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 get_user_pages(). This determines
46 * 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 */
89 dio_submit_t *submit_io; /* IO submition function */
91 loff_t logical_offset_in_bio; /* current first logical block in bio */
92 sector_t final_block_in_bio; /* current final block in bio + 1 */
93 sector_t next_block_for_io; /* next block to be put under IO,
94 in dio_blocks units */
97 * Deferred addition of a page to the dio. These variables are
98 * private to dio_send_cur_page(), submit_page_section() and
101 struct page *cur_page; /* The page */
102 unsigned cur_page_offset; /* Offset into it, in bytes */
103 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
104 sector_t cur_page_block; /* Where it starts */
105 loff_t cur_page_fs_offset; /* Offset in file */
107 struct iov_iter *iter;
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 */
117 /* dio_state communicated between submission path and end_io */
119 int flags; /* doesn't change */
123 struct gendisk *bio_disk;
125 loff_t i_size; /* i_size when submitted */
126 dio_iodone_t *end_io; /* IO completion function */
128 void *private; /* copy from map_bh.b_private */
130 /* BIO completion state */
131 spinlock_t bio_lock; /* protects BIO fields below */
132 int page_errors; /* errno from get_user_pages() */
133 int is_async; /* is IO async ? */
134 bool defer_completion; /* defer AIO completion to workqueue? */
135 bool should_dirty; /* if pages should be dirtied */
136 int io_error; /* IO error in completion path */
137 unsigned long refcount; /* direct_io_worker() and bios */
138 struct bio *bio_list; /* singly linked via bi_private */
139 struct task_struct *waiter; /* waiting task (NULL if none) */
141 /* AIO related stuff */
142 struct kiocb *iocb; /* kiocb */
143 ssize_t result; /* IO result */
146 * pages[] (and any fields placed after it) are not zeroed out at
147 * allocation time. Don't add new fields after pages[] unless you
148 * wish that they not be zeroed.
151 struct page *pages[DIO_PAGES]; /* page buffer */
152 struct work_struct complete_work;/* deferred AIO completion */
154 } ____cacheline_aligned_in_smp;
156 static struct kmem_cache *dio_cache __read_mostly;
159 * How many pages are in the queue?
161 static inline unsigned dio_pages_present(struct dio_submit *sdio)
163 return sdio->tail - sdio->head;
167 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
169 static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
173 ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES,
176 if (ret < 0 && sdio->blocks_available && (dio->op == REQ_OP_WRITE)) {
177 struct page *page = ZERO_PAGE(0);
179 * A memory fault, but the filesystem has some outstanding
180 * mapped blocks. We need to use those blocks up to avoid
181 * leaking stale data in the file.
183 if (dio->page_errors == 0)
184 dio->page_errors = ret;
186 dio->pages[0] = page;
190 sdio->to = PAGE_SIZE;
195 iov_iter_advance(sdio->iter, ret);
198 sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
199 sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1;
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
228 * This drops i_dio_count, lets interested parties know that a DIO operation
229 * has completed, and calculates the resulting return code for the operation.
231 * It lets the filesystem know if it registered an interest earlier via
232 * get_block. Pass the private field of the map buffer_head so that
233 * filesystems can use it to hold additional state between get_block calls and
236 static ssize_t dio_complete(struct dio *dio, ssize_t ret, unsigned int flags)
238 loff_t offset = dio->iocb->ki_pos;
239 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->op == REQ_OP_READ) &&
256 ((offset + transferred) > dio->i_size))
257 transferred = dio->i_size - offset;
258 /* ignore EFAULT if some IO has been done */
259 if (unlikely(ret == -EFAULT) && transferred)
264 ret = dio->page_errors;
272 err = dio->end_io(dio->iocb, offset, ret, dio->private);
278 * Try again to invalidate clean pages which might have been cached by
279 * non-direct readahead, or faulted in by get_user_pages() if the source
280 * of the write was an mmap'ed region of the file we're writing. Either
281 * one is a pretty crazy thing to do, so we don't support it 100%. If
282 * this invalidation fails, tough, the write still worked...
284 * And this page cache invalidation has to be after dio->end_io(), as
285 * some filesystems convert unwritten extents to real allocations in
286 * end_io() when necessary, otherwise a racing buffer read would cache
287 * zeros from unwritten extents.
289 if (flags & DIO_COMPLETE_INVALIDATE &&
290 ret > 0 && dio->op == REQ_OP_WRITE &&
291 dio->inode->i_mapping->nrpages) {
292 err = invalidate_inode_pages2_range(dio->inode->i_mapping,
293 offset >> PAGE_SHIFT,
294 (offset + ret - 1) >> PAGE_SHIFT);
296 dio_warn_stale_pagecache(dio->iocb->ki_filp);
299 inode_dio_end(dio->inode);
301 if (flags & DIO_COMPLETE_ASYNC) {
303 * generic_write_sync expects ki_pos to have been updated
304 * already, but the submission path only does this for
307 dio->iocb->ki_pos += transferred;
309 if (ret > 0 && dio->op == REQ_OP_WRITE)
310 ret = generic_write_sync(dio->iocb, ret);
311 dio->iocb->ki_complete(dio->iocb, ret, 0);
314 kmem_cache_free(dio_cache, dio);
318 static void dio_aio_complete_work(struct work_struct *work)
320 struct dio *dio = container_of(work, struct dio, complete_work);
322 dio_complete(dio, 0, DIO_COMPLETE_ASYNC | DIO_COMPLETE_INVALIDATE);
325 static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio);
328 * Asynchronous IO callback.
330 static void dio_bio_end_aio(struct bio *bio)
332 struct dio *dio = bio->bi_private;
333 unsigned long remaining;
335 bool defer_completion = false;
337 /* cleanup the bio */
338 dio_bio_complete(dio, bio);
340 spin_lock_irqsave(&dio->bio_lock, flags);
341 remaining = --dio->refcount;
342 if (remaining == 1 && dio->waiter)
343 wake_up_process(dio->waiter);
344 spin_unlock_irqrestore(&dio->bio_lock, flags);
346 if (remaining == 0) {
348 * Defer completion when defer_completion is set or
349 * when the inode has pages mapped and this is AIO write.
350 * We need to invalidate those pages because there is a
351 * chance they contain stale data in the case buffered IO
352 * went in between AIO submission and completion into the
356 defer_completion = dio->defer_completion ||
357 (dio->op == REQ_OP_WRITE &&
358 dio->inode->i_mapping->nrpages);
359 if (defer_completion) {
360 INIT_WORK(&dio->complete_work, dio_aio_complete_work);
361 queue_work(dio->inode->i_sb->s_dio_done_wq,
362 &dio->complete_work);
364 dio_complete(dio, 0, DIO_COMPLETE_ASYNC);
370 * The BIO completion handler simply queues the BIO up for the process-context
373 * During I/O bi_private points at the dio. After I/O, bi_private is used to
374 * implement a singly-linked list of completed BIOs, at dio->bio_list.
376 static void dio_bio_end_io(struct bio *bio)
378 struct dio *dio = bio->bi_private;
381 spin_lock_irqsave(&dio->bio_lock, flags);
382 bio->bi_private = dio->bio_list;
384 if (--dio->refcount == 1 && dio->waiter)
385 wake_up_process(dio->waiter);
386 spin_unlock_irqrestore(&dio->bio_lock, flags);
390 dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
391 struct block_device *bdev,
392 sector_t first_sector, int nr_vecs)
397 * bio_alloc() is guaranteed to return a bio when allowed to sleep and
398 * we request a valid number of vectors.
400 bio = bio_alloc(GFP_KERNEL, nr_vecs);
402 bio_set_dev(bio, bdev);
403 bio->bi_iter.bi_sector = first_sector;
404 bio_set_op_attrs(bio, dio->op, dio->op_flags);
406 bio->bi_end_io = dio_bio_end_aio;
408 bio->bi_end_io = dio_bio_end_io;
410 bio->bi_write_hint = dio->iocb->ki_hint;
413 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
417 * In the AIO read case we speculatively dirty the pages before starting IO.
418 * During IO completion, any of these pages which happen to have been written
419 * back will be redirtied by bio_check_pages_dirty().
421 * bios hold a dio reference between submit_bio and ->end_io.
423 static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
425 struct bio *bio = sdio->bio;
428 bio->bi_private = dio;
430 spin_lock_irqsave(&dio->bio_lock, flags);
432 spin_unlock_irqrestore(&dio->bio_lock, flags);
434 if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty)
435 bio_set_pages_dirty(bio);
437 dio->bio_disk = bio->bi_disk;
439 if (sdio->submit_io) {
440 sdio->submit_io(bio, dio->inode, sdio->logical_offset_in_bio);
441 dio->bio_cookie = BLK_QC_T_NONE;
443 dio->bio_cookie = submit_bio(bio);
447 sdio->logical_offset_in_bio = 0;
451 * Release any resources in case of a failure
453 static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
455 while (sdio->head < sdio->tail)
456 put_page(dio->pages[sdio->head++]);
460 * Wait for the next BIO to complete. Remove it and return it. NULL is
461 * returned once all BIOs have been completed. This must only be called once
462 * all bios have been issued so that dio->refcount can only decrease. This
463 * requires that that the caller hold a reference on the dio.
465 static struct bio *dio_await_one(struct dio *dio)
468 struct bio *bio = NULL;
470 spin_lock_irqsave(&dio->bio_lock, flags);
473 * Wait as long as the list is empty and there are bios in flight. bio
474 * completion drops the count, maybe adds to the list, and wakes while
475 * holding the bio_lock so we don't need set_current_state()'s barrier
476 * and can call it after testing our condition.
478 while (dio->refcount > 1 && dio->bio_list == NULL) {
479 __set_current_state(TASK_UNINTERRUPTIBLE);
480 dio->waiter = current;
481 spin_unlock_irqrestore(&dio->bio_lock, flags);
482 if (!(dio->iocb->ki_flags & IOCB_HIPRI) ||
483 !blk_poll(dio->bio_disk->queue, dio->bio_cookie, true))
485 /* wake up sets us TASK_RUNNING */
486 spin_lock_irqsave(&dio->bio_lock, flags);
491 dio->bio_list = bio->bi_private;
493 spin_unlock_irqrestore(&dio->bio_lock, flags);
498 * Process one completed BIO. No locks are held.
500 static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio)
502 blk_status_t err = bio->bi_status;
503 bool should_dirty = dio->op == REQ_OP_READ && dio->should_dirty;
506 if (err == BLK_STS_AGAIN && (bio->bi_opf & REQ_NOWAIT))
507 dio->io_error = -EAGAIN;
509 dio->io_error = -EIO;
512 if (dio->is_async && should_dirty) {
513 bio_check_pages_dirty(bio); /* transfers ownership */
515 bio_release_pages(bio, should_dirty);
522 * Wait on and process all in-flight BIOs. This must only be called once
523 * all bios have been issued so that the refcount can only decrease.
524 * This just waits for all bios to make it through dio_bio_complete. IO
525 * errors are propagated through dio->io_error and should be propagated via
528 static void dio_await_completion(struct dio *dio)
532 bio = dio_await_one(dio);
534 dio_bio_complete(dio, bio);
539 * A really large O_DIRECT read or write can generate a lot of BIOs. So
540 * to keep the memory consumption sane we periodically reap any completed BIOs
541 * during the BIO generation phase.
543 * This also helps to limit the peak amount of pinned userspace memory.
545 static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
549 if (sdio->reap_counter++ >= 64) {
550 while (dio->bio_list) {
555 spin_lock_irqsave(&dio->bio_lock, flags);
557 dio->bio_list = bio->bi_private;
558 spin_unlock_irqrestore(&dio->bio_lock, flags);
559 ret2 = blk_status_to_errno(dio_bio_complete(dio, bio));
563 sdio->reap_counter = 0;
569 * Create workqueue for deferred direct IO completions. We allocate the
570 * workqueue when it's first needed. This avoids creating workqueue for
571 * filesystems that don't need it and also allows us to create the workqueue
572 * late enough so the we can include s_id in the name of the workqueue.
574 int sb_init_dio_done_wq(struct super_block *sb)
576 struct workqueue_struct *old;
577 struct workqueue_struct *wq = alloc_workqueue("dio/%s",
583 * This has to be atomic as more DIOs can race to create the workqueue
585 old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
586 /* Someone created workqueue before us? Free ours... */
588 destroy_workqueue(wq);
592 static int dio_set_defer_completion(struct dio *dio)
594 struct super_block *sb = dio->inode->i_sb;
596 if (dio->defer_completion)
598 dio->defer_completion = true;
599 if (!sb->s_dio_done_wq)
600 return sb_init_dio_done_wq(sb);
605 * Call into the fs to map some more disk blocks. We record the current number
606 * of available blocks at sdio->blocks_available. These are in units of the
607 * fs blocksize, i_blocksize(inode).
609 * The fs is allowed to map lots of blocks at once. If it wants to do that,
610 * it uses the passed inode-relative block number as the file offset, as usual.
612 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
613 * has remaining to do. The fs should not map more than this number of blocks.
615 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
616 * indicate how much contiguous disk space has been made available at
619 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
620 * This isn't very efficient...
622 * In the case of filesystem holes: the fs may return an arbitrarily-large
623 * hole by returning an appropriate value in b_size and by clearing
624 * buffer_mapped(). However the direct-io code will only process holes one
625 * block at a time - it will repeatedly call get_block() as it walks the hole.
627 static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
628 struct buffer_head *map_bh)
631 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
632 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
633 unsigned long fs_count; /* Number of filesystem-sized blocks */
635 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
639 * If there was a memory error and we've overwritten all the
640 * mapped blocks then we can now return that memory error
642 ret = dio->page_errors;
644 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
645 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
646 fs_endblk = (sdio->final_block_in_request - 1) >>
648 fs_count = fs_endblk - fs_startblk + 1;
651 map_bh->b_size = fs_count << i_blkbits;
654 * For writes that could fill holes inside i_size on a
655 * DIO_SKIP_HOLES filesystem we forbid block creations: only
656 * overwrites are permitted. We will return early to the caller
657 * once we see an unmapped buffer head returned, and the caller
658 * will fall back to buffered I/O.
660 * Otherwise the decision is left to the get_blocks method,
661 * which may decide to handle it or also return an unmapped
664 create = dio->op == REQ_OP_WRITE;
665 if (dio->flags & DIO_SKIP_HOLES) {
666 i_size = i_size_read(dio->inode);
667 if (i_size && fs_startblk <= (i_size - 1) >> i_blkbits)
671 ret = (*sdio->get_block)(dio->inode, fs_startblk,
674 /* Store for completion */
675 dio->private = map_bh->b_private;
677 if (ret == 0 && buffer_defer_completion(map_bh))
678 ret = dio_set_defer_completion(dio);
684 * There is no bio. Make one now.
686 static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
687 sector_t start_sector, struct buffer_head *map_bh)
692 ret = dio_bio_reap(dio, sdio);
695 sector = start_sector << (sdio->blkbits - 9);
696 nr_pages = min(sdio->pages_in_io, BIO_MAX_PAGES);
697 BUG_ON(nr_pages <= 0);
698 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
705 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
706 * that was successful then update final_block_in_bio and take a ref against
707 * the just-added page.
709 * Return zero on success. Non-zero means the caller needs to start a new BIO.
711 static inline int dio_bio_add_page(struct dio_submit *sdio)
715 ret = bio_add_page(sdio->bio, sdio->cur_page,
716 sdio->cur_page_len, sdio->cur_page_offset);
717 if (ret == sdio->cur_page_len) {
719 * Decrement count only, if we are done with this page
721 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
723 get_page(sdio->cur_page);
724 sdio->final_block_in_bio = sdio->cur_page_block +
725 (sdio->cur_page_len >> sdio->blkbits);
734 * Put cur_page under IO. The section of cur_page which is described by
735 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
736 * starts on-disk at cur_page_block.
738 * We take a ref against the page here (on behalf of its presence in the bio).
740 * The caller of this function is responsible for removing cur_page from the
741 * dio, and for dropping the refcount which came from that presence.
743 static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
744 struct buffer_head *map_bh)
749 loff_t cur_offset = sdio->cur_page_fs_offset;
750 loff_t bio_next_offset = sdio->logical_offset_in_bio +
751 sdio->bio->bi_iter.bi_size;
754 * See whether this new request is contiguous with the old.
756 * Btrfs cannot handle having logically non-contiguous requests
757 * submitted. For example if you have
759 * Logical: [0-4095][HOLE][8192-12287]
760 * Physical: [0-4095] [4096-8191]
762 * We cannot submit those pages together as one BIO. So if our
763 * current logical offset in the file does not equal what would
764 * be the next logical offset in the bio, submit the bio we
767 if (sdio->final_block_in_bio != sdio->cur_page_block ||
768 cur_offset != bio_next_offset)
769 dio_bio_submit(dio, sdio);
772 if (sdio->bio == NULL) {
773 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
778 if (dio_bio_add_page(sdio) != 0) {
779 dio_bio_submit(dio, sdio);
780 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
782 ret = dio_bio_add_page(sdio);
791 * An autonomous function to put a chunk of a page under deferred IO.
793 * The caller doesn't actually know (or care) whether this piece of page is in
794 * a BIO, or is under IO or whatever. We just take care of all possible
795 * situations here. The separation between the logic of do_direct_IO() and
796 * that of submit_page_section() is important for clarity. Please don't break.
798 * The chunk of page starts on-disk at blocknr.
800 * We perform deferred IO, by recording the last-submitted page inside our
801 * private part of the dio structure. If possible, we just expand the IO
802 * across that page here.
804 * If that doesn't work out then we put the old page into the bio and add this
805 * page to the dio instead.
808 submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
809 unsigned offset, unsigned len, sector_t blocknr,
810 struct buffer_head *map_bh)
814 if (dio->op == REQ_OP_WRITE) {
816 * Read accounting is performed in submit_bio()
818 task_io_account_write(len);
822 * Can we just grow the current page's presence in the dio?
824 if (sdio->cur_page == page &&
825 sdio->cur_page_offset + sdio->cur_page_len == offset &&
826 sdio->cur_page_block +
827 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
828 sdio->cur_page_len += len;
833 * If there's a deferred page already there then send it.
835 if (sdio->cur_page) {
836 ret = dio_send_cur_page(dio, sdio, map_bh);
837 put_page(sdio->cur_page);
838 sdio->cur_page = NULL;
843 get_page(page); /* It is in dio */
844 sdio->cur_page = page;
845 sdio->cur_page_offset = offset;
846 sdio->cur_page_len = len;
847 sdio->cur_page_block = blocknr;
848 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
851 * If sdio->boundary then we want to schedule the IO now to
852 * avoid metadata seeks.
854 if (sdio->boundary) {
855 ret = dio_send_cur_page(dio, sdio, map_bh);
857 dio_bio_submit(dio, sdio);
858 put_page(sdio->cur_page);
859 sdio->cur_page = NULL;
865 * If we are not writing the entire block and get_block() allocated
866 * the block for us, we need to fill-in the unused portion of the
867 * block with zeros. This happens only if user-buffer, fileoffset or
868 * io length is not filesystem block-size multiple.
870 * `end' is zero if we're doing the start of the IO, 1 at the end of the
873 static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
874 int end, struct buffer_head *map_bh)
876 unsigned dio_blocks_per_fs_block;
877 unsigned this_chunk_blocks; /* In dio_blocks */
878 unsigned this_chunk_bytes;
881 sdio->start_zero_done = 1;
882 if (!sdio->blkfactor || !buffer_new(map_bh))
885 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
886 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
888 if (!this_chunk_blocks)
892 * We need to zero out part of an fs block. It is either at the
893 * beginning or the end of the fs block.
896 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
898 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
901 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
902 sdio->next_block_for_io, map_bh))
905 sdio->next_block_for_io += this_chunk_blocks;
909 * Walk the user pages, and the file, mapping blocks to disk and generating
910 * a sequence of (page,offset,len,block) mappings. These mappings are injected
911 * into submit_page_section(), which takes care of the next stage of submission
913 * Direct IO against a blockdev is different from a file. Because we can
914 * happily perform page-sized but 512-byte aligned IOs. It is important that
915 * blockdev IO be able to have fine alignment and large sizes.
917 * So what we do is to permit the ->get_block function to populate bh.b_size
918 * with the size of IO which is permitted at this offset and this i_blkbits.
920 * For best results, the blockdev should be set up with 512-byte i_blkbits and
921 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
922 * fine alignment but still allows this function to work in PAGE_SIZE units.
924 static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
925 struct buffer_head *map_bh)
927 const unsigned blkbits = sdio->blkbits;
928 const unsigned i_blkbits = blkbits + sdio->blkfactor;
931 while (sdio->block_in_file < sdio->final_block_in_request) {
935 page = dio_get_page(dio, sdio);
940 from = sdio->head ? 0 : sdio->from;
941 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
945 unsigned this_chunk_bytes; /* # of bytes mapped */
946 unsigned this_chunk_blocks; /* # of blocks */
949 if (sdio->blocks_available == 0) {
951 * Need to go and map some more disk
953 unsigned long blkmask;
954 unsigned long dio_remainder;
956 ret = get_more_blocks(dio, sdio, map_bh);
961 if (!buffer_mapped(map_bh))
964 sdio->blocks_available =
965 map_bh->b_size >> blkbits;
966 sdio->next_block_for_io =
967 map_bh->b_blocknr << sdio->blkfactor;
968 if (buffer_new(map_bh)) {
972 map_bh->b_size >> i_blkbits);
975 if (!sdio->blkfactor)
978 blkmask = (1 << sdio->blkfactor) - 1;
979 dio_remainder = (sdio->block_in_file & blkmask);
982 * If we are at the start of IO and that IO
983 * starts partway into a fs-block,
984 * dio_remainder will be non-zero. If the IO
985 * is a read then we can simply advance the IO
986 * cursor to the first block which is to be
987 * read. But if the IO is a write and the
988 * block was newly allocated we cannot do that;
989 * the start of the fs block must be zeroed out
992 if (!buffer_new(map_bh))
993 sdio->next_block_for_io += dio_remainder;
994 sdio->blocks_available -= dio_remainder;
998 if (!buffer_mapped(map_bh)) {
999 loff_t i_size_aligned;
1001 /* AKPM: eargh, -ENOTBLK is a hack */
1002 if (dio->op == REQ_OP_WRITE) {
1008 * Be sure to account for a partial block as the
1009 * last block in the file
1011 i_size_aligned = ALIGN(i_size_read(dio->inode),
1013 if (sdio->block_in_file >=
1014 i_size_aligned >> blkbits) {
1019 zero_user(page, from, 1 << blkbits);
1020 sdio->block_in_file++;
1021 from += 1 << blkbits;
1022 dio->result += 1 << blkbits;
1027 * If we're performing IO which has an alignment which
1028 * is finer than the underlying fs, go check to see if
1029 * we must zero out the start of this block.
1031 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
1032 dio_zero_block(dio, sdio, 0, map_bh);
1035 * Work out, in this_chunk_blocks, how much disk we
1036 * can add to this page
1038 this_chunk_blocks = sdio->blocks_available;
1039 u = (to - from) >> blkbits;
1040 if (this_chunk_blocks > u)
1041 this_chunk_blocks = u;
1042 u = sdio->final_block_in_request - sdio->block_in_file;
1043 if (this_chunk_blocks > u)
1044 this_chunk_blocks = u;
1045 this_chunk_bytes = this_chunk_blocks << blkbits;
1046 BUG_ON(this_chunk_bytes == 0);
1048 if (this_chunk_blocks == sdio->blocks_available)
1049 sdio->boundary = buffer_boundary(map_bh);
1050 ret = submit_page_section(dio, sdio, page,
1053 sdio->next_block_for_io,
1059 sdio->next_block_for_io += this_chunk_blocks;
1061 sdio->block_in_file += this_chunk_blocks;
1062 from += this_chunk_bytes;
1063 dio->result += this_chunk_bytes;
1064 sdio->blocks_available -= this_chunk_blocks;
1066 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1067 if (sdio->block_in_file == sdio->final_block_in_request)
1071 /* Drop the ref which was taken in get_user_pages() */
1078 static inline int drop_refcount(struct dio *dio)
1081 unsigned long flags;
1084 * Sync will always be dropping the final ref and completing the
1085 * operation. AIO can if it was a broken operation described above or
1086 * in fact if all the bios race to complete before we get here. In
1087 * that case dio_complete() translates the EIOCBQUEUED into the proper
1088 * return code that the caller will hand to ->complete().
1090 * This is managed by the bio_lock instead of being an atomic_t so that
1091 * completion paths can drop their ref and use the remaining count to
1092 * decide to wake the submission path atomically.
1094 spin_lock_irqsave(&dio->bio_lock, flags);
1095 ret2 = --dio->refcount;
1096 spin_unlock_irqrestore(&dio->bio_lock, flags);
1101 * This is a library function for use by filesystem drivers.
1103 * The locking rules are governed by the flags parameter:
1104 * - if the flags value contains DIO_LOCKING we use a fancy locking
1105 * scheme for dumb filesystems.
1106 * For writes this function is called under i_mutex and returns with
1107 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1108 * taken and dropped again before returning.
1109 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1110 * internal locking but rather rely on the filesystem to synchronize
1111 * direct I/O reads/writes versus each other and truncate.
1113 * To help with locking against truncate we incremented the i_dio_count
1114 * counter before starting direct I/O, and decrement it once we are done.
1115 * Truncate can wait for it to reach zero to provide exclusion. It is
1116 * expected that filesystem provide exclusion between new direct I/O
1117 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1118 * but other filesystems need to take care of this on their own.
1120 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1121 * is always inlined. Otherwise gcc is unable to split the structure into
1122 * individual fields and will generate much worse code. This is important
1123 * for the whole file.
1125 static inline ssize_t
1126 do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1127 struct block_device *bdev, struct iov_iter *iter,
1128 get_block_t get_block, dio_iodone_t end_io,
1129 dio_submit_t submit_io, int flags)
1131 unsigned i_blkbits = READ_ONCE(inode->i_blkbits);
1132 unsigned blkbits = i_blkbits;
1133 unsigned blocksize_mask = (1 << blkbits) - 1;
1134 ssize_t retval = -EINVAL;
1135 const size_t count = iov_iter_count(iter);
1136 loff_t offset = iocb->ki_pos;
1137 const loff_t end = offset + count;
1139 struct dio_submit sdio = { 0, };
1140 struct buffer_head map_bh = { 0, };
1141 struct blk_plug plug;
1142 unsigned long align = offset | iov_iter_alignment(iter);
1145 * Avoid references to bdev if not absolutely needed to give
1146 * the early prefetch in the caller enough time.
1149 /* watch out for a 0 len io from a tricksy fs */
1150 if (iov_iter_rw(iter) == READ && !count)
1153 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
1157 * Believe it or not, zeroing out the page array caused a .5%
1158 * performance regression in a database benchmark. So, we take
1159 * care to only zero out what's needed.
1161 memset(dio, 0, offsetof(struct dio, pages));
1164 if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ) {
1165 /* will be released by direct_io_worker */
1169 /* Once we sampled i_size check for reads beyond EOF */
1170 dio->i_size = i_size_read(inode);
1171 if (iov_iter_rw(iter) == READ && offset >= dio->i_size) {
1176 if (align & blocksize_mask) {
1178 blkbits = blksize_bits(bdev_logical_block_size(bdev));
1179 blocksize_mask = (1 << blkbits) - 1;
1180 if (align & blocksize_mask)
1184 if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ) {
1185 struct address_space *mapping = iocb->ki_filp->f_mapping;
1187 retval = filemap_write_and_wait_range(mapping, offset, end - 1);
1193 * For file extending writes updating i_size before data writeouts
1194 * complete can expose uninitialized blocks in dumb filesystems.
1195 * In that case we need to wait for I/O completion even if asked
1196 * for an asynchronous write.
1198 if (is_sync_kiocb(iocb))
1199 dio->is_async = false;
1200 else if (iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
1201 dio->is_async = false;
1203 dio->is_async = true;
1206 if (iov_iter_rw(iter) == WRITE) {
1207 dio->op = REQ_OP_WRITE;
1208 dio->op_flags = REQ_SYNC | REQ_IDLE;
1209 if (iocb->ki_flags & IOCB_NOWAIT)
1210 dio->op_flags |= REQ_NOWAIT;
1212 dio->op = REQ_OP_READ;
1214 if (iocb->ki_flags & IOCB_HIPRI)
1215 dio->op_flags |= REQ_HIPRI;
1218 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1219 * so that we can call ->fsync.
1221 if (dio->is_async && iov_iter_rw(iter) == WRITE) {
1223 if (iocb->ki_flags & IOCB_DSYNC)
1224 retval = dio_set_defer_completion(dio);
1225 else if (!dio->inode->i_sb->s_dio_done_wq) {
1227 * In case of AIO write racing with buffered read we
1228 * need to defer completion. We can't decide this now,
1229 * however the workqueue needs to be initialized here.
1231 retval = sb_init_dio_done_wq(dio->inode->i_sb);
1238 * Will be decremented at I/O completion time.
1240 inode_dio_begin(inode);
1243 sdio.blkbits = blkbits;
1244 sdio.blkfactor = i_blkbits - blkbits;
1245 sdio.block_in_file = offset >> blkbits;
1247 sdio.get_block = get_block;
1248 dio->end_io = end_io;
1249 sdio.submit_io = submit_io;
1250 sdio.final_block_in_bio = -1;
1251 sdio.next_block_for_io = -1;
1255 spin_lock_init(&dio->bio_lock);
1258 dio->should_dirty = iter_is_iovec(iter) && iov_iter_rw(iter) == READ;
1260 sdio.final_block_in_request = end >> blkbits;
1263 * In case of non-aligned buffers, we may need 2 more
1264 * pages since we need to zero out first and last block.
1266 if (unlikely(sdio.blkfactor))
1267 sdio.pages_in_io = 2;
1269 sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
1271 blk_start_plug(&plug);
1273 retval = do_direct_IO(dio, &sdio, &map_bh);
1275 dio_cleanup(dio, &sdio);
1277 if (retval == -ENOTBLK) {
1279 * The remaining part of the request will be
1280 * be handled by buffered I/O when we return
1285 * There may be some unwritten disk at the end of a part-written
1286 * fs-block-sized block. Go zero that now.
1288 dio_zero_block(dio, &sdio, 1, &map_bh);
1290 if (sdio.cur_page) {
1293 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1296 put_page(sdio.cur_page);
1297 sdio.cur_page = NULL;
1300 dio_bio_submit(dio, &sdio);
1302 blk_finish_plug(&plug);
1305 * It is possible that, we return short IO due to end of file.
1306 * In that case, we need to release all the pages we got hold on.
1308 dio_cleanup(dio, &sdio);
1311 * All block lookups have been performed. For READ requests
1312 * we can let i_mutex go now that its achieved its purpose
1313 * of protecting us from looking up uninitialized blocks.
1315 if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING))
1316 inode_unlock(dio->inode);
1319 * The only time we want to leave bios in flight is when a successful
1320 * partial aio read or full aio write have been setup. In that case
1321 * bio completion will call aio_complete. The only time it's safe to
1322 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1323 * This had *better* be the only place that raises -EIOCBQUEUED.
1325 BUG_ON(retval == -EIOCBQUEUED);
1326 if (dio->is_async && retval == 0 && dio->result &&
1327 (iov_iter_rw(iter) == READ || dio->result == count))
1328 retval = -EIOCBQUEUED;
1330 dio_await_completion(dio);
1332 if (drop_refcount(dio) == 0) {
1333 retval = dio_complete(dio, retval, DIO_COMPLETE_INVALIDATE);
1335 BUG_ON(retval != -EIOCBQUEUED);
1340 if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ)
1341 inode_unlock(inode);
1343 kmem_cache_free(dio_cache, dio);
1347 ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1348 struct block_device *bdev, struct iov_iter *iter,
1349 get_block_t get_block,
1350 dio_iodone_t end_io, dio_submit_t submit_io,
1354 * The block device state is needed in the end to finally
1355 * submit everything. Since it's likely to be cache cold
1356 * prefetch it here as first thing to hide some of the
1359 * Attempt to prefetch the pieces we likely need later.
1361 prefetch(&bdev->bd_disk->part_tbl);
1362 prefetch(bdev->bd_disk->queue);
1363 prefetch((char *)bdev->bd_disk->queue + SMP_CACHE_BYTES);
1365 return do_blockdev_direct_IO(iocb, inode, bdev, iter, get_block,
1366 end_io, submit_io, flags);
1369 EXPORT_SYMBOL(__blockdev_direct_IO);
1371 static __init int dio_init(void)
1373 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1376 module_init(dio_init)