t transferred = 0;
253 * AIO submission can race with bio completion to get here while
254 * expecting to have the last io completed by bio completion.
255 * In that case -EIOCBQUEUED is in fact not an error we want
256 * to preserve through this call.
258 if (ret == -EIOCBQUEUED)
262 transferred = dio->result;
264 /* Check for short read case */
265 if (dio_op == REQ_OP_READ &&
266 ((offset + transferred) > dio->i_size))
267 transferred = dio->i_size - offset;
268 /* ignore EFAULT if some IO has been done */
269 if (unlikely(ret == -EFAULT) && transferred)
274 ret = dio->page_errors;
282 err = dio->end_io(dio->iocb, offset, ret, dio->private);
288 * Try again to invalidate clean pages which might have been cached by
289 * non-direct readahead, or faulted in by get_user_pages() if the source
290 * of the write was an mmap'ed region of the file we're writing. Either
291 * one is a pretty crazy thing to do, so we don't support it 100%. If
292 * this invalidation fails, tough, the write still worked...
294 * And this page cache invalidation has to be after dio->end_io(), as
295 * some filesystems convert unwritten extents to real allocations in
296 * end_io() when necessary, otherwise a racing buffer read would cache
297 * zeros from unwritten extents.
299 if (flags & DIO_COMPLETE_INVALIDATE &&
300 ret > 0 && dio_op == REQ_OP_WRITE &&
301 dio->inode->i_mapping->nrpages) {
302 err = invalidate_inode_pages2_range(dio->inode->i_mapping,
303 offset >> PAGE_SHIFT,
304 (offset + ret - 1) >> PAGE_SHIFT);
306 dio_warn_stale_pagecache(dio->iocb->ki_filp);
309 inode_dio_end(dio->inode);
311 if (flags & DIO_COMPLETE_ASYNC) {
313 * generic_write_sync expects ki_pos to have been updated
314 * already, but the submission path only does this for
317 dio->iocb->ki_pos += transferred;
319 if (ret > 0 && dio_op == REQ_OP_WRITE)
320 ret = generic_write_sync(dio->iocb, ret);
321 dio->iocb->ki_complete(dio->iocb, ret);
324 kmem_cache_free(dio_cache, dio);
328 static void dio_aio_complete_work(struct work_struct *work)
330 struct dio *dio = container_of(work, struct dio, complete_work);
332 dio_complete(dio, 0, DIO_COMPLETE_ASYNC | DIO_COMPLETE_INVALIDATE);
335 static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio);
338 * Asynchronous IO callback.
340 static void dio_bio_end_aio(struct bio *bio)
342 struct dio *dio = bio->bi_private;
343 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
344 unsigned long remaining;
346 bool defer_completion = false;
348 /* cleanup the bio */
349 dio_bio_complete(dio, bio);
351 spin_lock_irqsave(&dio->bio_lock, flags);
352 remaining = --dio->refcount;
353 if (remaining == 1 && dio->waiter)
354 wake_up_process(dio->waiter);
355 spin_unlock_irqrestore(&dio->bio_lock, flags);
357 if (remaining == 0) {
359 * Defer completion when defer_completion is set or
360 * when the inode has pages mapped and this is AIO write.
361 * We need to invalidate those pages because there is a
362 * chance they contain stale data in the case buffered IO
363 * went in between AIO submission and completion into the
367 defer_completion = dio->defer_completion ||
368 (dio_op == REQ_OP_WRITE &&
369 dio->inode->i_mapping->nrpages);
370 if (defer_completion) {
371 INIT_WORK(&dio->complete_work, dio_aio_complete_work);
372 queue_work(dio->inode->i_sb->s_dio_done_wq,
373 &dio->complete_work);
375 dio_complete(dio, 0, DIO_COMPLETE_ASYNC);
381 * The BIO completion handler simply queues the BIO up for the process-context
384 * During I/O bi_private points at the dio. After I/O, bi_private is used to
385 * implement a singly-linked list of completed BIOs, at dio->bio_list.
387 static void dio_bio_end_io(struct bio *bio)
389 struct dio *dio = bio->bi_private;
392 spin_lock_irqsave(&dio->bio_lock, flags);
393 bio->bi_private = dio->bio_list;
395 if (--dio->refcount == 1 && dio->waiter)
396 wake_up_process(dio->waiter);
397 spin_unlock_irqrestore(&dio->bio_lock, flags);
401 dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
402 struct block_device *bdev,
403 sector_t first_sector, int nr_vecs)
408 * bio_alloc() is guaranteed to return a bio when allowed to sleep and
409 * we request a valid number of vectors.
411 bio = bio_alloc(bdev, nr_vecs, dio->opf, GFP_KERNEL);
412 bio->bi_iter.bi_sector = first_sector;
414 bio->bi_end_io = dio_bio_end_aio;
416 bio->bi_end_io = dio_bio_end_io;
418 bio_set_flag(bio, BIO_PAGE_PINNED);
420 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
424 * In the AIO read case we speculatively dirty the pages before starting IO.
425 * During IO completion, any of these pages which happen to have been written
426 * back will be redirtied by bio_check_pages_dirty().
428 * bios hold a dio reference between submit_bio and ->end_io.
430 static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
432 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
433 struct bio *bio = sdio->bio;
436 bio->bi_private = dio;
438 spin_lock_irqsave(&dio->bio_lock, flags);
440 spin_unlock_irqrestore(&dio->bio_lock, flags);
442 if (dio->is_async && dio_op == REQ_OP_READ && dio->should_dirty)
443 bio_set_pages_dirty(bio);
445 dio->bio_disk = bio->bi_bdev->bd_disk;
451 sdio->logical_offset_in_bio = 0;
455 * Release any resources in case of a failure
457 static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
460 unpin_user_pages(dio->pages + sdio->head,
461 sdio->tail - sdio->head);
465 * Wait for the next BIO to complete. Remove it and return it. NULL is
466 * returned once all BIOs have been completed. This must only be called once
467 * all bios have been issued so that dio->refcount can only decrease. This
468 * requires that the caller hold a reference on the dio.
470 static struct bio *dio_await_one(struct dio *dio)
473 struct bio *bio = NULL;
475 spin_lock_irqsave(&dio->bio_lock, flags);
478 * Wait as long as the list is empty and there are bios in flight. bio
479 * completion drops the count, maybe adds to the list, and wakes while
480 * holding the bio_lock so we don't need set_current_state()'s barrier
481 * and can call it after testing our condition.
483 while (dio->refcount > 1 && dio->bio_list == NULL) {
484 __set_current_state(TASK_UNINTERRUPTIBLE);
485 dio->waiter = current;
486 spin_unlock_irqrestore(&dio->bio_lock, flags);
488 /* wake up sets us TASK_RUNNING */
489 spin_lock_irqsave(&dio->bio_lock, flags);
494 dio->bio_list = bio->bi_private;
496 spin_unlock_irqrestore(&dio->bio_lock, flags);
501 * Process one completed BIO. No locks are held.
503 static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio)
505 blk_status_t err = bio->bi_status;
506 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
507 bool should_dirty = dio_op == REQ_OP_READ && dio->should_dirty;
510 if (err == BLK_STS_AGAIN && (bio->bi_opf & REQ_NOWAIT))
511 dio->io_error = -EAGAIN;
513 dio->io_error = -EIO;
516 if (dio->is_async && should_dirty) {
517 bio_check_pages_dirty(bio); /* transfers ownership */
519 bio_release_pages(bio, should_dirty);
526 * Wait on and process all in-flight BIOs. This must only be called once
527 * all bios have been issued so that the refcount can only decrease.
528 * This just waits for all bios to make it through dio_bio_complete. IO
529 * errors are propagated through dio->io_error and should be propagated via
532 static void dio_await_completion(struct dio *dio)
536 bio = dio_await_one(dio);
538 dio_bio_complete(dio, bio);
543 * A really large O_DIRECT read or write can generate a lot of BIOs. So
544 * to keep the memory consumption sane we periodically reap any completed BIOs
545 * during the BIO generation phase.
547 * This also helps to limit the peak amount of pinned userspace memory.
549 static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
553 if (sdio->reap_counter++ >= 64) {
554 while (dio->bio_list) {
559 spin_lock_irqsave(&dio->bio_lock, flags);
561 dio->bio_list = bio->bi_private;
562 spin_unlock_irqrestore(&dio->bio_lock, flags);
563 ret2 = blk_status_to_errno(dio_bio_complete(dio, bio));
567 sdio->reap_counter = 0;
572 static int dio_set_defer_completion(struct dio *dio)
574 struct super_block *sb = dio->inode->i_sb;
576 if (dio->defer_completion)
578 dio->defer_completion = true;
579 if (!sb->s_dio_done_wq)
580 return sb_init_dio_done_wq(sb);
585 * Call into the fs to map some more disk blocks. We record the current number
586 * of available blocks at sdio->blocks_available. These are in units of the
587 * fs blocksize, i_blocksize(inode).
589 * The fs is allowed to map lots of blocks at once. If it wants to do that,
590 * it uses the passed inode-relative block number as the file offset, as usual.
592 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
593 * has remaining to do. The fs should not map more than this number of blocks.
595 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
596 * indicate how much contiguous disk space has been made available at
599 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
600 * This isn't very efficient...
602 * In the case of filesystem holes: the fs may return an arbitrarily-large
603 * hole by returning an appropriate value in b_size and by clearing
604 * buffer_mapped(). However the direct-io code will only process holes one
605 * block at a time - it will repeatedly call get_block() as it walks the hole.
607 static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
608 struct buffer_head *map_bh)
610 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
612 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
613 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
614 unsigned long fs_count; /* Number of filesystem-sized blocks */
616 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
620 * If there was a memory error and we've overwritten all the
621 * mapped blocks then we can now return that memory error
623 ret = dio->page_errors;
625 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
626 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
627 fs_endblk = (sdio->final_block_in_request - 1) >>
629 fs_count = fs_endblk - fs_startblk + 1;
632 map_bh->b_size = fs_count << i_blkbits;
635 * For writes that could fill holes inside i_size on a
636 * DIO_SKIP_HOLES filesystem we forbid block creations: only
637 * overwrites are permitted. We will return early to the caller
638 * once we see an unmapped buffer head returned, and the caller
639 * will fall back to buffered I/O.
641 * Otherwise the decision is left to the get_blocks method,
642 * which may decide to handle it or also return an unmapped
645 create = dio_op == REQ_OP_WRITE;
646 if (dio->flags & DIO_SKIP_HOLES) {
647 i_size = i_size_read(dio->inode);
648 if (i_size && fs_startblk <= (i_size - 1) >> i_blkbits)
652 ret = (*sdio->get_block)(dio->inode, fs_startblk,
655 /* Store for completion */
656 dio->private = map_bh->b_private;
658 if (ret == 0 && buffer_defer_completion(map_bh))
659 ret = dio_set_defer_completion(dio);
665 * There is no bio. Make one now.
667 static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
668 sector_t start_sector, struct buffer_head *map_bh)
673 ret = dio_bio_reap(dio, sdio);
676 sector = start_sector << (sdio->blkbits - 9);
677 nr_pages = bio_max_segs(sdio->pages_in_io);
678 BUG_ON(nr_pages <= 0);
679 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
686 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
687 * that was successful then update final_block_in_bio and take a ref against
688 * the just-added page.
690 * Return zero on success. Non-zero means the caller needs to start a new BIO.
692 static inline int dio_bio_add_page(struct dio *dio, struct dio_submit *sdio)
696 ret = bio_add_page(sdio->bio, sdio->cur_page,
697 sdio->cur_page_len, sdio->cur_page_offset);
698 if (ret == sdio->cur_page_len) {
700 * Decrement count only, if we are done with this page
702 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
704 dio_pin_page(dio, sdio->cur_page);
705 sdio->final_block_in_bio = sdio->cur_page_block +
706 (sdio->cur_page_len >> sdio->blkbits);
715 * Put cur_page under IO. The section of cur_page which is described by
716 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
717 * starts on-disk at cur_page_block.
719 * We take a ref against the page here (on behalf of its presence in the bio).
721 * The caller of this function is responsible for removing cur_page from the
722 * dio, and for dropping the refcount which came from that presence.
724 static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
725 struct buffer_head *map_bh)
730 loff_t cur_offset = sdio->cur_page_fs_offset;
731 loff_t bio_next_offset = sdio->logical_offset_in_bio +
732 sdio->bio->bi_iter.bi_size;
735 * See whether this new request is contiguous with the old.
737 * Btrfs cannot handle having logically non-contiguous requests
738 * submitted. For example if you have
740 * Logical: [0-4095][HOLE][8192-12287]
741 * Physical: [0-4095] [4096-8191]
743 * We cannot submit those pages together as one BIO. So if our
744 * current logical offset in the file does not equal what would
745 * be the next logical offset in the bio, submit the bio we
748 if (sdio->final_block_in_bio != sdio->cur_page_block ||
749 cur_offset != bio_next_offset)
750 dio_bio_submit(dio, sdio);
753 if (sdio->bio == NULL) {
754 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
759 if (dio_bio_add_page(dio, sdio) != 0) {
760 dio_bio_submit(dio, sdio);
761 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
763 ret = dio_bio_add_page(dio, sdio);
772 * An autonomous function to put a chunk of a page under deferred IO.
774 * The caller doesn't actually know (or care) whether this piece of page is in
775 * a BIO, or is under IO or whatever. We just take care of all possible
776 * situations here. The separation between the logic of do_direct_IO() and
777 * that of submit_page_section() is important for clarity. Please don't break.
779 * The chunk of page starts on-disk at blocknr.
781 * We perform deferred IO, by recording the last-submitted page inside our
782 * private part of the dio structure. If possible, we just expand the IO
783 * across that page here.
785 * If that doesn't work out then we put the old page into the bio and add this
786 * page to the dio instead.
789 submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
790 unsigned offset, unsigned len, sector_t blocknr,
791 struct buffer_head *map_bh)
793 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
795 int boundary = sdio->boundary; /* dio_send_cur_page may clear it */
797 if (dio_op == REQ_OP_WRITE) {
799 * Read accounting is performed in submit_bio()
801 task_io_account_write(len);
805 * Can we just grow the current page's presence in the dio?
807 if (sdio->cur_page == page &&
808 sdio->cur_page_offset + sdio->cur_page_len == offset &&
809 sdio->cur_page_block +
810 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
811 sdio->cur_page_len += len;
816 * If there's a deferred page already there then send it.
818 if (sdio->cur_page) {
819 ret = dio_send_cur_page(dio, sdio, map_bh);
820 dio_unpin_page(dio, sdio->cur_page);
821 sdio->cur_page = NULL;
826 dio_pin_page(dio, page); /* It is in dio */
827 sdio->cur_page = page;
828 sdio->cur_page_offset = offset;
829 sdio->cur_page_len = len;
830 sdio->cur_page_block = blocknr;
831 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
834 * If boundary then we want to schedule the IO now to
835 * avoid metadata seeks.
838 ret = dio_send_cur_page(dio, sdio, map_bh);
840 dio_bio_submit(dio, sdio);
841 dio_unpin_page(dio, sdio->cur_page);
842 sdio->cur_page = NULL;
848 * If we are not writing the entire block and get_block() allocated
849 * the block for us, we need to fill-in the unused portion of the
850 * block with zeros. This happens only if user-buffer, fileoffset or
851 * io length is not filesystem block-size multiple.
853 * `end' is zero if we're doing the start of the IO, 1 at the end of the
856 static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
857 int end, struct buffer_head *map_bh)
859 unsigned dio_blocks_per_fs_block;
860 unsigned this_chunk_blocks; /* In dio_blocks */
861 unsigned this_chunk_bytes;
864 sdio->start_zero_done = 1;
865 if (!sdio->blkfactor || !buffer_new(map_bh))
868 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
869 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
871 if (!this_chunk_blocks)
875 * We need to zero out part of an fs block. It is either at the
876 * beginning or the end of the fs block.
879 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
881 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
884 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
885 sdio->next_block_for_io, map_bh))
888 sdio->next_block_for_io += this_chunk_blocks;
892 * Walk the user pages, and the file, mapping blocks to disk and generating
893 * a sequence of (page,offset,len,block) mappings. These mappings are injected
894 * into submit_page_section(), which takes care of the next stage of submission
896 * Direct IO against a blockdev is different from a file. Because we can
897 * happily perform page-sized but 512-byte aligned IOs. It is important that
898 * blockdev IO be able to have fine alignment and large sizes.
900 * So what we do is to permit the ->get_block function to populate bh.b_size
901 * with the size of IO which is permitted at this offset and this i_blkbits.
903 * For best results, the blockdev should be set up with 512-byte i_blkbits and
904 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
905 * fine alignment but still allows this function to work in PAGE_SIZE units.
907 static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
908 struct buffer_head *map_bh)
910 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
911 const unsigned blkbits = sdio->blkbits;
912 const unsigned i_blkbits = blkbits + sdio->blkfactor;
915 while (sdio->block_in_file < sdio->final_block_in_request) {
919 page = dio_get_page(dio, sdio);
924 from = sdio->head ? 0 : sdio->from;
925 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
929 unsigned this_chunk_bytes; /* # of bytes mapped */
930 unsigned this_chunk_blocks; /* # of blocks */
933 if (sdio->blocks_available == 0) {
935 * Need to go and map some more disk
937 unsigned long blkmask;
938 unsigned long dio_remainder;
940 ret = get_more_blocks(dio, sdio, map_bh);
942 dio_unpin_page(dio, page);
945 if (!buffer_mapped(map_bh))
948 sdio->blocks_available =
949 map_bh->b_size >> blkbits;
950 sdio->next_block_for_io =
951 map_bh->b_blocknr << sdio->blkfactor;
952 if (buffer_new(map_bh)) {
956 map_bh->b_size >> i_blkbits);
959 if (!sdio->blkfactor)
962 blkmask = (1 << sdio->blkfactor) - 1;
963 dio_remainder = (sdio->block_in_file & blkmask);
966 * If we are at the start of IO and that IO
967 * starts partway into a fs-block,
968 * dio_remainder will be non-zero. If the IO
969 * is a read then we can simply advance the IO
970 * cursor to the first block which is to be
971 * read. But if the IO is a write and the
972 * block was newly allocated we cannot do that;
973 * the start of the fs block must be zeroed out
976 if (!buffer_new(map_bh))
977 sdio->next_block_for_io += dio_remainder;
978 sdio->blocks_available -= dio_remainder;
982 if (!buffer_mapped(map_bh)) {
983 loff_t i_size_aligned;
985 /* AKPM: eargh, -ENOTBLK is a hack */
986 if (dio_op == REQ_OP_WRITE) {
987 dio_unpin_page(dio, page);
992 * Be sure to account for a partial block as the
993 * last block in the file
995 i_size_aligned = ALIGN(i_size_read(dio->inode),
997 if (sdio->block_in_file >=
998 i_size_aligned >> blkbits) {
1000 dio_unpin_page(dio, page);
1003 zero_user(page, from, 1 << blkbits);
1004 sdio->block_in_file++;
1005 from += 1 << blkbits;
1006 dio->result += 1 << blkbits;
1011 * If we're performing IO which has an alignment which
1012 * is finer than the underlying fs, go check to see if
1013 * we must zero out the start of this block.
1015 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
1016 dio_zero_block(dio, sdio, 0, map_bh);
1019 * Work out, in this_chunk_blocks, how much disk we
1020 * can add to this page
1022 this_chunk_blocks = sdio->blocks_available;
1023 u = (to - from) >> blkbits;
1024 if (this_chunk_blocks > u)
1025 this_chunk_blocks = u;
1026 u = sdio->final_block_in_request - sdio->block_in_file;
1027 if (this_chunk_blocks > u)
1028 this_chunk_blocks = u;
1029 this_chunk_bytes = this_chunk_blocks << blkbits;
1030 BUG_ON(this_chunk_bytes == 0);
1032 if (this_chunk_blocks == sdio->blocks_available)
1033 sdio->boundary = buffer_boundary(map_bh);
1034 ret = submit_page_section(dio, sdio, page,
1037 sdio->next_block_for_io,
1040 dio_unpin_page(dio, page);
1043 sdio->next_block_for_io += this_chunk_blocks;
1045 sdio->block_in_file += this_chunk_blocks;
1046 from += this_chunk_bytes;
1047 dio->result += this_chunk_bytes;
1048 sdio->blocks_available -= this_chunk_blocks;
1050 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1051 if (sdio->block_in_file == sdio->final_block_in_request)
1055 /* Drop the pin which was taken in get_user_pages() */
1056 dio_unpin_page(dio, page);
1062 static inline int drop_refcount(struct dio *dio)
1065 unsigned long flags;
1068 * Sync will always be dropping the final ref and completing the
1069 * operation. AIO can if it was a broken operation described above or
1070 * in fact if all the bios race to complete before we get here. In
1071 * that case dio_complete() translates the EIOCBQUEUED into the proper
1072 * return code that the caller will hand to ->complete().
1074 * This is managed by the bio_lock instead of being an atomic_t so that
1075 * completion paths can drop their ref and use the remaining count to
1076 * decide to wake the submission path atomically.
1078 spin_lock_irqsave(&dio->bio_lock, flags);
1079 ret2 = --dio->refcount;
1080 spin_unlock_irqrestore(&dio->bio_lock, flags);
1085 * This is a library function for use by filesystem drivers.
1087 * The locking rules are governed by the flags parameter:
1088 * - if the flags value contains DIO_LOCKING we use a fancy locking
1089 * scheme for dumb filesystems.
1090 * For writes this function is called under i_mutex and returns with
1091 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1092 * taken and dropped again before returning.
1093 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1094 * internal locking but rather rely on the filesystem to synchronize
1095 * direct I/O reads/writes versus each other and truncate.
1097 * To help with locking against truncate we incremented the i_dio_count
1098 * counter before starting direct I/O, and decrement it once we are done.
1099 * Truncate can wait for it to reach zero to provide exclusion. It is
1100 * expected that filesystem provide exclusion between new direct I/O
1101 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1102 * but other filesystems need to take care of this on their own.
1104 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1105 * is always inlined. Otherwise gcc is unable to split the structure into
1106 * individual fields and will generate much worse code. This is important
1107 * for the whole file.
1109 ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1110 struct block_device *bdev, struct iov_iter *iter,
1111 get_block_t get_block, dio_iodone_t end_io,
1114 unsigned i_blkbits = READ_ONCE(inode->i_blkbits);
1115 unsigned blkbits = i_blkbits;
1116 unsigned blocksize_mask = (1 << blkbits) - 1;
1117 ssize_t retval = -EINVAL;
1118 const size_t count = iov_iter_count(iter);
1119 loff_t offset = iocb->ki_pos;
1120 const loff_t end = offset + count;
1122 struct dio_submit sdio = { 0, };
1123 struct buffer_head map_bh = { 0, };
1124 struct blk_plug plug;
1125 unsigned long align = offset | iov_iter_alignment(iter);
1128 * Avoid references to bdev if not absolutely needed to give
1129 * the early prefetch in the caller enough time.
1132 /* watch out for a 0 len io from a tricksy fs */
1133 if (iov_iter_rw(iter) == READ && !count)
1136 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
1140 * Believe it or not, zeroing out the page array caused a .5%
1141 * performance regression in a database benchmark. So, we take
1142 * care to only zero out what's needed.
1144 memset(dio, 0, offsetof(struct dio, pages));
1147 if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ) {
1148 /* will be released by direct_io_worker */
1151 dio->is_pinned = iov_iter_extract_will_pin(iter);
1153 /* Once we sampled i_size check for reads beyond EOF */
1154 dio->i_size = i_size_read(inode);
1155 if (iov_iter_rw(iter) == READ && offset >= dio->i_size) {
1160 if (align & blocksize_mask) {
1162 blkbits = blksize_bits(bdev_logical_block_size(bdev));
1163 blocksize_mask = (1 << blkbits) - 1;
1164 if (align & blocksize_mask)
1168 if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ) {
1169 struct address_space *mapping = iocb->ki_filp->f_mapping;
1171 retval = filemap_write_and_wait_range(mapping, offset, end - 1);
1177 * For file extending writes updating i_size before data writeouts
1178 * complete can expose uninitialized blocks in dumb filesystems.
1179 * In that case we need to wait for I/O completion even if asked
1180 * for an asynchronous write.
1182 if (is_sync_kiocb(iocb))
1183 dio->is_async = false;
1184 else if (iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
1185 dio->is_async = false;
1187 dio->is_async = true;
1190 if (iov_iter_rw(iter) == WRITE) {
1191 dio->opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
1192 if (iocb->ki_flags & IOCB_NOWAIT)
1193 dio->opf |= REQ_NOWAIT;
1195 dio->opf = REQ_OP_READ;
1199 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1200 * so that we can call ->fsync.
1202 if (dio->is_async && iov_iter_rw(iter) == WRITE) {
1204 if (iocb_is_dsync(iocb))
1205 retval = dio_set_defer_completion(dio);
1206 else if (!dio->inode->i_sb->s_dio_done_wq) {
1208 * In case of AIO write racing with buffered read we
1209 * need to defer completion. We can't decide this now,
1210 * however the workqueue needs to be initialized here.
1212 retval = sb_init_dio_done_wq(dio->inode->i_sb);
1219 * Will be decremented at I/O completion time.
1221 inode_dio_begin(inode);
1224 sdio.blkbits = blkbits;
1225 sdio.blkfactor = i_blkbits - blkbits;
1226 sdio.block_in_file = offset >> blkbits;
1228 sdio.get_block = get_block;
1229 dio->end_io = end_io;
1230 sdio.final_block_in_bio = -1;
1231 sdio.next_block_for_io = -1;
1235 spin_lock_init(&dio->bio_lock);
1238 dio->should_dirty = user_backed_iter(iter) && iov_iter_rw(iter) == READ;
1240 sdio.final_block_in_request = end >> blkbits;
1243 * In case of non-aligned buffers, we may need 2 more
1244 * pages since we need to zero out first and last block.
1246 if (unlikely(sdio.blkfactor))
1247 sdio.pages_in_io = 2;
1249 sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
1251 blk_start_plug(&plug);
1253 retval = do_direct_IO(dio, &sdio, &map_bh);
1255 dio_cleanup(dio, &sdio);
1257 if (retval == -ENOTBLK) {
1259 * The remaining part of the request will be
1260 * handled by buffered I/O when we return
1265 * There may be some unwritten disk at the end of a part-written
1266 * fs-block-sized block. Go zero that now.
1268 dio_zero_block(dio, &sdio, 1, &map_bh);
1270 if (sdio.cur_page) {
1273 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1276 dio_unpin_page(dio, sdio.cur_page);
1277 sdio.cur_page = NULL;
1280 dio_bio_submit(dio, &sdio);
1282 blk_finish_plug(&plug);
1285 * It is possible that, we return short IO due to end of file.
1286 * In that case, we need to release all the pages we got hold on.
1288 dio_cleanup(dio, &sdio);
1291 * All block lookups have been performed. For READ requests
1292 * we can let i_mutex go now that its achieved its purpose
1293 * of protecting us from looking up uninitialized blocks.
1295 if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING))
1296 inode_unlock(dio->inode);
1299 * The only time we want to leave bios in flight is when a successful
1300 * partial aio read or full aio write have been setup. In that case
1301 * bio completion will call aio_complete. The only time it's safe to
1302 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1303 * This had *better* be the only place that raises -EIOCBQUEUED.
1305 BUG_ON(retval == -EIOCBQUEUED);
1306 if (dio->is_async && retval == 0 && dio->result &&
1307 (iov_iter_rw(iter) == READ || dio->result == count))
1308 retval = -EIOCBQUEUED;
1310 dio_await_completion(dio);
1312 if (drop_refcount(dio) == 0) {
1313 retval = dio_complete(dio, retval, DIO_COMPLETE_INVALIDATE);
1315 BUG_ON(retval != -EIOCBQUEUED);
1320 if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ)
1321 inode_unlock(inode);
1323 kmem_cache_free(dio_cache, dio);
1326 EXPORT_SYMBOL(__blockdev_direct_IO);
1328 static __init int dio_init(void)
1330 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1333 module_init(dio_init)