pull master
[hh.org.git] / fs / direct-io.c
blob5981e17f46f051afa98ec0b7af67c1b25a0dc897
1 /*
2 * fs/direct-io.c
4 * Copyright (C) 2002, Linus Torvalds.
6 * O_DIRECT
8 * 04Jul2002 akpm@zip.com.au
9 * Initial version
10 * 11Sep2002 janetinc@us.ibm.com
11 * added readv/writev support.
12 * 29Oct2002 akpm@zip.com.au
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>
25 #include <linux/fs.h>
26 #include <linux/mm.h>
27 #include <linux/slab.h>
28 #include <linux/highmem.h>
29 #include <linux/pagemap.h>
30 #include <linux/bio.h>
31 #include <linux/wait.h>
32 #include <linux/err.h>
33 #include <linux/blkdev.h>
34 #include <linux/buffer_head.h>
35 #include <linux/rwsem.h>
36 #include <linux/uio.h>
37 #include <asm/atomic.h>
40 * How many user pages to map in one call to get_user_pages(). This determines
41 * the size of a structure on the stack.
43 #define DIO_PAGES 64
46 * This code generally works in units of "dio_blocks". A dio_block is
47 * somewhere between the hard sector size and the filesystem block size. it
48 * is determined on a per-invocation basis. When talking to the filesystem
49 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
50 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
51 * to bio_block quantities by shifting left by blkfactor.
53 * If blkfactor is zero then the user's request was aligned to the filesystem's
54 * blocksize.
56 * lock_type is DIO_LOCKING for regular files on direct-IO-naive filesystems.
57 * This determines whether we need to do the fancy locking which prevents
58 * direct-IO from being able to read uninitialised disk blocks. If its zero
59 * (blockdev) this locking is not done, and if it is DIO_OWN_LOCKING i_mutex is
60 * not held for the entire direct write (taken briefly, initially, during a
61 * direct read though, but its never held for the duration of a direct-IO).
64 struct dio {
65 /* BIO submission state */
66 struct bio *bio; /* bio under assembly */
67 struct inode *inode;
68 int rw;
69 loff_t i_size; /* i_size when submitted */
70 int lock_type; /* doesn't change */
71 unsigned blkbits; /* doesn't change */
72 unsigned blkfactor; /* When we're using an alignment which
73 is finer than the filesystem's soft
74 blocksize, this specifies how much
75 finer. blkfactor=2 means 1/4-block
76 alignment. Does not change */
77 unsigned start_zero_done; /* flag: sub-blocksize zeroing has
78 been performed at the start of a
79 write */
80 int pages_in_io; /* approximate total IO pages */
81 size_t size; /* total request size (doesn't change)*/
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 sector_t final_block_in_request;/* doesn't change */
86 unsigned first_block_in_page; /* doesn't change, Used only once */
87 int boundary; /* prev block is at a boundary */
88 int reap_counter; /* rate limit reaping */
89 get_block_t *get_block; /* block mapping function */
90 dio_iodone_t *end_io; /* IO completion function */
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 */
94 struct buffer_head map_bh; /* last get_block() result */
97 * Deferred addition of a page to the dio. These variables are
98 * private to dio_send_cur_page(), submit_page_section() and
99 * dio_bio_add_page().
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 */
107 * Page fetching state. These variables belong to dio_refill_pages().
109 int curr_page; /* changes */
110 int total_pages; /* doesn't change */
111 unsigned long curr_user_address;/* changes */
114 * Page queue. These variables belong to dio_refill_pages() and
115 * dio_get_page().
117 struct page *pages[DIO_PAGES]; /* page buffer */
118 unsigned head; /* next page to process */
119 unsigned tail; /* last valid page + 1 */
120 int page_errors; /* errno from get_user_pages() */
122 /* BIO completion state */
123 spinlock_t bio_lock; /* protects BIO fields below */
124 int bio_count; /* nr bios to be completed */
125 int bios_in_flight; /* nr bios in flight */
126 struct bio *bio_list; /* singly linked via bi_private */
127 struct task_struct *waiter; /* waiting task (NULL if none) */
129 /* AIO related stuff */
130 struct kiocb *iocb; /* kiocb */
131 int is_async; /* is IO async ? */
132 int io_error; /* IO error in completion path */
133 ssize_t result; /* IO result */
137 * How many pages are in the queue?
139 static inline unsigned dio_pages_present(struct dio *dio)
141 return dio->tail - dio->head;
145 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
147 static int dio_refill_pages(struct dio *dio)
149 int ret;
150 int nr_pages;
152 nr_pages = min(dio->total_pages - dio->curr_page, DIO_PAGES);
153 down_read(&current->mm->mmap_sem);
154 ret = get_user_pages(
155 current, /* Task for fault acounting */
156 current->mm, /* whose pages? */
157 dio->curr_user_address, /* Where from? */
158 nr_pages, /* How many pages? */
159 dio->rw == READ, /* Write to memory? */
160 0, /* force (?) */
161 &dio->pages[0],
162 NULL); /* vmas */
163 up_read(&current->mm->mmap_sem);
165 if (ret < 0 && dio->blocks_available && (dio->rw & WRITE)) {
166 struct page *page = ZERO_PAGE(dio->curr_user_address);
168 * A memory fault, but the filesystem has some outstanding
169 * mapped blocks. We need to use those blocks up to avoid
170 * leaking stale data in the file.
172 if (dio->page_errors == 0)
173 dio->page_errors = ret;
174 page_cache_get(page);
175 dio->pages[0] = page;
176 dio->head = 0;
177 dio->tail = 1;
178 ret = 0;
179 goto out;
182 if (ret >= 0) {
183 dio->curr_user_address += ret * PAGE_SIZE;
184 dio->curr_page += ret;
185 dio->head = 0;
186 dio->tail = ret;
187 ret = 0;
189 out:
190 return ret;
194 * Get another userspace page. Returns an ERR_PTR on error. Pages are
195 * buffered inside the dio so that we can call get_user_pages() against a
196 * decent number of pages, less frequently. To provide nicer use of the
197 * L1 cache.
199 static struct page *dio_get_page(struct dio *dio)
201 if (dio_pages_present(dio) == 0) {
202 int ret;
204 ret = dio_refill_pages(dio);
205 if (ret)
206 return ERR_PTR(ret);
207 BUG_ON(dio_pages_present(dio) == 0);
209 return dio->pages[dio->head++];
213 * Called when all DIO BIO I/O has been completed - let the filesystem
214 * know, if it registered an interest earlier via get_block. Pass the
215 * private field of the map buffer_head so that filesystems can use it
216 * to hold additional state between get_block calls and dio_complete.
218 static void dio_complete(struct dio *dio, loff_t offset, ssize_t bytes)
220 if (dio->end_io && dio->result)
221 dio->end_io(dio->iocb, offset, bytes, dio->map_bh.b_private);
222 if (dio->lock_type == DIO_LOCKING)
223 /* lockdep: non-owner release */
224 up_read_non_owner(&dio->inode->i_alloc_sem);
228 * Called when a BIO has been processed. If the count goes to zero then IO is
229 * complete and we can signal this to the AIO layer.
231 static void finished_one_bio(struct dio *dio)
233 unsigned long flags;
235 spin_lock_irqsave(&dio->bio_lock, flags);
236 if (dio->bio_count == 1) {
237 if (dio->is_async) {
238 ssize_t transferred;
239 loff_t offset;
242 * Last reference to the dio is going away.
243 * Drop spinlock and complete the DIO.
245 spin_unlock_irqrestore(&dio->bio_lock, flags);
247 /* Check for short read case */
248 transferred = dio->result;
249 offset = dio->iocb->ki_pos;
251 if ((dio->rw == READ) &&
252 ((offset + transferred) > dio->i_size))
253 transferred = dio->i_size - offset;
255 /* check for error in completion path */
256 if (dio->io_error)
257 transferred = dio->io_error;
259 dio_complete(dio, offset, transferred);
261 /* Complete AIO later if falling back to buffered i/o */
262 if (dio->result == dio->size ||
263 ((dio->rw == READ) && dio->result)) {
264 aio_complete(dio->iocb, transferred, 0);
265 kfree(dio);
266 return;
267 } else {
269 * Falling back to buffered
271 spin_lock_irqsave(&dio->bio_lock, flags);
272 dio->bio_count--;
273 if (dio->waiter)
274 wake_up_process(dio->waiter);
275 spin_unlock_irqrestore(&dio->bio_lock, flags);
276 return;
280 dio->bio_count--;
281 spin_unlock_irqrestore(&dio->bio_lock, flags);
284 static int dio_bio_complete(struct dio *dio, struct bio *bio);
286 * Asynchronous IO callback.
288 static int dio_bio_end_aio(struct bio *bio, unsigned int bytes_done, int error)
290 struct dio *dio = bio->bi_private;
292 if (bio->bi_size)
293 return 1;
295 /* cleanup the bio */
296 dio_bio_complete(dio, bio);
297 return 0;
301 * The BIO completion handler simply queues the BIO up for the process-context
302 * handler.
304 * During I/O bi_private points at the dio. After I/O, bi_private is used to
305 * implement a singly-linked list of completed BIOs, at dio->bio_list.
307 static int dio_bio_end_io(struct bio *bio, unsigned int bytes_done, int error)
309 struct dio *dio = bio->bi_private;
310 unsigned long flags;
312 if (bio->bi_size)
313 return 1;
315 spin_lock_irqsave(&dio->bio_lock, flags);
316 bio->bi_private = dio->bio_list;
317 dio->bio_list = bio;
318 dio->bios_in_flight--;
319 if (dio->waiter && dio->bios_in_flight == 0)
320 wake_up_process(dio->waiter);
321 spin_unlock_irqrestore(&dio->bio_lock, flags);
322 return 0;
325 static int
326 dio_bio_alloc(struct dio *dio, struct block_device *bdev,
327 sector_t first_sector, int nr_vecs)
329 struct bio *bio;
331 bio = bio_alloc(GFP_KERNEL, nr_vecs);
332 if (bio == NULL)
333 return -ENOMEM;
335 bio->bi_bdev = bdev;
336 bio->bi_sector = first_sector;
337 if (dio->is_async)
338 bio->bi_end_io = dio_bio_end_aio;
339 else
340 bio->bi_end_io = dio_bio_end_io;
342 dio->bio = bio;
343 return 0;
347 * In the AIO read case we speculatively dirty the pages before starting IO.
348 * During IO completion, any of these pages which happen to have been written
349 * back will be redirtied by bio_check_pages_dirty().
351 static void dio_bio_submit(struct dio *dio)
353 struct bio *bio = dio->bio;
354 unsigned long flags;
356 bio->bi_private = dio;
357 spin_lock_irqsave(&dio->bio_lock, flags);
358 dio->bio_count++;
359 dio->bios_in_flight++;
360 spin_unlock_irqrestore(&dio->bio_lock, flags);
361 if (dio->is_async && dio->rw == READ)
362 bio_set_pages_dirty(bio);
363 submit_bio(dio->rw, bio);
365 dio->bio = NULL;
366 dio->boundary = 0;
370 * Release any resources in case of a failure
372 static void dio_cleanup(struct dio *dio)
374 while (dio_pages_present(dio))
375 page_cache_release(dio_get_page(dio));
379 * Wait for the next BIO to complete. Remove it and return it.
381 static struct bio *dio_await_one(struct dio *dio)
383 unsigned long flags;
384 struct bio *bio;
386 spin_lock_irqsave(&dio->bio_lock, flags);
387 while (dio->bio_list == NULL) {
388 set_current_state(TASK_UNINTERRUPTIBLE);
389 if (dio->bio_list == NULL) {
390 dio->waiter = current;
391 spin_unlock_irqrestore(&dio->bio_lock, flags);
392 blk_run_address_space(dio->inode->i_mapping);
393 io_schedule();
394 spin_lock_irqsave(&dio->bio_lock, flags);
395 dio->waiter = NULL;
397 set_current_state(TASK_RUNNING);
399 bio = dio->bio_list;
400 dio->bio_list = bio->bi_private;
401 spin_unlock_irqrestore(&dio->bio_lock, flags);
402 return bio;
406 * Process one completed BIO. No locks are held.
408 static int dio_bio_complete(struct dio *dio, struct bio *bio)
410 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
411 struct bio_vec *bvec = bio->bi_io_vec;
412 int page_no;
414 if (!uptodate)
415 dio->io_error = -EIO;
417 if (dio->is_async && dio->rw == READ) {
418 bio_check_pages_dirty(bio); /* transfers ownership */
419 } else {
420 for (page_no = 0; page_no < bio->bi_vcnt; page_no++) {
421 struct page *page = bvec[page_no].bv_page;
423 if (dio->rw == READ && !PageCompound(page))
424 set_page_dirty_lock(page);
425 page_cache_release(page);
427 bio_put(bio);
429 finished_one_bio(dio);
430 return uptodate ? 0 : -EIO;
434 * Wait on and process all in-flight BIOs.
436 static int dio_await_completion(struct dio *dio)
438 int ret = 0;
440 if (dio->bio)
441 dio_bio_submit(dio);
444 * The bio_lock is not held for the read of bio_count.
445 * This is ok since it is the dio_bio_complete() that changes
446 * bio_count.
448 while (dio->bio_count) {
449 struct bio *bio = dio_await_one(dio);
450 int ret2;
452 ret2 = dio_bio_complete(dio, bio);
453 if (ret == 0)
454 ret = ret2;
456 return ret;
460 * A really large O_DIRECT read or write can generate a lot of BIOs. So
461 * to keep the memory consumption sane we periodically reap any completed BIOs
462 * during the BIO generation phase.
464 * This also helps to limit the peak amount of pinned userspace memory.
466 static int dio_bio_reap(struct dio *dio)
468 int ret = 0;
470 if (dio->reap_counter++ >= 64) {
471 while (dio->bio_list) {
472 unsigned long flags;
473 struct bio *bio;
474 int ret2;
476 spin_lock_irqsave(&dio->bio_lock, flags);
477 bio = dio->bio_list;
478 dio->bio_list = bio->bi_private;
479 spin_unlock_irqrestore(&dio->bio_lock, flags);
480 ret2 = dio_bio_complete(dio, bio);
481 if (ret == 0)
482 ret = ret2;
484 dio->reap_counter = 0;
486 return ret;
490 * Call into the fs to map some more disk blocks. We record the current number
491 * of available blocks at dio->blocks_available. These are in units of the
492 * fs blocksize, (1 << inode->i_blkbits).
494 * The fs is allowed to map lots of blocks at once. If it wants to do that,
495 * it uses the passed inode-relative block number as the file offset, as usual.
497 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
498 * has remaining to do. The fs should not map more than this number of blocks.
500 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
501 * indicate how much contiguous disk space has been made available at
502 * bh->b_blocknr.
504 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
505 * This isn't very efficient...
507 * In the case of filesystem holes: the fs may return an arbitrarily-large
508 * hole by returning an appropriate value in b_size and by clearing
509 * buffer_mapped(). However the direct-io code will only process holes one
510 * block at a time - it will repeatedly call get_block() as it walks the hole.
512 static int get_more_blocks(struct dio *dio)
514 int ret;
515 struct buffer_head *map_bh = &dio->map_bh;
516 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
517 unsigned long fs_count; /* Number of filesystem-sized blocks */
518 unsigned long dio_count;/* Number of dio_block-sized blocks */
519 unsigned long blkmask;
520 int create;
523 * If there was a memory error and we've overwritten all the
524 * mapped blocks then we can now return that memory error
526 ret = dio->page_errors;
527 if (ret == 0) {
528 BUG_ON(dio->block_in_file >= dio->final_block_in_request);
529 fs_startblk = dio->block_in_file >> dio->blkfactor;
530 dio_count = dio->final_block_in_request - dio->block_in_file;
531 fs_count = dio_count >> dio->blkfactor;
532 blkmask = (1 << dio->blkfactor) - 1;
533 if (dio_count & blkmask)
534 fs_count++;
536 map_bh->b_state = 0;
537 map_bh->b_size = fs_count << dio->inode->i_blkbits;
539 create = dio->rw & WRITE;
540 if (dio->lock_type == DIO_LOCKING) {
541 if (dio->block_in_file < (i_size_read(dio->inode) >>
542 dio->blkbits))
543 create = 0;
544 } else if (dio->lock_type == DIO_NO_LOCKING) {
545 create = 0;
549 * For writes inside i_size we forbid block creations: only
550 * overwrites are permitted. We fall back to buffered writes
551 * at a higher level for inside-i_size block-instantiating
552 * writes.
554 ret = (*dio->get_block)(dio->inode, fs_startblk,
555 map_bh, create);
557 return ret;
561 * There is no bio. Make one now.
563 static int dio_new_bio(struct dio *dio, sector_t start_sector)
565 sector_t sector;
566 int ret, nr_pages;
568 ret = dio_bio_reap(dio);
569 if (ret)
570 goto out;
571 sector = start_sector << (dio->blkbits - 9);
572 nr_pages = min(dio->pages_in_io, bio_get_nr_vecs(dio->map_bh.b_bdev));
573 BUG_ON(nr_pages <= 0);
574 ret = dio_bio_alloc(dio, dio->map_bh.b_bdev, sector, nr_pages);
575 dio->boundary = 0;
576 out:
577 return ret;
581 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
582 * that was successful then update final_block_in_bio and take a ref against
583 * the just-added page.
585 * Return zero on success. Non-zero means the caller needs to start a new BIO.
587 static int dio_bio_add_page(struct dio *dio)
589 int ret;
591 ret = bio_add_page(dio->bio, dio->cur_page,
592 dio->cur_page_len, dio->cur_page_offset);
593 if (ret == dio->cur_page_len) {
595 * Decrement count only, if we are done with this page
597 if ((dio->cur_page_len + dio->cur_page_offset) == PAGE_SIZE)
598 dio->pages_in_io--;
599 page_cache_get(dio->cur_page);
600 dio->final_block_in_bio = dio->cur_page_block +
601 (dio->cur_page_len >> dio->blkbits);
602 ret = 0;
603 } else {
604 ret = 1;
606 return ret;
610 * Put cur_page under IO. The section of cur_page which is described by
611 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
612 * starts on-disk at cur_page_block.
614 * We take a ref against the page here (on behalf of its presence in the bio).
616 * The caller of this function is responsible for removing cur_page from the
617 * dio, and for dropping the refcount which came from that presence.
619 static int dio_send_cur_page(struct dio *dio)
621 int ret = 0;
623 if (dio->bio) {
625 * See whether this new request is contiguous with the old
627 if (dio->final_block_in_bio != dio->cur_page_block)
628 dio_bio_submit(dio);
630 * Submit now if the underlying fs is about to perform a
631 * metadata read
633 if (dio->boundary)
634 dio_bio_submit(dio);
637 if (dio->bio == NULL) {
638 ret = dio_new_bio(dio, dio->cur_page_block);
639 if (ret)
640 goto out;
643 if (dio_bio_add_page(dio) != 0) {
644 dio_bio_submit(dio);
645 ret = dio_new_bio(dio, dio->cur_page_block);
646 if (ret == 0) {
647 ret = dio_bio_add_page(dio);
648 BUG_ON(ret != 0);
651 out:
652 return ret;
656 * An autonomous function to put a chunk of a page under deferred IO.
658 * The caller doesn't actually know (or care) whether this piece of page is in
659 * a BIO, or is under IO or whatever. We just take care of all possible
660 * situations here. The separation between the logic of do_direct_IO() and
661 * that of submit_page_section() is important for clarity. Please don't break.
663 * The chunk of page starts on-disk at blocknr.
665 * We perform deferred IO, by recording the last-submitted page inside our
666 * private part of the dio structure. If possible, we just expand the IO
667 * across that page here.
669 * If that doesn't work out then we put the old page into the bio and add this
670 * page to the dio instead.
672 static int
673 submit_page_section(struct dio *dio, struct page *page,
674 unsigned offset, unsigned len, sector_t blocknr)
676 int ret = 0;
679 * Can we just grow the current page's presence in the dio?
681 if ( (dio->cur_page == page) &&
682 (dio->cur_page_offset + dio->cur_page_len == offset) &&
683 (dio->cur_page_block +
684 (dio->cur_page_len >> dio->blkbits) == blocknr)) {
685 dio->cur_page_len += len;
688 * If dio->boundary then we want to schedule the IO now to
689 * avoid metadata seeks.
691 if (dio->boundary) {
692 ret = dio_send_cur_page(dio);
693 page_cache_release(dio->cur_page);
694 dio->cur_page = NULL;
696 goto out;
700 * If there's a deferred page already there then send it.
702 if (dio->cur_page) {
703 ret = dio_send_cur_page(dio);
704 page_cache_release(dio->cur_page);
705 dio->cur_page = NULL;
706 if (ret)
707 goto out;
710 page_cache_get(page); /* It is in dio */
711 dio->cur_page = page;
712 dio->cur_page_offset = offset;
713 dio->cur_page_len = len;
714 dio->cur_page_block = blocknr;
715 out:
716 return ret;
720 * Clean any dirty buffers in the blockdev mapping which alias newly-created
721 * file blocks. Only called for S_ISREG files - blockdevs do not set
722 * buffer_new
724 static void clean_blockdev_aliases(struct dio *dio)
726 unsigned i;
727 unsigned nblocks;
729 nblocks = dio->map_bh.b_size >> dio->inode->i_blkbits;
731 for (i = 0; i < nblocks; i++) {
732 unmap_underlying_metadata(dio->map_bh.b_bdev,
733 dio->map_bh.b_blocknr + i);
738 * If we are not writing the entire block and get_block() allocated
739 * the block for us, we need to fill-in the unused portion of the
740 * block with zeros. This happens only if user-buffer, fileoffset or
741 * io length is not filesystem block-size multiple.
743 * `end' is zero if we're doing the start of the IO, 1 at the end of the
744 * IO.
746 static void dio_zero_block(struct dio *dio, int end)
748 unsigned dio_blocks_per_fs_block;
749 unsigned this_chunk_blocks; /* In dio_blocks */
750 unsigned this_chunk_bytes;
751 struct page *page;
753 dio->start_zero_done = 1;
754 if (!dio->blkfactor || !buffer_new(&dio->map_bh))
755 return;
757 dio_blocks_per_fs_block = 1 << dio->blkfactor;
758 this_chunk_blocks = dio->block_in_file & (dio_blocks_per_fs_block - 1);
760 if (!this_chunk_blocks)
761 return;
764 * We need to zero out part of an fs block. It is either at the
765 * beginning or the end of the fs block.
767 if (end)
768 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
770 this_chunk_bytes = this_chunk_blocks << dio->blkbits;
772 page = ZERO_PAGE(dio->curr_user_address);
773 if (submit_page_section(dio, page, 0, this_chunk_bytes,
774 dio->next_block_for_io))
775 return;
777 dio->next_block_for_io += this_chunk_blocks;
781 * Walk the user pages, and the file, mapping blocks to disk and generating
782 * a sequence of (page,offset,len,block) mappings. These mappings are injected
783 * into submit_page_section(), which takes care of the next stage of submission
785 * Direct IO against a blockdev is different from a file. Because we can
786 * happily perform page-sized but 512-byte aligned IOs. It is important that
787 * blockdev IO be able to have fine alignment and large sizes.
789 * So what we do is to permit the ->get_block function to populate bh.b_size
790 * with the size of IO which is permitted at this offset and this i_blkbits.
792 * For best results, the blockdev should be set up with 512-byte i_blkbits and
793 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
794 * fine alignment but still allows this function to work in PAGE_SIZE units.
796 static int do_direct_IO(struct dio *dio)
798 const unsigned blkbits = dio->blkbits;
799 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
800 struct page *page;
801 unsigned block_in_page;
802 struct buffer_head *map_bh = &dio->map_bh;
803 int ret = 0;
805 /* The I/O can start at any block offset within the first page */
806 block_in_page = dio->first_block_in_page;
808 while (dio->block_in_file < dio->final_block_in_request) {
809 page = dio_get_page(dio);
810 if (IS_ERR(page)) {
811 ret = PTR_ERR(page);
812 goto out;
815 while (block_in_page < blocks_per_page) {
816 unsigned offset_in_page = block_in_page << blkbits;
817 unsigned this_chunk_bytes; /* # of bytes mapped */
818 unsigned this_chunk_blocks; /* # of blocks */
819 unsigned u;
821 if (dio->blocks_available == 0) {
823 * Need to go and map some more disk
825 unsigned long blkmask;
826 unsigned long dio_remainder;
828 ret = get_more_blocks(dio);
829 if (ret) {
830 page_cache_release(page);
831 goto out;
833 if (!buffer_mapped(map_bh))
834 goto do_holes;
836 dio->blocks_available =
837 map_bh->b_size >> dio->blkbits;
838 dio->next_block_for_io =
839 map_bh->b_blocknr << dio->blkfactor;
840 if (buffer_new(map_bh))
841 clean_blockdev_aliases(dio);
843 if (!dio->blkfactor)
844 goto do_holes;
846 blkmask = (1 << dio->blkfactor) - 1;
847 dio_remainder = (dio->block_in_file & blkmask);
850 * If we are at the start of IO and that IO
851 * starts partway into a fs-block,
852 * dio_remainder will be non-zero. If the IO
853 * is a read then we can simply advance the IO
854 * cursor to the first block which is to be
855 * read. But if the IO is a write and the
856 * block was newly allocated we cannot do that;
857 * the start of the fs block must be zeroed out
858 * on-disk
860 if (!buffer_new(map_bh))
861 dio->next_block_for_io += dio_remainder;
862 dio->blocks_available -= dio_remainder;
864 do_holes:
865 /* Handle holes */
866 if (!buffer_mapped(map_bh)) {
867 char *kaddr;
868 loff_t i_size_aligned;
870 /* AKPM: eargh, -ENOTBLK is a hack */
871 if (dio->rw & WRITE) {
872 page_cache_release(page);
873 return -ENOTBLK;
877 * Be sure to account for a partial block as the
878 * last block in the file
880 i_size_aligned = ALIGN(i_size_read(dio->inode),
881 1 << blkbits);
882 if (dio->block_in_file >=
883 i_size_aligned >> blkbits) {
884 /* We hit eof */
885 page_cache_release(page);
886 goto out;
888 kaddr = kmap_atomic(page, KM_USER0);
889 memset(kaddr + (block_in_page << blkbits),
890 0, 1 << blkbits);
891 flush_dcache_page(page);
892 kunmap_atomic(kaddr, KM_USER0);
893 dio->block_in_file++;
894 block_in_page++;
895 goto next_block;
899 * If we're performing IO which has an alignment which
900 * is finer than the underlying fs, go check to see if
901 * we must zero out the start of this block.
903 if (unlikely(dio->blkfactor && !dio->start_zero_done))
904 dio_zero_block(dio, 0);
907 * Work out, in this_chunk_blocks, how much disk we
908 * can add to this page
910 this_chunk_blocks = dio->blocks_available;
911 u = (PAGE_SIZE - offset_in_page) >> blkbits;
912 if (this_chunk_blocks > u)
913 this_chunk_blocks = u;
914 u = dio->final_block_in_request - dio->block_in_file;
915 if (this_chunk_blocks > u)
916 this_chunk_blocks = u;
917 this_chunk_bytes = this_chunk_blocks << blkbits;
918 BUG_ON(this_chunk_bytes == 0);
920 dio->boundary = buffer_boundary(map_bh);
921 ret = submit_page_section(dio, page, offset_in_page,
922 this_chunk_bytes, dio->next_block_for_io);
923 if (ret) {
924 page_cache_release(page);
925 goto out;
927 dio->next_block_for_io += this_chunk_blocks;
929 dio->block_in_file += this_chunk_blocks;
930 block_in_page += this_chunk_blocks;
931 dio->blocks_available -= this_chunk_blocks;
932 next_block:
933 BUG_ON(dio->block_in_file > dio->final_block_in_request);
934 if (dio->block_in_file == dio->final_block_in_request)
935 break;
938 /* Drop the ref which was taken in get_user_pages() */
939 page_cache_release(page);
940 block_in_page = 0;
942 out:
943 return ret;
947 * Releases both i_mutex and i_alloc_sem
949 static ssize_t
950 direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
951 const struct iovec *iov, loff_t offset, unsigned long nr_segs,
952 unsigned blkbits, get_block_t get_block, dio_iodone_t end_io,
953 struct dio *dio)
955 unsigned long user_addr;
956 int seg;
957 ssize_t ret = 0;
958 ssize_t ret2;
959 size_t bytes;
961 dio->bio = NULL;
962 dio->inode = inode;
963 dio->rw = rw;
964 dio->blkbits = blkbits;
965 dio->blkfactor = inode->i_blkbits - blkbits;
966 dio->start_zero_done = 0;
967 dio->size = 0;
968 dio->block_in_file = offset >> blkbits;
969 dio->blocks_available = 0;
970 dio->cur_page = NULL;
972 dio->boundary = 0;
973 dio->reap_counter = 0;
974 dio->get_block = get_block;
975 dio->end_io = end_io;
976 dio->map_bh.b_private = NULL;
977 dio->final_block_in_bio = -1;
978 dio->next_block_for_io = -1;
980 dio->page_errors = 0;
981 dio->io_error = 0;
982 dio->result = 0;
983 dio->iocb = iocb;
984 dio->i_size = i_size_read(inode);
987 * BIO completion state.
989 * ->bio_count starts out at one, and we decrement it to zero after all
990 * BIOs are submitted. This to avoid the situation where a really fast
991 * (or synchronous) device could take the count to zero while we're
992 * still submitting BIOs.
994 dio->bio_count = 1;
995 dio->bios_in_flight = 0;
996 spin_lock_init(&dio->bio_lock);
997 dio->bio_list = NULL;
998 dio->waiter = NULL;
1001 * In case of non-aligned buffers, we may need 2 more
1002 * pages since we need to zero out first and last block.
1004 if (unlikely(dio->blkfactor))
1005 dio->pages_in_io = 2;
1006 else
1007 dio->pages_in_io = 0;
1009 for (seg = 0; seg < nr_segs; seg++) {
1010 user_addr = (unsigned long)iov[seg].iov_base;
1011 dio->pages_in_io +=
1012 ((user_addr+iov[seg].iov_len +PAGE_SIZE-1)/PAGE_SIZE
1013 - user_addr/PAGE_SIZE);
1016 for (seg = 0; seg < nr_segs; seg++) {
1017 user_addr = (unsigned long)iov[seg].iov_base;
1018 dio->size += bytes = iov[seg].iov_len;
1020 /* Index into the first page of the first block */
1021 dio->first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
1022 dio->final_block_in_request = dio->block_in_file +
1023 (bytes >> blkbits);
1024 /* Page fetching state */
1025 dio->head = 0;
1026 dio->tail = 0;
1027 dio->curr_page = 0;
1029 dio->total_pages = 0;
1030 if (user_addr & (PAGE_SIZE-1)) {
1031 dio->total_pages++;
1032 bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
1034 dio->total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
1035 dio->curr_user_address = user_addr;
1037 ret = do_direct_IO(dio);
1039 dio->result += iov[seg].iov_len -
1040 ((dio->final_block_in_request - dio->block_in_file) <<
1041 blkbits);
1043 if (ret) {
1044 dio_cleanup(dio);
1045 break;
1047 } /* end iovec loop */
1049 if (ret == -ENOTBLK && (rw & WRITE)) {
1051 * The remaining part of the request will be
1052 * be handled by buffered I/O when we return
1054 ret = 0;
1057 * There may be some unwritten disk at the end of a part-written
1058 * fs-block-sized block. Go zero that now.
1060 dio_zero_block(dio, 1);
1062 if (dio->cur_page) {
1063 ret2 = dio_send_cur_page(dio);
1064 if (ret == 0)
1065 ret = ret2;
1066 page_cache_release(dio->cur_page);
1067 dio->cur_page = NULL;
1069 if (dio->bio)
1070 dio_bio_submit(dio);
1073 * It is possible that, we return short IO due to end of file.
1074 * In that case, we need to release all the pages we got hold on.
1076 dio_cleanup(dio);
1079 * All block lookups have been performed. For READ requests
1080 * we can let i_mutex go now that its achieved its purpose
1081 * of protecting us from looking up uninitialized blocks.
1083 if ((rw == READ) && (dio->lock_type == DIO_LOCKING))
1084 mutex_unlock(&dio->inode->i_mutex);
1087 * OK, all BIOs are submitted, so we can decrement bio_count to truly
1088 * reflect the number of to-be-processed BIOs.
1090 if (dio->is_async) {
1091 int should_wait = 0;
1093 if (dio->result < dio->size && (rw & WRITE)) {
1094 dio->waiter = current;
1095 should_wait = 1;
1097 if (ret == 0)
1098 ret = dio->result;
1099 finished_one_bio(dio); /* This can free the dio */
1100 blk_run_address_space(inode->i_mapping);
1101 if (should_wait) {
1102 unsigned long flags;
1104 * Wait for already issued I/O to drain out and
1105 * release its references to user-space pages
1106 * before returning to fallback on buffered I/O
1109 spin_lock_irqsave(&dio->bio_lock, flags);
1110 set_current_state(TASK_UNINTERRUPTIBLE);
1111 while (dio->bio_count) {
1112 spin_unlock_irqrestore(&dio->bio_lock, flags);
1113 io_schedule();
1114 spin_lock_irqsave(&dio->bio_lock, flags);
1115 set_current_state(TASK_UNINTERRUPTIBLE);
1117 spin_unlock_irqrestore(&dio->bio_lock, flags);
1118 set_current_state(TASK_RUNNING);
1119 kfree(dio);
1121 } else {
1122 ssize_t transferred = 0;
1124 finished_one_bio(dio);
1125 ret2 = dio_await_completion(dio);
1126 if (ret == 0)
1127 ret = ret2;
1128 if (ret == 0)
1129 ret = dio->page_errors;
1130 if (dio->result) {
1131 loff_t i_size = i_size_read(inode);
1133 transferred = dio->result;
1135 * Adjust the return value if the read crossed a
1136 * non-block-aligned EOF.
1138 if (rw == READ && (offset + transferred > i_size))
1139 transferred = i_size - offset;
1141 dio_complete(dio, offset, transferred);
1142 if (ret == 0)
1143 ret = transferred;
1145 /* We could have also come here on an AIO file extend */
1146 if (!is_sync_kiocb(iocb) && (rw & WRITE) &&
1147 ret >= 0 && dio->result == dio->size)
1149 * For AIO writes where we have completed the
1150 * i/o, we have to mark the the aio complete.
1152 aio_complete(iocb, ret, 0);
1153 kfree(dio);
1155 return ret;
1159 * This is a library function for use by filesystem drivers.
1160 * The locking rules are governed by the dio_lock_type parameter.
1162 * DIO_NO_LOCKING (no locking, for raw block device access)
1163 * For writes, i_mutex is not held on entry; it is never taken.
1165 * DIO_LOCKING (simple locking for regular files)
1166 * For writes we are called under i_mutex and return with i_mutex held, even
1167 * though it is internally dropped.
1168 * For reads, i_mutex is not held on entry, but it is taken and dropped before
1169 * returning.
1171 * DIO_OWN_LOCKING (filesystem provides synchronisation and handling of
1172 * uninitialised data, allowing parallel direct readers and writers)
1173 * For writes we are called without i_mutex, return without it, never touch it.
1174 * For reads we are called under i_mutex and return with i_mutex held, even
1175 * though it may be internally dropped.
1177 * Additional i_alloc_sem locking requirements described inline below.
1179 ssize_t
1180 __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1181 struct block_device *bdev, const struct iovec *iov, loff_t offset,
1182 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1183 int dio_lock_type)
1185 int seg;
1186 size_t size;
1187 unsigned long addr;
1188 unsigned blkbits = inode->i_blkbits;
1189 unsigned bdev_blkbits = 0;
1190 unsigned blocksize_mask = (1 << blkbits) - 1;
1191 ssize_t retval = -EINVAL;
1192 loff_t end = offset;
1193 struct dio *dio;
1194 int release_i_mutex = 0;
1195 int acquire_i_mutex = 0;
1197 if (rw & WRITE)
1198 rw = WRITE_SYNC;
1200 if (bdev)
1201 bdev_blkbits = blksize_bits(bdev_hardsect_size(bdev));
1203 if (offset & blocksize_mask) {
1204 if (bdev)
1205 blkbits = bdev_blkbits;
1206 blocksize_mask = (1 << blkbits) - 1;
1207 if (offset & blocksize_mask)
1208 goto out;
1211 /* Check the memory alignment. Blocks cannot straddle pages */
1212 for (seg = 0; seg < nr_segs; seg++) {
1213 addr = (unsigned long)iov[seg].iov_base;
1214 size = iov[seg].iov_len;
1215 end += size;
1216 if ((addr & blocksize_mask) || (size & blocksize_mask)) {
1217 if (bdev)
1218 blkbits = bdev_blkbits;
1219 blocksize_mask = (1 << blkbits) - 1;
1220 if ((addr & blocksize_mask) || (size & blocksize_mask))
1221 goto out;
1225 dio = kmalloc(sizeof(*dio), GFP_KERNEL);
1226 retval = -ENOMEM;
1227 if (!dio)
1228 goto out;
1231 * For block device access DIO_NO_LOCKING is used,
1232 * neither readers nor writers do any locking at all
1233 * For regular files using DIO_LOCKING,
1234 * readers need to grab i_mutex and i_alloc_sem
1235 * writers need to grab i_alloc_sem only (i_mutex is already held)
1236 * For regular files using DIO_OWN_LOCKING,
1237 * neither readers nor writers take any locks here
1239 dio->lock_type = dio_lock_type;
1240 if (dio_lock_type != DIO_NO_LOCKING) {
1241 /* watch out for a 0 len io from a tricksy fs */
1242 if (rw == READ && end > offset) {
1243 struct address_space *mapping;
1245 mapping = iocb->ki_filp->f_mapping;
1246 if (dio_lock_type != DIO_OWN_LOCKING) {
1247 mutex_lock(&inode->i_mutex);
1248 release_i_mutex = 1;
1251 retval = filemap_write_and_wait_range(mapping, offset,
1252 end - 1);
1253 if (retval) {
1254 kfree(dio);
1255 goto out;
1258 if (dio_lock_type == DIO_OWN_LOCKING) {
1259 mutex_unlock(&inode->i_mutex);
1260 acquire_i_mutex = 1;
1264 if (dio_lock_type == DIO_LOCKING)
1265 /* lockdep: not the owner will release it */
1266 down_read_non_owner(&inode->i_alloc_sem);
1270 * For file extending writes updating i_size before data
1271 * writeouts complete can expose uninitialized blocks. So
1272 * even for AIO, we need to wait for i/o to complete before
1273 * returning in this case.
1275 dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
1276 (end > i_size_read(inode)));
1278 retval = direct_io_worker(rw, iocb, inode, iov, offset,
1279 nr_segs, blkbits, get_block, end_io, dio);
1281 if (rw == READ && dio_lock_type == DIO_LOCKING)
1282 release_i_mutex = 0;
1284 out:
1285 if (release_i_mutex)
1286 mutex_unlock(&inode->i_mutex);
1287 else if (acquire_i_mutex)
1288 mutex_lock(&inode->i_mutex);
1289 return retval;
1291 EXPORT_SYMBOL(__blockdev_direct_IO);