4 * Copyright (C) 2002, Linus Torvalds.
8 * 04Jul2002 Andrew Morton
10 * 11Sep2002 janetinc@us.ibm.com
11 * added readv/writev support.
12 * 29Oct2002 Andrew Morton
13 * rewrote bio_add_page() support.
14 * 30Oct2002 pbadari@us.ibm.com
15 * added support for non-aligned IO.
16 * 06Nov2002 pbadari@us.ibm.com
17 * added asynchronous IO support.
18 * 21Jul2003 nathans@sgi.com
19 * added IO completion notifier.
22 #include <linux/kernel.h>
23 #include <linux/module.h>
24 #include <linux/types.h>
27 #include <linux/slab.h>
28 #include <linux/highmem.h>
29 #include <linux/pagemap.h>
30 #include <linux/task_io_accounting_ops.h>
31 #include <linux/bio.h>
32 #include <linux/wait.h>
33 #include <linux/err.h>
34 #include <linux/blkdev.h>
35 #include <linux/buffer_head.h>
36 #include <linux/rwsem.h>
37 #include <linux/uio.h>
38 #include <linux/atomic.h>
41 * How many user pages to map in one call to get_user_pages(). This determines
42 * the size of a structure on the stack.
47 * This code generally works in units of "dio_blocks". A dio_block is
48 * somewhere between the hard sector size and the filesystem block size. it
49 * is determined on a per-invocation basis. When talking to the filesystem
50 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
51 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
52 * to bio_block quantities by shifting left by blkfactor.
54 * If blkfactor is zero then the user's request was aligned to the filesystem's
58 /* dio_state only used in the submission path */
61 struct bio *bio; /* bio under assembly */
62 unsigned blkbits; /* doesn't change */
63 unsigned blkfactor; /* When we're using an alignment which
64 is finer than the filesystem's soft
65 blocksize, this specifies how much
66 finer. blkfactor=2 means 1/4-block
67 alignment. Does not change */
68 unsigned start_zero_done; /* flag: sub-blocksize zeroing has
69 been performed at the start of a
71 int pages_in_io; /* approximate total IO pages */
72 size_t size; /* total request size (doesn't change)*/
73 sector_t block_in_file; /* Current offset into the underlying
74 file in dio_block units. */
75 unsigned blocks_available; /* At block_in_file. changes */
76 sector_t final_block_in_request;/* doesn't change */
77 unsigned first_block_in_page; /* doesn't change, Used only once */
78 int boundary; /* prev block is at a boundary */
79 int reap_counter; /* rate limit reaping */
80 get_block_t *get_block; /* block mapping function */
81 dio_submit_t *submit_io; /* IO submition function */
83 loff_t logical_offset_in_bio; /* current first logical block in bio */
84 sector_t final_block_in_bio; /* current final block in bio + 1 */
85 sector_t next_block_for_io; /* next block to be put under IO,
86 in dio_blocks units */
89 * Deferred addition of a page to the dio. These variables are
90 * private to dio_send_cur_page(), submit_page_section() and
93 struct page *cur_page; /* The page */
94 unsigned cur_page_offset; /* Offset into it, in bytes */
95 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
96 sector_t cur_page_block; /* Where it starts */
97 loff_t cur_page_fs_offset; /* Offset in file */
100 * Page fetching state. These variables belong to dio_refill_pages().
102 int curr_page; /* changes */
103 int total_pages; /* doesn't change */
104 unsigned long curr_user_address;/* changes */
107 * Page queue. These variables belong to dio_refill_pages() and
110 unsigned head; /* next page to process */
111 unsigned tail; /* last valid page + 1 */
114 /* dio_state communicated between submission path and end_io */
116 int flags; /* doesn't change */
119 loff_t i_size; /* i_size when submitted */
120 dio_iodone_t *end_io; /* IO completion function */
121 struct buffer_head map_bh; /* last get_block() result */
124 /* BIO completion state */
125 spinlock_t bio_lock; /* protects BIO fields below */
126 unsigned long refcount; /* direct_io_worker() and bios */
127 struct bio *bio_list; /* singly linked via bi_private */
128 struct task_struct *waiter; /* waiting task (NULL if none) */
130 /* AIO related stuff */
131 struct kiocb *iocb; /* kiocb */
132 int is_async; /* is IO async ? */
133 int io_error; /* IO error in completion path */
134 ssize_t result; /* IO result */
136 int page_errors; /* errno from get_user_pages() */
139 * pages[] (and any fields placed after it) are not zeroed out at
140 * allocation time. Don't add new fields after pages[] unless you
141 * wish that they not be zeroed.
143 struct page *pages[DIO_PAGES]; /* page buffer */
146 static void __inode_dio_wait(struct inode *inode)
148 wait_queue_head_t *wq = bit_waitqueue(&inode->i_state, __I_DIO_WAKEUP);
149 DEFINE_WAIT_BIT(q, &inode->i_state, __I_DIO_WAKEUP);
152 prepare_to_wait(wq, &q.wait, TASK_UNINTERRUPTIBLE);
153 if (atomic_read(&inode->i_dio_count))
155 } while (atomic_read(&inode->i_dio_count));
156 finish_wait(wq, &q.wait);
160 * inode_dio_wait - wait for outstanding DIO requests to finish
161 * @inode: inode to wait for
163 * Waits for all pending direct I/O requests to finish so that we can
164 * proceed with a truncate or equivalent operation.
166 * Must be called under a lock that serializes taking new references
167 * to i_dio_count, usually by inode->i_mutex.
169 void inode_dio_wait(struct inode *inode)
171 if (atomic_read(&inode->i_dio_count))
172 __inode_dio_wait(inode);
174 EXPORT_SYMBOL_GPL(inode_dio_wait);
177 * inode_dio_done - signal finish of a direct I/O requests
178 * @inode: inode the direct I/O happens on
180 * This is called once we've finished processing a direct I/O request,
181 * and is used to wake up callers waiting for direct I/O to be quiesced.
183 void inode_dio_done(struct inode *inode)
185 if (atomic_dec_and_test(&inode->i_dio_count))
186 wake_up_bit(&inode->i_state, __I_DIO_WAKEUP);
188 EXPORT_SYMBOL_GPL(inode_dio_done);
191 * How many pages are in the queue?
193 static inline unsigned dio_pages_present(struct dio *dio, struct dio_submit *sdio)
195 return sdio->tail - sdio->head;
199 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
201 static int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
206 nr_pages = min(sdio->total_pages - sdio->curr_page, DIO_PAGES);
207 ret = get_user_pages_fast(
208 sdio->curr_user_address, /* Where from? */
209 nr_pages, /* How many pages? */
210 dio->rw == READ, /* Write to memory? */
211 &dio->pages[0]); /* Put results here */
213 if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) {
214 struct page *page = ZERO_PAGE(0);
216 * A memory fault, but the filesystem has some outstanding
217 * mapped blocks. We need to use those blocks up to avoid
218 * leaking stale data in the file.
220 if (dio->page_errors == 0)
221 dio->page_errors = ret;
222 page_cache_get(page);
223 dio->pages[0] = page;
231 sdio->curr_user_address += ret * PAGE_SIZE;
232 sdio->curr_page += ret;
242 * Get another userspace page. Returns an ERR_PTR on error. Pages are
243 * buffered inside the dio so that we can call get_user_pages() against a
244 * decent number of pages, less frequently. To provide nicer use of the
247 static struct page *dio_get_page(struct dio *dio, struct dio_submit *sdio)
249 if (dio_pages_present(dio, sdio) == 0) {
252 ret = dio_refill_pages(dio, sdio);
255 BUG_ON(dio_pages_present(dio, sdio) == 0);
257 return dio->pages[sdio->head++];
261 * dio_complete() - called when all DIO BIO I/O has been completed
262 * @offset: the byte offset in the file of the completed operation
264 * This releases locks as dictated by the locking type, lets interested parties
265 * know that a DIO operation has completed, and calculates the resulting return
266 * code for the operation.
268 * It lets the filesystem know if it registered an interest earlier via
269 * get_block. Pass the private field of the map buffer_head so that
270 * filesystems can use it to hold additional state between get_block calls and
273 static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret, bool is_async)
275 ssize_t transferred = 0;
278 * AIO submission can race with bio completion to get here while
279 * expecting to have the last io completed by bio completion.
280 * In that case -EIOCBQUEUED is in fact not an error we want
281 * to preserve through this call.
283 if (ret == -EIOCBQUEUED)
287 transferred = dio->result;
289 /* Check for short read case */
290 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
291 transferred = dio->i_size - offset;
295 ret = dio->page_errors;
301 if (dio->end_io && dio->result) {
302 dio->end_io(dio->iocb, offset, transferred,
303 dio->map_bh.b_private, ret, is_async);
306 aio_complete(dio->iocb, ret, 0);
307 inode_dio_done(dio->inode);
313 static int dio_bio_complete(struct dio *dio, struct bio *bio);
315 * Asynchronous IO callback.
317 static void dio_bio_end_aio(struct bio *bio, int error)
319 struct dio *dio = bio->bi_private;
320 unsigned long remaining;
323 /* cleanup the bio */
324 dio_bio_complete(dio, bio);
326 spin_lock_irqsave(&dio->bio_lock, flags);
327 remaining = --dio->refcount;
328 if (remaining == 1 && dio->waiter)
329 wake_up_process(dio->waiter);
330 spin_unlock_irqrestore(&dio->bio_lock, flags);
332 if (remaining == 0) {
333 dio_complete(dio, dio->iocb->ki_pos, 0, true);
339 * The BIO completion handler simply queues the BIO up for the process-context
342 * During I/O bi_private points at the dio. After I/O, bi_private is used to
343 * implement a singly-linked list of completed BIOs, at dio->bio_list.
345 static void dio_bio_end_io(struct bio *bio, int error)
347 struct dio *dio = bio->bi_private;
350 spin_lock_irqsave(&dio->bio_lock, flags);
351 bio->bi_private = dio->bio_list;
353 if (--dio->refcount == 1 && dio->waiter)
354 wake_up_process(dio->waiter);
355 spin_unlock_irqrestore(&dio->bio_lock, flags);
359 * dio_end_io - handle the end io action for the given bio
360 * @bio: The direct io bio thats being completed
361 * @error: Error if there was one
363 * This is meant to be called by any filesystem that uses their own dio_submit_t
364 * so that the DIO specific endio actions are dealt with after the filesystem
365 * has done it's completion work.
367 void dio_end_io(struct bio *bio, int error)
369 struct dio *dio = bio->bi_private;
372 dio_bio_end_aio(bio, error);
374 dio_bio_end_io(bio, error);
376 EXPORT_SYMBOL_GPL(dio_end_io);
379 dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
380 struct block_device *bdev,
381 sector_t first_sector, int nr_vecs)
386 * bio_alloc() is guaranteed to return a bio when called with
387 * __GFP_WAIT and we request a valid number of vectors.
389 bio = bio_alloc(GFP_KERNEL, nr_vecs);
392 bio->bi_sector = first_sector;
394 bio->bi_end_io = dio_bio_end_aio;
396 bio->bi_end_io = dio_bio_end_io;
399 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
403 * In the AIO read case we speculatively dirty the pages before starting IO.
404 * During IO completion, any of these pages which happen to have been written
405 * back will be redirtied by bio_check_pages_dirty().
407 * bios hold a dio reference between submit_bio and ->end_io.
409 static void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
411 struct bio *bio = sdio->bio;
414 bio->bi_private = dio;
416 spin_lock_irqsave(&dio->bio_lock, flags);
418 spin_unlock_irqrestore(&dio->bio_lock, flags);
420 if (dio->is_async && dio->rw == READ)
421 bio_set_pages_dirty(bio);
424 sdio->submit_io(dio->rw, bio, dio->inode,
425 sdio->logical_offset_in_bio);
427 submit_bio(dio->rw, bio);
431 sdio->logical_offset_in_bio = 0;
435 * Release any resources in case of a failure
437 static void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
439 while (dio_pages_present(dio, sdio))
440 page_cache_release(dio_get_page(dio, sdio));
444 * Wait for the next BIO to complete. Remove it and return it. NULL is
445 * returned once all BIOs have been completed. This must only be called once
446 * all bios have been issued so that dio->refcount can only decrease. This
447 * requires that that the caller hold a reference on the dio.
449 static struct bio *dio_await_one(struct dio *dio)
452 struct bio *bio = NULL;
454 spin_lock_irqsave(&dio->bio_lock, flags);
457 * Wait as long as the list is empty and there are bios in flight. bio
458 * completion drops the count, maybe adds to the list, and wakes while
459 * holding the bio_lock so we don't need set_current_state()'s barrier
460 * and can call it after testing our condition.
462 while (dio->refcount > 1 && dio->bio_list == NULL) {
463 __set_current_state(TASK_UNINTERRUPTIBLE);
464 dio->waiter = current;
465 spin_unlock_irqrestore(&dio->bio_lock, flags);
467 /* wake up sets us TASK_RUNNING */
468 spin_lock_irqsave(&dio->bio_lock, flags);
473 dio->bio_list = bio->bi_private;
475 spin_unlock_irqrestore(&dio->bio_lock, flags);
480 * Process one completed BIO. No locks are held.
482 static int dio_bio_complete(struct dio *dio, struct bio *bio)
484 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
485 struct bio_vec *bvec = bio->bi_io_vec;
489 dio->io_error = -EIO;
491 if (dio->is_async && dio->rw == READ) {
492 bio_check_pages_dirty(bio); /* transfers ownership */
494 for (page_no = 0; page_no < bio->bi_vcnt; page_no++) {
495 struct page *page = bvec[page_no].bv_page;
497 if (dio->rw == READ && !PageCompound(page))
498 set_page_dirty_lock(page);
499 page_cache_release(page);
503 return uptodate ? 0 : -EIO;
507 * Wait on and process all in-flight BIOs. This must only be called once
508 * all bios have been issued so that the refcount can only decrease.
509 * This just waits for all bios to make it through dio_bio_complete. IO
510 * errors are propagated through dio->io_error and should be propagated via
513 static void dio_await_completion(struct dio *dio)
517 bio = dio_await_one(dio);
519 dio_bio_complete(dio, bio);
524 * A really large O_DIRECT read or write can generate a lot of BIOs. So
525 * to keep the memory consumption sane we periodically reap any completed BIOs
526 * during the BIO generation phase.
528 * This also helps to limit the peak amount of pinned userspace memory.
530 static int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
534 if (sdio->reap_counter++ >= 64) {
535 while (dio->bio_list) {
540 spin_lock_irqsave(&dio->bio_lock, flags);
542 dio->bio_list = bio->bi_private;
543 spin_unlock_irqrestore(&dio->bio_lock, flags);
544 ret2 = dio_bio_complete(dio, bio);
548 sdio->reap_counter = 0;
554 * Call into the fs to map some more disk blocks. We record the current number
555 * of available blocks at sdio->blocks_available. These are in units of the
556 * fs blocksize, (1 << inode->i_blkbits).
558 * The fs is allowed to map lots of blocks at once. If it wants to do that,
559 * it uses the passed inode-relative block number as the file offset, as usual.
561 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
562 * has remaining to do. The fs should not map more than this number of blocks.
564 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
565 * indicate how much contiguous disk space has been made available at
568 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
569 * This isn't very efficient...
571 * In the case of filesystem holes: the fs may return an arbitrarily-large
572 * hole by returning an appropriate value in b_size and by clearing
573 * buffer_mapped(). However the direct-io code will only process holes one
574 * block at a time - it will repeatedly call get_block() as it walks the hole.
576 static int get_more_blocks(struct dio *dio, struct dio_submit *sdio)
579 struct buffer_head *map_bh = &dio->map_bh;
580 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
581 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
582 unsigned long fs_count; /* Number of filesystem-sized blocks */
586 * If there was a memory error and we've overwritten all the
587 * mapped blocks then we can now return that memory error
589 ret = dio->page_errors;
591 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
592 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
593 fs_endblk = (sdio->final_block_in_request-1) >> sdio->blkfactor;
594 fs_count = fs_endblk - fs_startblk + 1;
597 map_bh->b_size = fs_count << dio->inode->i_blkbits;
600 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we
601 * forbid block creations: only overwrites are permitted.
602 * We will return early to the caller once we see an
603 * unmapped buffer head returned, and the caller will fall
604 * back to buffered I/O.
606 * Otherwise the decision is left to the get_blocks method,
607 * which may decide to handle it or also return an unmapped
610 create = dio->rw & WRITE;
611 if (dio->flags & DIO_SKIP_HOLES) {
612 if (sdio->block_in_file < (i_size_read(dio->inode) >>
617 ret = (*sdio->get_block)(dio->inode, fs_startblk,
624 * There is no bio. Make one now.
626 static int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
627 sector_t start_sector)
632 ret = dio_bio_reap(dio, sdio);
635 sector = start_sector << (sdio->blkbits - 9);
636 nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(dio->map_bh.b_bdev));
637 nr_pages = min(nr_pages, BIO_MAX_PAGES);
638 BUG_ON(nr_pages <= 0);
639 dio_bio_alloc(dio, sdio, dio->map_bh.b_bdev, sector, nr_pages);
646 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
647 * that was successful then update final_block_in_bio and take a ref against
648 * the just-added page.
650 * Return zero on success. Non-zero means the caller needs to start a new BIO.
652 static int dio_bio_add_page(struct dio *dio, struct dio_submit *sdio)
656 ret = bio_add_page(sdio->bio, sdio->cur_page,
657 sdio->cur_page_len, sdio->cur_page_offset);
658 if (ret == sdio->cur_page_len) {
660 * Decrement count only, if we are done with this page
662 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
664 page_cache_get(sdio->cur_page);
665 sdio->final_block_in_bio = sdio->cur_page_block +
666 (sdio->cur_page_len >> sdio->blkbits);
675 * Put cur_page under IO. The section of cur_page which is described by
676 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
677 * starts on-disk at cur_page_block.
679 * We take a ref against the page here (on behalf of its presence in the bio).
681 * The caller of this function is responsible for removing cur_page from the
682 * dio, and for dropping the refcount which came from that presence.
684 static int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio)
689 loff_t cur_offset = sdio->cur_page_fs_offset;
690 loff_t bio_next_offset = sdio->logical_offset_in_bio +
694 * See whether this new request is contiguous with the old.
696 * Btrfs cannot handle having logically non-contiguous requests
697 * submitted. For example if you have
699 * Logical: [0-4095][HOLE][8192-12287]
700 * Physical: [0-4095] [4096-8191]
702 * We cannot submit those pages together as one BIO. So if our
703 * current logical offset in the file does not equal what would
704 * be the next logical offset in the bio, submit the bio we
707 if (sdio->final_block_in_bio != sdio->cur_page_block ||
708 cur_offset != bio_next_offset)
709 dio_bio_submit(dio, sdio);
711 * Submit now if the underlying fs is about to perform a
714 else if (sdio->boundary)
715 dio_bio_submit(dio, sdio);
718 if (sdio->bio == NULL) {
719 ret = dio_new_bio(dio, sdio, sdio->cur_page_block);
724 if (dio_bio_add_page(dio, sdio) != 0) {
725 dio_bio_submit(dio, sdio);
726 ret = dio_new_bio(dio, sdio, sdio->cur_page_block);
728 ret = dio_bio_add_page(dio, sdio);
737 * An autonomous function to put a chunk of a page under deferred IO.
739 * The caller doesn't actually know (or care) whether this piece of page is in
740 * a BIO, or is under IO or whatever. We just take care of all possible
741 * situations here. The separation between the logic of do_direct_IO() and
742 * that of submit_page_section() is important for clarity. Please don't break.
744 * The chunk of page starts on-disk at blocknr.
746 * We perform deferred IO, by recording the last-submitted page inside our
747 * private part of the dio structure. If possible, we just expand the IO
748 * across that page here.
750 * If that doesn't work out then we put the old page into the bio and add this
751 * page to the dio instead.
754 submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
755 unsigned offset, unsigned len, sector_t blocknr)
759 if (dio->rw & WRITE) {
761 * Read accounting is performed in submit_bio()
763 task_io_account_write(len);
767 * Can we just grow the current page's presence in the dio?
769 if ((sdio->cur_page == page) &&
770 (sdio->cur_page_offset + sdio->cur_page_len == offset) &&
771 (sdio->cur_page_block +
772 (sdio->cur_page_len >> sdio->blkbits) == blocknr)) {
773 sdio->cur_page_len += len;
776 * If sdio->boundary then we want to schedule the IO now to
777 * avoid metadata seeks.
779 if (sdio->boundary) {
780 ret = dio_send_cur_page(dio, sdio);
781 page_cache_release(sdio->cur_page);
782 sdio->cur_page = NULL;
788 * If there's a deferred page already there then send it.
790 if (sdio->cur_page) {
791 ret = dio_send_cur_page(dio, sdio);
792 page_cache_release(sdio->cur_page);
793 sdio->cur_page = NULL;
798 page_cache_get(page); /* It is in dio */
799 sdio->cur_page = page;
800 sdio->cur_page_offset = offset;
801 sdio->cur_page_len = len;
802 sdio->cur_page_block = blocknr;
803 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
809 * Clean any dirty buffers in the blockdev mapping which alias newly-created
810 * file blocks. Only called for S_ISREG files - blockdevs do not set
813 static void clean_blockdev_aliases(struct dio *dio)
818 nblocks = dio->map_bh.b_size >> dio->inode->i_blkbits;
820 for (i = 0; i < nblocks; i++) {
821 unmap_underlying_metadata(dio->map_bh.b_bdev,
822 dio->map_bh.b_blocknr + i);
827 * If we are not writing the entire block and get_block() allocated
828 * the block for us, we need to fill-in the unused portion of the
829 * block with zeros. This happens only if user-buffer, fileoffset or
830 * io length is not filesystem block-size multiple.
832 * `end' is zero if we're doing the start of the IO, 1 at the end of the
835 static void dio_zero_block(struct dio *dio, struct dio_submit *sdio, int end)
837 unsigned dio_blocks_per_fs_block;
838 unsigned this_chunk_blocks; /* In dio_blocks */
839 unsigned this_chunk_bytes;
842 sdio->start_zero_done = 1;
843 if (!sdio->blkfactor || !buffer_new(&dio->map_bh))
846 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
847 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
849 if (!this_chunk_blocks)
853 * We need to zero out part of an fs block. It is either at the
854 * beginning or the end of the fs block.
857 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
859 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
862 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
863 sdio->next_block_for_io))
866 sdio->next_block_for_io += this_chunk_blocks;
870 * Walk the user pages, and the file, mapping blocks to disk and generating
871 * a sequence of (page,offset,len,block) mappings. These mappings are injected
872 * into submit_page_section(), which takes care of the next stage of submission
874 * Direct IO against a blockdev is different from a file. Because we can
875 * happily perform page-sized but 512-byte aligned IOs. It is important that
876 * blockdev IO be able to have fine alignment and large sizes.
878 * So what we do is to permit the ->get_block function to populate bh.b_size
879 * with the size of IO which is permitted at this offset and this i_blkbits.
881 * For best results, the blockdev should be set up with 512-byte i_blkbits and
882 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
883 * fine alignment but still allows this function to work in PAGE_SIZE units.
885 static int do_direct_IO(struct dio *dio, struct dio_submit *sdio)
887 const unsigned blkbits = sdio->blkbits;
888 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
890 unsigned block_in_page;
891 struct buffer_head *map_bh = &dio->map_bh;
894 /* The I/O can start at any block offset within the first page */
895 block_in_page = sdio->first_block_in_page;
897 while (sdio->block_in_file < sdio->final_block_in_request) {
898 page = dio_get_page(dio, sdio);
904 while (block_in_page < blocks_per_page) {
905 unsigned offset_in_page = block_in_page << blkbits;
906 unsigned this_chunk_bytes; /* # of bytes mapped */
907 unsigned this_chunk_blocks; /* # of blocks */
910 if (sdio->blocks_available == 0) {
912 * Need to go and map some more disk
914 unsigned long blkmask;
915 unsigned long dio_remainder;
917 ret = get_more_blocks(dio, sdio);
919 page_cache_release(page);
922 if (!buffer_mapped(map_bh))
925 sdio->blocks_available =
926 map_bh->b_size >> sdio->blkbits;
927 sdio->next_block_for_io =
928 map_bh->b_blocknr << sdio->blkfactor;
929 if (buffer_new(map_bh))
930 clean_blockdev_aliases(dio);
932 if (!sdio->blkfactor)
935 blkmask = (1 << sdio->blkfactor) - 1;
936 dio_remainder = (sdio->block_in_file & blkmask);
939 * If we are at the start of IO and that IO
940 * starts partway into a fs-block,
941 * dio_remainder will be non-zero. If the IO
942 * is a read then we can simply advance the IO
943 * cursor to the first block which is to be
944 * read. But if the IO is a write and the
945 * block was newly allocated we cannot do that;
946 * the start of the fs block must be zeroed out
949 if (!buffer_new(map_bh))
950 sdio->next_block_for_io += dio_remainder;
951 sdio->blocks_available -= dio_remainder;
955 if (!buffer_mapped(map_bh)) {
956 loff_t i_size_aligned;
958 /* AKPM: eargh, -ENOTBLK is a hack */
959 if (dio->rw & WRITE) {
960 page_cache_release(page);
965 * Be sure to account for a partial block as the
966 * last block in the file
968 i_size_aligned = ALIGN(i_size_read(dio->inode),
970 if (sdio->block_in_file >=
971 i_size_aligned >> blkbits) {
973 page_cache_release(page);
976 zero_user(page, block_in_page << blkbits,
978 sdio->block_in_file++;
984 * If we're performing IO which has an alignment which
985 * is finer than the underlying fs, go check to see if
986 * we must zero out the start of this block.
988 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
989 dio_zero_block(dio, sdio, 0);
992 * Work out, in this_chunk_blocks, how much disk we
993 * can add to this page
995 this_chunk_blocks = sdio->blocks_available;
996 u = (PAGE_SIZE - offset_in_page) >> blkbits;
997 if (this_chunk_blocks > u)
998 this_chunk_blocks = u;
999 u = sdio->final_block_in_request - sdio->block_in_file;
1000 if (this_chunk_blocks > u)
1001 this_chunk_blocks = u;
1002 this_chunk_bytes = this_chunk_blocks << blkbits;
1003 BUG_ON(this_chunk_bytes == 0);
1005 sdio->boundary = buffer_boundary(map_bh);
1006 ret = submit_page_section(dio, sdio, page, offset_in_page,
1007 this_chunk_bytes, sdio->next_block_for_io);
1009 page_cache_release(page);
1012 sdio->next_block_for_io += this_chunk_blocks;
1014 sdio->block_in_file += this_chunk_blocks;
1015 block_in_page += this_chunk_blocks;
1016 sdio->blocks_available -= this_chunk_blocks;
1018 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1019 if (sdio->block_in_file == sdio->final_block_in_request)
1023 /* Drop the ref which was taken in get_user_pages() */
1024 page_cache_release(page);
1032 direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
1033 const struct iovec *iov, loff_t offset, unsigned long nr_segs,
1034 unsigned blkbits, get_block_t get_block, dio_iodone_t end_io,
1035 dio_submit_t submit_io, struct dio *dio, struct dio_submit *sdio)
1037 unsigned long user_addr;
1038 unsigned long flags;
1046 sdio->blkbits = blkbits;
1047 sdio->blkfactor = inode->i_blkbits - blkbits;
1048 sdio->block_in_file = offset >> blkbits;
1050 sdio->get_block = get_block;
1051 dio->end_io = end_io;
1052 sdio->submit_io = submit_io;
1053 sdio->final_block_in_bio = -1;
1054 sdio->next_block_for_io = -1;
1057 dio->i_size = i_size_read(inode);
1059 spin_lock_init(&dio->bio_lock);
1063 * In case of non-aligned buffers, we may need 2 more
1064 * pages since we need to zero out first and last block.
1066 if (unlikely(sdio->blkfactor))
1067 sdio->pages_in_io = 2;
1069 for (seg = 0; seg < nr_segs; seg++) {
1070 user_addr = (unsigned long)iov[seg].iov_base;
1071 sdio->pages_in_io +=
1072 ((user_addr+iov[seg].iov_len +PAGE_SIZE-1)/PAGE_SIZE
1073 - user_addr/PAGE_SIZE);
1076 for (seg = 0; seg < nr_segs; seg++) {
1077 user_addr = (unsigned long)iov[seg].iov_base;
1078 sdio->size += bytes = iov[seg].iov_len;
1080 /* Index into the first page of the first block */
1081 sdio->first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
1082 sdio->final_block_in_request = sdio->block_in_file +
1084 /* Page fetching state */
1087 sdio->curr_page = 0;
1089 sdio->total_pages = 0;
1090 if (user_addr & (PAGE_SIZE-1)) {
1091 sdio->total_pages++;
1092 bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
1094 sdio->total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
1095 sdio->curr_user_address = user_addr;
1097 ret = do_direct_IO(dio, sdio);
1099 dio->result += iov[seg].iov_len -
1100 ((sdio->final_block_in_request - sdio->block_in_file) <<
1104 dio_cleanup(dio, sdio);
1107 } /* end iovec loop */
1109 if (ret == -ENOTBLK) {
1111 * The remaining part of the request will be
1112 * be handled by buffered I/O when we return
1117 * There may be some unwritten disk at the end of a part-written
1118 * fs-block-sized block. Go zero that now.
1120 dio_zero_block(dio, sdio, 1);
1122 if (sdio->cur_page) {
1123 ret2 = dio_send_cur_page(dio, sdio);
1126 page_cache_release(sdio->cur_page);
1127 sdio->cur_page = NULL;
1130 dio_bio_submit(dio, sdio);
1133 * It is possible that, we return short IO due to end of file.
1134 * In that case, we need to release all the pages we got hold on.
1136 dio_cleanup(dio, sdio);
1139 * All block lookups have been performed. For READ requests
1140 * we can let i_mutex go now that its achieved its purpose
1141 * of protecting us from looking up uninitialized blocks.
1143 if (rw == READ && (dio->flags & DIO_LOCKING))
1144 mutex_unlock(&dio->inode->i_mutex);
1147 * The only time we want to leave bios in flight is when a successful
1148 * partial aio read or full aio write have been setup. In that case
1149 * bio completion will call aio_complete. The only time it's safe to
1150 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1151 * This had *better* be the only place that raises -EIOCBQUEUED.
1153 BUG_ON(ret == -EIOCBQUEUED);
1154 if (dio->is_async && ret == 0 && dio->result &&
1155 ((rw & READ) || (dio->result == sdio->size)))
1158 if (ret != -EIOCBQUEUED)
1159 dio_await_completion(dio);
1162 * Sync will always be dropping the final ref and completing the
1163 * operation. AIO can if it was a broken operation described above or
1164 * in fact if all the bios race to complete before we get here. In
1165 * that case dio_complete() translates the EIOCBQUEUED into the proper
1166 * return code that the caller will hand to aio_complete().
1168 * This is managed by the bio_lock instead of being an atomic_t so that
1169 * completion paths can drop their ref and use the remaining count to
1170 * decide to wake the submission path atomically.
1172 spin_lock_irqsave(&dio->bio_lock, flags);
1173 ret2 = --dio->refcount;
1174 spin_unlock_irqrestore(&dio->bio_lock, flags);
1177 ret = dio_complete(dio, offset, ret, false);
1180 BUG_ON(ret != -EIOCBQUEUED);
1186 * This is a library function for use by filesystem drivers.
1188 * The locking rules are governed by the flags parameter:
1189 * - if the flags value contains DIO_LOCKING we use a fancy locking
1190 * scheme for dumb filesystems.
1191 * For writes this function is called under i_mutex and returns with
1192 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1193 * taken and dropped again before returning.
1194 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1195 * internal locking but rather rely on the filesystem to synchronize
1196 * direct I/O reads/writes versus each other and truncate.
1198 * To help with locking against truncate we incremented the i_dio_count
1199 * counter before starting direct I/O, and decrement it once we are done.
1200 * Truncate can wait for it to reach zero to provide exclusion. It is
1201 * expected that filesystem provide exclusion between new direct I/O
1202 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1203 * but other filesystems need to take care of this on their own.
1206 __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1207 struct block_device *bdev, const struct iovec *iov, loff_t offset,
1208 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1209 dio_submit_t submit_io, int flags)
1214 unsigned blkbits = inode->i_blkbits;
1215 unsigned bdev_blkbits = 0;
1216 unsigned blocksize_mask = (1 << blkbits) - 1;
1217 ssize_t retval = -EINVAL;
1218 loff_t end = offset;
1220 struct dio_submit sdio = { 0, };
1226 bdev_blkbits = blksize_bits(bdev_logical_block_size(bdev));
1228 if (offset & blocksize_mask) {
1230 blkbits = bdev_blkbits;
1231 blocksize_mask = (1 << blkbits) - 1;
1232 if (offset & blocksize_mask)
1236 /* Check the memory alignment. Blocks cannot straddle pages */
1237 for (seg = 0; seg < nr_segs; seg++) {
1238 addr = (unsigned long)iov[seg].iov_base;
1239 size = iov[seg].iov_len;
1241 if ((addr & blocksize_mask) || (size & blocksize_mask)) {
1243 blkbits = bdev_blkbits;
1244 blocksize_mask = (1 << blkbits) - 1;
1245 if ((addr & blocksize_mask) || (size & blocksize_mask))
1250 /* watch out for a 0 len io from a tricksy fs */
1251 if (rw == READ && end == offset)
1254 dio = kmalloc(sizeof(*dio), GFP_KERNEL);
1259 * Believe it or not, zeroing out the page array caused a .5%
1260 * performance regression in a database benchmark. So, we take
1261 * care to only zero out what's needed.
1263 memset(dio, 0, offsetof(struct dio, pages));
1266 if (dio->flags & DIO_LOCKING) {
1268 struct address_space *mapping =
1269 iocb->ki_filp->f_mapping;
1271 /* will be released by direct_io_worker */
1272 mutex_lock(&inode->i_mutex);
1274 retval = filemap_write_and_wait_range(mapping, offset,
1277 mutex_unlock(&inode->i_mutex);
1285 * Will be decremented at I/O completion time.
1287 atomic_inc(&inode->i_dio_count);
1290 * For file extending writes updating i_size before data
1291 * writeouts complete can expose uninitialized blocks. So
1292 * even for AIO, we need to wait for i/o to complete before
1293 * returning in this case.
1295 dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
1296 (end > i_size_read(inode)));
1298 retval = direct_io_worker(rw, iocb, inode, iov, offset,
1299 nr_segs, blkbits, get_block, end_io,
1300 submit_io, dio, &sdio);
1305 EXPORT_SYMBOL(__blockdev_direct_IO);