Linux 2.6.31.6
[linux/fpc-iii.git] / fs / direct-io.c
blob8b10b87dc01a294845aac2afd1a2d22ae5a269b2
1 /*
2 * fs/direct-io.c
4 * Copyright (C) 2002, Linus Torvalds.
6 * O_DIRECT
8 * 04Jul2002 Andrew Morton
9 * Initial version
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>
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/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 <asm/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.
44 #define DIO_PAGES 64
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
55 * blocksize.
57 * lock_type is DIO_LOCKING for regular files on direct-IO-naive filesystems.
58 * This determines whether we need to do the fancy locking which prevents
59 * direct-IO from being able to read uninitialised disk blocks. If its zero
60 * (blockdev) this locking is not done, and if it is DIO_OWN_LOCKING i_mutex is
61 * not held for the entire direct write (taken briefly, initially, during a
62 * direct read though, but its never held for the duration of a direct-IO).
65 struct dio {
66 /* BIO submission state */
67 struct bio *bio; /* bio under assembly */
68 struct inode *inode;
69 int rw;
70 loff_t i_size; /* i_size when submitted */
71 int lock_type; /* doesn't change */
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
80 write */
81 int pages_in_io; /* approximate total IO pages */
82 size_t size; /* total request size (doesn't change)*/
83 sector_t block_in_file; /* Current offset into the underlying
84 file in dio_block units. */
85 unsigned blocks_available; /* At block_in_file. changes */
86 sector_t final_block_in_request;/* doesn't change */
87 unsigned first_block_in_page; /* doesn't change, Used only once */
88 int boundary; /* prev block is at a boundary */
89 int reap_counter; /* rate limit reaping */
90 get_block_t *get_block; /* block mapping function */
91 dio_iodone_t *end_io; /* IO completion function */
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 */
95 struct buffer_head map_bh; /* last get_block() result */
98 * Deferred addition of a page to the dio. These variables are
99 * private to dio_send_cur_page(), submit_page_section() and
100 * dio_bio_add_page().
102 struct page *cur_page; /* The page */
103 unsigned cur_page_offset; /* Offset into it, in bytes */
104 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
105 sector_t cur_page_block; /* Where it starts */
108 * Page fetching state. These variables belong to dio_refill_pages().
110 int curr_page; /* changes */
111 int total_pages; /* doesn't change */
112 unsigned long curr_user_address;/* changes */
115 * Page queue. These variables belong to dio_refill_pages() and
116 * dio_get_page().
118 struct page *pages[DIO_PAGES]; /* page buffer */
119 unsigned head; /* next page to process */
120 unsigned tail; /* last valid page + 1 */
121 int page_errors; /* errno from get_user_pages() */
123 /* BIO completion state */
124 spinlock_t bio_lock; /* protects BIO fields below */
125 unsigned long refcount; /* direct_io_worker() and bios */
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 ret = get_user_pages_fast(
154 dio->curr_user_address, /* Where from? */
155 nr_pages, /* How many pages? */
156 dio->rw == READ, /* Write to memory? */
157 &dio->pages[0]); /* Put results here */
159 if (ret < 0 && dio->blocks_available && (dio->rw & WRITE)) {
160 struct page *page = ZERO_PAGE(0);
162 * A memory fault, but the filesystem has some outstanding
163 * mapped blocks. We need to use those blocks up to avoid
164 * leaking stale data in the file.
166 if (dio->page_errors == 0)
167 dio->page_errors = ret;
168 page_cache_get(page);
169 dio->pages[0] = page;
170 dio->head = 0;
171 dio->tail = 1;
172 ret = 0;
173 goto out;
176 if (ret >= 0) {
177 dio->curr_user_address += ret * PAGE_SIZE;
178 dio->curr_page += ret;
179 dio->head = 0;
180 dio->tail = ret;
181 ret = 0;
183 out:
184 return ret;
188 * Get another userspace page. Returns an ERR_PTR on error. Pages are
189 * buffered inside the dio so that we can call get_user_pages() against a
190 * decent number of pages, less frequently. To provide nicer use of the
191 * L1 cache.
193 static struct page *dio_get_page(struct dio *dio)
195 if (dio_pages_present(dio) == 0) {
196 int ret;
198 ret = dio_refill_pages(dio);
199 if (ret)
200 return ERR_PTR(ret);
201 BUG_ON(dio_pages_present(dio) == 0);
203 return dio->pages[dio->head++];
207 * dio_complete() - called when all DIO BIO I/O has been completed
208 * @offset: the byte offset in the file of the completed operation
210 * This releases locks as dictated by the locking type, lets interested parties
211 * know that a DIO operation has completed, and calculates the resulting return
212 * code for the operation.
214 * It lets the filesystem know if it registered an interest earlier via
215 * get_block. Pass the private field of the map buffer_head so that
216 * filesystems can use it to hold additional state between get_block calls and
217 * dio_complete.
219 static int dio_complete(struct dio *dio, loff_t offset, int ret)
221 ssize_t transferred = 0;
224 * AIO submission can race with bio completion to get here while
225 * expecting to have the last io completed by bio completion.
226 * In that case -EIOCBQUEUED is in fact not an error we want
227 * to preserve through this call.
229 if (ret == -EIOCBQUEUED)
230 ret = 0;
232 if (dio->result) {
233 transferred = dio->result;
235 /* Check for short read case */
236 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
237 transferred = dio->i_size - offset;
240 if (dio->end_io && dio->result)
241 dio->end_io(dio->iocb, offset, transferred,
242 dio->map_bh.b_private);
243 if (dio->lock_type == DIO_LOCKING)
244 /* lockdep: non-owner release */
245 up_read_non_owner(&dio->inode->i_alloc_sem);
247 if (ret == 0)
248 ret = dio->page_errors;
249 if (ret == 0)
250 ret = dio->io_error;
251 if (ret == 0)
252 ret = transferred;
254 return ret;
257 static int dio_bio_complete(struct dio *dio, struct bio *bio);
259 * Asynchronous IO callback.
261 static void dio_bio_end_aio(struct bio *bio, int error)
263 struct dio *dio = bio->bi_private;
264 unsigned long remaining;
265 unsigned long flags;
267 /* cleanup the bio */
268 dio_bio_complete(dio, bio);
270 spin_lock_irqsave(&dio->bio_lock, flags);
271 remaining = --dio->refcount;
272 if (remaining == 1 && dio->waiter)
273 wake_up_process(dio->waiter);
274 spin_unlock_irqrestore(&dio->bio_lock, flags);
276 if (remaining == 0) {
277 int ret = dio_complete(dio, dio->iocb->ki_pos, 0);
278 aio_complete(dio->iocb, ret, 0);
279 kfree(dio);
284 * The BIO completion handler simply queues the BIO up for the process-context
285 * handler.
287 * During I/O bi_private points at the dio. After I/O, bi_private is used to
288 * implement a singly-linked list of completed BIOs, at dio->bio_list.
290 static void dio_bio_end_io(struct bio *bio, int error)
292 struct dio *dio = bio->bi_private;
293 unsigned long flags;
295 spin_lock_irqsave(&dio->bio_lock, flags);
296 bio->bi_private = dio->bio_list;
297 dio->bio_list = bio;
298 if (--dio->refcount == 1 && dio->waiter)
299 wake_up_process(dio->waiter);
300 spin_unlock_irqrestore(&dio->bio_lock, flags);
303 static int
304 dio_bio_alloc(struct dio *dio, struct block_device *bdev,
305 sector_t first_sector, int nr_vecs)
307 struct bio *bio;
309 bio = bio_alloc(GFP_KERNEL, nr_vecs);
311 bio->bi_bdev = bdev;
312 bio->bi_sector = first_sector;
313 if (dio->is_async)
314 bio->bi_end_io = dio_bio_end_aio;
315 else
316 bio->bi_end_io = dio_bio_end_io;
318 dio->bio = bio;
319 return 0;
323 * In the AIO read case we speculatively dirty the pages before starting IO.
324 * During IO completion, any of these pages which happen to have been written
325 * back will be redirtied by bio_check_pages_dirty().
327 * bios hold a dio reference between submit_bio and ->end_io.
329 static void dio_bio_submit(struct dio *dio)
331 struct bio *bio = dio->bio;
332 unsigned long flags;
334 bio->bi_private = dio;
336 spin_lock_irqsave(&dio->bio_lock, flags);
337 dio->refcount++;
338 spin_unlock_irqrestore(&dio->bio_lock, flags);
340 if (dio->is_async && dio->rw == READ)
341 bio_set_pages_dirty(bio);
343 submit_bio(dio->rw, bio);
345 dio->bio = NULL;
346 dio->boundary = 0;
350 * Release any resources in case of a failure
352 static void dio_cleanup(struct dio *dio)
354 while (dio_pages_present(dio))
355 page_cache_release(dio_get_page(dio));
359 * Wait for the next BIO to complete. Remove it and return it. NULL is
360 * returned once all BIOs have been completed. This must only be called once
361 * all bios have been issued so that dio->refcount can only decrease. This
362 * requires that that the caller hold a reference on the dio.
364 static struct bio *dio_await_one(struct dio *dio)
366 unsigned long flags;
367 struct bio *bio = NULL;
369 spin_lock_irqsave(&dio->bio_lock, flags);
372 * Wait as long as the list is empty and there are bios in flight. bio
373 * completion drops the count, maybe adds to the list, and wakes while
374 * holding the bio_lock so we don't need set_current_state()'s barrier
375 * and can call it after testing our condition.
377 while (dio->refcount > 1 && dio->bio_list == NULL) {
378 __set_current_state(TASK_UNINTERRUPTIBLE);
379 dio->waiter = current;
380 spin_unlock_irqrestore(&dio->bio_lock, flags);
381 io_schedule();
382 /* wake up sets us TASK_RUNNING */
383 spin_lock_irqsave(&dio->bio_lock, flags);
384 dio->waiter = NULL;
386 if (dio->bio_list) {
387 bio = dio->bio_list;
388 dio->bio_list = bio->bi_private;
390 spin_unlock_irqrestore(&dio->bio_lock, flags);
391 return bio;
395 * Process one completed BIO. No locks are held.
397 static int dio_bio_complete(struct dio *dio, struct bio *bio)
399 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
400 struct bio_vec *bvec = bio->bi_io_vec;
401 int page_no;
403 if (!uptodate)
404 dio->io_error = -EIO;
406 if (dio->is_async && dio->rw == READ) {
407 bio_check_pages_dirty(bio); /* transfers ownership */
408 } else {
409 for (page_no = 0; page_no < bio->bi_vcnt; page_no++) {
410 struct page *page = bvec[page_no].bv_page;
412 if (dio->rw == READ && !PageCompound(page))
413 set_page_dirty_lock(page);
414 page_cache_release(page);
416 bio_put(bio);
418 return uptodate ? 0 : -EIO;
422 * Wait on and process all in-flight BIOs. This must only be called once
423 * all bios have been issued so that the refcount can only decrease.
424 * This just waits for all bios to make it through dio_bio_complete. IO
425 * errors are propagated through dio->io_error and should be propagated via
426 * dio_complete().
428 static void dio_await_completion(struct dio *dio)
430 struct bio *bio;
431 do {
432 bio = dio_await_one(dio);
433 if (bio)
434 dio_bio_complete(dio, bio);
435 } while (bio);
439 * A really large O_DIRECT read or write can generate a lot of BIOs. So
440 * to keep the memory consumption sane we periodically reap any completed BIOs
441 * during the BIO generation phase.
443 * This also helps to limit the peak amount of pinned userspace memory.
445 static int dio_bio_reap(struct dio *dio)
447 int ret = 0;
449 if (dio->reap_counter++ >= 64) {
450 while (dio->bio_list) {
451 unsigned long flags;
452 struct bio *bio;
453 int ret2;
455 spin_lock_irqsave(&dio->bio_lock, flags);
456 bio = dio->bio_list;
457 dio->bio_list = bio->bi_private;
458 spin_unlock_irqrestore(&dio->bio_lock, flags);
459 ret2 = dio_bio_complete(dio, bio);
460 if (ret == 0)
461 ret = ret2;
463 dio->reap_counter = 0;
465 return ret;
469 * Call into the fs to map some more disk blocks. We record the current number
470 * of available blocks at dio->blocks_available. These are in units of the
471 * fs blocksize, (1 << inode->i_blkbits).
473 * The fs is allowed to map lots of blocks at once. If it wants to do that,
474 * it uses the passed inode-relative block number as the file offset, as usual.
476 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
477 * has remaining to do. The fs should not map more than this number of blocks.
479 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
480 * indicate how much contiguous disk space has been made available at
481 * bh->b_blocknr.
483 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
484 * This isn't very efficient...
486 * In the case of filesystem holes: the fs may return an arbitrarily-large
487 * hole by returning an appropriate value in b_size and by clearing
488 * buffer_mapped(). However the direct-io code will only process holes one
489 * block at a time - it will repeatedly call get_block() as it walks the hole.
491 static int get_more_blocks(struct dio *dio)
493 int ret;
494 struct buffer_head *map_bh = &dio->map_bh;
495 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
496 unsigned long fs_count; /* Number of filesystem-sized blocks */
497 unsigned long dio_count;/* Number of dio_block-sized blocks */
498 unsigned long blkmask;
499 int create;
502 * If there was a memory error and we've overwritten all the
503 * mapped blocks then we can now return that memory error
505 ret = dio->page_errors;
506 if (ret == 0) {
507 BUG_ON(dio->block_in_file >= dio->final_block_in_request);
508 fs_startblk = dio->block_in_file >> dio->blkfactor;
509 dio_count = dio->final_block_in_request - dio->block_in_file;
510 fs_count = dio_count >> dio->blkfactor;
511 blkmask = (1 << dio->blkfactor) - 1;
512 if (dio_count & blkmask)
513 fs_count++;
515 map_bh->b_state = 0;
516 map_bh->b_size = fs_count << dio->inode->i_blkbits;
518 create = dio->rw & WRITE;
519 if (dio->lock_type == DIO_LOCKING) {
520 if (dio->block_in_file < (i_size_read(dio->inode) >>
521 dio->blkbits))
522 create = 0;
523 } else if (dio->lock_type == DIO_NO_LOCKING) {
524 create = 0;
528 * For writes inside i_size we forbid block creations: only
529 * overwrites are permitted. We fall back to buffered writes
530 * at a higher level for inside-i_size block-instantiating
531 * writes.
533 ret = (*dio->get_block)(dio->inode, fs_startblk,
534 map_bh, create);
536 return ret;
540 * There is no bio. Make one now.
542 static int dio_new_bio(struct dio *dio, sector_t start_sector)
544 sector_t sector;
545 int ret, nr_pages;
547 ret = dio_bio_reap(dio);
548 if (ret)
549 goto out;
550 sector = start_sector << (dio->blkbits - 9);
551 nr_pages = min(dio->pages_in_io, bio_get_nr_vecs(dio->map_bh.b_bdev));
552 BUG_ON(nr_pages <= 0);
553 ret = dio_bio_alloc(dio, dio->map_bh.b_bdev, sector, nr_pages);
554 dio->boundary = 0;
555 out:
556 return ret;
560 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
561 * that was successful then update final_block_in_bio and take a ref against
562 * the just-added page.
564 * Return zero on success. Non-zero means the caller needs to start a new BIO.
566 static int dio_bio_add_page(struct dio *dio)
568 int ret;
570 ret = bio_add_page(dio->bio, dio->cur_page,
571 dio->cur_page_len, dio->cur_page_offset);
572 if (ret == dio->cur_page_len) {
574 * Decrement count only, if we are done with this page
576 if ((dio->cur_page_len + dio->cur_page_offset) == PAGE_SIZE)
577 dio->pages_in_io--;
578 page_cache_get(dio->cur_page);
579 dio->final_block_in_bio = dio->cur_page_block +
580 (dio->cur_page_len >> dio->blkbits);
581 ret = 0;
582 } else {
583 ret = 1;
585 return ret;
589 * Put cur_page under IO. The section of cur_page which is described by
590 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
591 * starts on-disk at cur_page_block.
593 * We take a ref against the page here (on behalf of its presence in the bio).
595 * The caller of this function is responsible for removing cur_page from the
596 * dio, and for dropping the refcount which came from that presence.
598 static int dio_send_cur_page(struct dio *dio)
600 int ret = 0;
602 if (dio->bio) {
604 * See whether this new request is contiguous with the old
606 if (dio->final_block_in_bio != dio->cur_page_block)
607 dio_bio_submit(dio);
609 * Submit now if the underlying fs is about to perform a
610 * metadata read
612 if (dio->boundary)
613 dio_bio_submit(dio);
616 if (dio->bio == NULL) {
617 ret = dio_new_bio(dio, dio->cur_page_block);
618 if (ret)
619 goto out;
622 if (dio_bio_add_page(dio) != 0) {
623 dio_bio_submit(dio);
624 ret = dio_new_bio(dio, dio->cur_page_block);
625 if (ret == 0) {
626 ret = dio_bio_add_page(dio);
627 BUG_ON(ret != 0);
630 out:
631 return ret;
635 * An autonomous function to put a chunk of a page under deferred IO.
637 * The caller doesn't actually know (or care) whether this piece of page is in
638 * a BIO, or is under IO or whatever. We just take care of all possible
639 * situations here. The separation between the logic of do_direct_IO() and
640 * that of submit_page_section() is important for clarity. Please don't break.
642 * The chunk of page starts on-disk at blocknr.
644 * We perform deferred IO, by recording the last-submitted page inside our
645 * private part of the dio structure. If possible, we just expand the IO
646 * across that page here.
648 * If that doesn't work out then we put the old page into the bio and add this
649 * page to the dio instead.
651 static int
652 submit_page_section(struct dio *dio, struct page *page,
653 unsigned offset, unsigned len, sector_t blocknr)
655 int ret = 0;
657 if (dio->rw & WRITE) {
659 * Read accounting is performed in submit_bio()
661 task_io_account_write(len);
665 * Can we just grow the current page's presence in the dio?
667 if ( (dio->cur_page == page) &&
668 (dio->cur_page_offset + dio->cur_page_len == offset) &&
669 (dio->cur_page_block +
670 (dio->cur_page_len >> dio->blkbits) == blocknr)) {
671 dio->cur_page_len += len;
674 * If dio->boundary then we want to schedule the IO now to
675 * avoid metadata seeks.
677 if (dio->boundary) {
678 ret = dio_send_cur_page(dio);
679 page_cache_release(dio->cur_page);
680 dio->cur_page = NULL;
682 goto out;
686 * If there's a deferred page already there then send it.
688 if (dio->cur_page) {
689 ret = dio_send_cur_page(dio);
690 page_cache_release(dio->cur_page);
691 dio->cur_page = NULL;
692 if (ret)
693 goto out;
696 page_cache_get(page); /* It is in dio */
697 dio->cur_page = page;
698 dio->cur_page_offset = offset;
699 dio->cur_page_len = len;
700 dio->cur_page_block = blocknr;
701 out:
702 return ret;
706 * Clean any dirty buffers in the blockdev mapping which alias newly-created
707 * file blocks. Only called for S_ISREG files - blockdevs do not set
708 * buffer_new
710 static void clean_blockdev_aliases(struct dio *dio)
712 unsigned i;
713 unsigned nblocks;
715 nblocks = dio->map_bh.b_size >> dio->inode->i_blkbits;
717 for (i = 0; i < nblocks; i++) {
718 unmap_underlying_metadata(dio->map_bh.b_bdev,
719 dio->map_bh.b_blocknr + i);
724 * If we are not writing the entire block and get_block() allocated
725 * the block for us, we need to fill-in the unused portion of the
726 * block with zeros. This happens only if user-buffer, fileoffset or
727 * io length is not filesystem block-size multiple.
729 * `end' is zero if we're doing the start of the IO, 1 at the end of the
730 * IO.
732 static void dio_zero_block(struct dio *dio, int end)
734 unsigned dio_blocks_per_fs_block;
735 unsigned this_chunk_blocks; /* In dio_blocks */
736 unsigned this_chunk_bytes;
737 struct page *page;
739 dio->start_zero_done = 1;
740 if (!dio->blkfactor || !buffer_new(&dio->map_bh))
741 return;
743 dio_blocks_per_fs_block = 1 << dio->blkfactor;
744 this_chunk_blocks = dio->block_in_file & (dio_blocks_per_fs_block - 1);
746 if (!this_chunk_blocks)
747 return;
750 * We need to zero out part of an fs block. It is either at the
751 * beginning or the end of the fs block.
753 if (end)
754 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
756 this_chunk_bytes = this_chunk_blocks << dio->blkbits;
758 page = ZERO_PAGE(0);
759 if (submit_page_section(dio, page, 0, this_chunk_bytes,
760 dio->next_block_for_io))
761 return;
763 dio->next_block_for_io += this_chunk_blocks;
767 * Walk the user pages, and the file, mapping blocks to disk and generating
768 * a sequence of (page,offset,len,block) mappings. These mappings are injected
769 * into submit_page_section(), which takes care of the next stage of submission
771 * Direct IO against a blockdev is different from a file. Because we can
772 * happily perform page-sized but 512-byte aligned IOs. It is important that
773 * blockdev IO be able to have fine alignment and large sizes.
775 * So what we do is to permit the ->get_block function to populate bh.b_size
776 * with the size of IO which is permitted at this offset and this i_blkbits.
778 * For best results, the blockdev should be set up with 512-byte i_blkbits and
779 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
780 * fine alignment but still allows this function to work in PAGE_SIZE units.
782 static int do_direct_IO(struct dio *dio)
784 const unsigned blkbits = dio->blkbits;
785 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
786 struct page *page;
787 unsigned block_in_page;
788 struct buffer_head *map_bh = &dio->map_bh;
789 int ret = 0;
791 /* The I/O can start at any block offset within the first page */
792 block_in_page = dio->first_block_in_page;
794 while (dio->block_in_file < dio->final_block_in_request) {
795 page = dio_get_page(dio);
796 if (IS_ERR(page)) {
797 ret = PTR_ERR(page);
798 goto out;
801 while (block_in_page < blocks_per_page) {
802 unsigned offset_in_page = block_in_page << blkbits;
803 unsigned this_chunk_bytes; /* # of bytes mapped */
804 unsigned this_chunk_blocks; /* # of blocks */
805 unsigned u;
807 if (dio->blocks_available == 0) {
809 * Need to go and map some more disk
811 unsigned long blkmask;
812 unsigned long dio_remainder;
814 ret = get_more_blocks(dio);
815 if (ret) {
816 page_cache_release(page);
817 goto out;
819 if (!buffer_mapped(map_bh))
820 goto do_holes;
822 dio->blocks_available =
823 map_bh->b_size >> dio->blkbits;
824 dio->next_block_for_io =
825 map_bh->b_blocknr << dio->blkfactor;
826 if (buffer_new(map_bh))
827 clean_blockdev_aliases(dio);
829 if (!dio->blkfactor)
830 goto do_holes;
832 blkmask = (1 << dio->blkfactor) - 1;
833 dio_remainder = (dio->block_in_file & blkmask);
836 * If we are at the start of IO and that IO
837 * starts partway into a fs-block,
838 * dio_remainder will be non-zero. If the IO
839 * is a read then we can simply advance the IO
840 * cursor to the first block which is to be
841 * read. But if the IO is a write and the
842 * block was newly allocated we cannot do that;
843 * the start of the fs block must be zeroed out
844 * on-disk
846 if (!buffer_new(map_bh))
847 dio->next_block_for_io += dio_remainder;
848 dio->blocks_available -= dio_remainder;
850 do_holes:
851 /* Handle holes */
852 if (!buffer_mapped(map_bh)) {
853 loff_t i_size_aligned;
855 /* AKPM: eargh, -ENOTBLK is a hack */
856 if (dio->rw & WRITE) {
857 page_cache_release(page);
858 return -ENOTBLK;
862 * Be sure to account for a partial block as the
863 * last block in the file
865 i_size_aligned = ALIGN(i_size_read(dio->inode),
866 1 << blkbits);
867 if (dio->block_in_file >=
868 i_size_aligned >> blkbits) {
869 /* We hit eof */
870 page_cache_release(page);
871 goto out;
873 zero_user(page, block_in_page << blkbits,
874 1 << blkbits);
875 dio->block_in_file++;
876 block_in_page++;
877 goto next_block;
881 * If we're performing IO which has an alignment which
882 * is finer than the underlying fs, go check to see if
883 * we must zero out the start of this block.
885 if (unlikely(dio->blkfactor && !dio->start_zero_done))
886 dio_zero_block(dio, 0);
889 * Work out, in this_chunk_blocks, how much disk we
890 * can add to this page
892 this_chunk_blocks = dio->blocks_available;
893 u = (PAGE_SIZE - offset_in_page) >> blkbits;
894 if (this_chunk_blocks > u)
895 this_chunk_blocks = u;
896 u = dio->final_block_in_request - dio->block_in_file;
897 if (this_chunk_blocks > u)
898 this_chunk_blocks = u;
899 this_chunk_bytes = this_chunk_blocks << blkbits;
900 BUG_ON(this_chunk_bytes == 0);
902 dio->boundary = buffer_boundary(map_bh);
903 ret = submit_page_section(dio, page, offset_in_page,
904 this_chunk_bytes, dio->next_block_for_io);
905 if (ret) {
906 page_cache_release(page);
907 goto out;
909 dio->next_block_for_io += this_chunk_blocks;
911 dio->block_in_file += this_chunk_blocks;
912 block_in_page += this_chunk_blocks;
913 dio->blocks_available -= this_chunk_blocks;
914 next_block:
915 BUG_ON(dio->block_in_file > dio->final_block_in_request);
916 if (dio->block_in_file == dio->final_block_in_request)
917 break;
920 /* Drop the ref which was taken in get_user_pages() */
921 page_cache_release(page);
922 block_in_page = 0;
924 out:
925 return ret;
929 * Releases both i_mutex and i_alloc_sem
931 static ssize_t
932 direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
933 const struct iovec *iov, loff_t offset, unsigned long nr_segs,
934 unsigned blkbits, get_block_t get_block, dio_iodone_t end_io,
935 struct dio *dio)
937 unsigned long user_addr;
938 unsigned long flags;
939 int seg;
940 ssize_t ret = 0;
941 ssize_t ret2;
942 size_t bytes;
944 dio->inode = inode;
945 dio->rw = rw;
946 dio->blkbits = blkbits;
947 dio->blkfactor = inode->i_blkbits - blkbits;
948 dio->block_in_file = offset >> blkbits;
950 dio->get_block = get_block;
951 dio->end_io = end_io;
952 dio->final_block_in_bio = -1;
953 dio->next_block_for_io = -1;
955 dio->iocb = iocb;
956 dio->i_size = i_size_read(inode);
958 spin_lock_init(&dio->bio_lock);
959 dio->refcount = 1;
962 * In case of non-aligned buffers, we may need 2 more
963 * pages since we need to zero out first and last block.
965 if (unlikely(dio->blkfactor))
966 dio->pages_in_io = 2;
968 for (seg = 0; seg < nr_segs; seg++) {
969 user_addr = (unsigned long)iov[seg].iov_base;
970 dio->pages_in_io +=
971 ((user_addr+iov[seg].iov_len +PAGE_SIZE-1)/PAGE_SIZE
972 - user_addr/PAGE_SIZE);
975 for (seg = 0; seg < nr_segs; seg++) {
976 user_addr = (unsigned long)iov[seg].iov_base;
977 dio->size += bytes = iov[seg].iov_len;
979 /* Index into the first page of the first block */
980 dio->first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
981 dio->final_block_in_request = dio->block_in_file +
982 (bytes >> blkbits);
983 /* Page fetching state */
984 dio->head = 0;
985 dio->tail = 0;
986 dio->curr_page = 0;
988 dio->total_pages = 0;
989 if (user_addr & (PAGE_SIZE-1)) {
990 dio->total_pages++;
991 bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
993 dio->total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
994 dio->curr_user_address = user_addr;
996 ret = do_direct_IO(dio);
998 dio->result += iov[seg].iov_len -
999 ((dio->final_block_in_request - dio->block_in_file) <<
1000 blkbits);
1002 if (ret) {
1003 dio_cleanup(dio);
1004 break;
1006 } /* end iovec loop */
1008 if (ret == -ENOTBLK && (rw & WRITE)) {
1010 * The remaining part of the request will be
1011 * be handled by buffered I/O when we return
1013 ret = 0;
1016 * There may be some unwritten disk at the end of a part-written
1017 * fs-block-sized block. Go zero that now.
1019 dio_zero_block(dio, 1);
1021 if (dio->cur_page) {
1022 ret2 = dio_send_cur_page(dio);
1023 if (ret == 0)
1024 ret = ret2;
1025 page_cache_release(dio->cur_page);
1026 dio->cur_page = NULL;
1028 if (dio->bio)
1029 dio_bio_submit(dio);
1031 /* All IO is now issued, send it on its way */
1032 blk_run_address_space(inode->i_mapping);
1035 * It is possible that, we return short IO due to end of file.
1036 * In that case, we need to release all the pages we got hold on.
1038 dio_cleanup(dio);
1041 * All block lookups have been performed. For READ requests
1042 * we can let i_mutex go now that its achieved its purpose
1043 * of protecting us from looking up uninitialized blocks.
1045 if ((rw == READ) && (dio->lock_type == DIO_LOCKING))
1046 mutex_unlock(&dio->inode->i_mutex);
1049 * The only time we want to leave bios in flight is when a successful
1050 * partial aio read or full aio write have been setup. In that case
1051 * bio completion will call aio_complete. The only time it's safe to
1052 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1053 * This had *better* be the only place that raises -EIOCBQUEUED.
1055 BUG_ON(ret == -EIOCBQUEUED);
1056 if (dio->is_async && ret == 0 && dio->result &&
1057 ((rw & READ) || (dio->result == dio->size)))
1058 ret = -EIOCBQUEUED;
1060 if (ret != -EIOCBQUEUED)
1061 dio_await_completion(dio);
1064 * Sync will always be dropping the final ref and completing the
1065 * operation. AIO can if it was a broken operation described above or
1066 * in fact if all the bios race to complete before we get here. In
1067 * that case dio_complete() translates the EIOCBQUEUED into the proper
1068 * return code that the caller will hand to aio_complete().
1070 * This is managed by the bio_lock instead of being an atomic_t so that
1071 * completion paths can drop their ref and use the remaining count to
1072 * decide to wake the submission path atomically.
1074 spin_lock_irqsave(&dio->bio_lock, flags);
1075 ret2 = --dio->refcount;
1076 spin_unlock_irqrestore(&dio->bio_lock, flags);
1078 if (ret2 == 0) {
1079 ret = dio_complete(dio, offset, ret);
1080 kfree(dio);
1081 } else
1082 BUG_ON(ret != -EIOCBQUEUED);
1084 return ret;
1088 * This is a library function for use by filesystem drivers.
1089 * The locking rules are governed by the dio_lock_type parameter.
1091 * DIO_NO_LOCKING (no locking, for raw block device access)
1092 * For writes, i_mutex is not held on entry; it is never taken.
1094 * DIO_LOCKING (simple locking for regular files)
1095 * For writes we are called under i_mutex and return with i_mutex held, even
1096 * though it is internally dropped.
1097 * For reads, i_mutex is not held on entry, but it is taken and dropped before
1098 * returning.
1100 * DIO_OWN_LOCKING (filesystem provides synchronisation and handling of
1101 * uninitialised data, allowing parallel direct readers and writers)
1102 * For writes we are called without i_mutex, return without it, never touch it.
1103 * For reads we are called under i_mutex and return with i_mutex held, even
1104 * though it may be internally dropped.
1106 * Additional i_alloc_sem locking requirements described inline below.
1108 ssize_t
1109 __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1110 struct block_device *bdev, const struct iovec *iov, loff_t offset,
1111 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1112 int dio_lock_type)
1114 int seg;
1115 size_t size;
1116 unsigned long addr;
1117 unsigned blkbits = inode->i_blkbits;
1118 unsigned bdev_blkbits = 0;
1119 unsigned blocksize_mask = (1 << blkbits) - 1;
1120 ssize_t retval = -EINVAL;
1121 loff_t end = offset;
1122 struct dio *dio;
1123 int release_i_mutex = 0;
1124 int acquire_i_mutex = 0;
1126 if (rw & WRITE)
1127 rw = WRITE_ODIRECT;
1129 if (bdev)
1130 bdev_blkbits = blksize_bits(bdev_logical_block_size(bdev));
1132 if (offset & blocksize_mask) {
1133 if (bdev)
1134 blkbits = bdev_blkbits;
1135 blocksize_mask = (1 << blkbits) - 1;
1136 if (offset & blocksize_mask)
1137 goto out;
1140 /* Check the memory alignment. Blocks cannot straddle pages */
1141 for (seg = 0; seg < nr_segs; seg++) {
1142 addr = (unsigned long)iov[seg].iov_base;
1143 size = iov[seg].iov_len;
1144 end += size;
1145 if ((addr & blocksize_mask) || (size & blocksize_mask)) {
1146 if (bdev)
1147 blkbits = bdev_blkbits;
1148 blocksize_mask = (1 << blkbits) - 1;
1149 if ((addr & blocksize_mask) || (size & blocksize_mask))
1150 goto out;
1154 dio = kzalloc(sizeof(*dio), GFP_KERNEL);
1155 retval = -ENOMEM;
1156 if (!dio)
1157 goto out;
1160 * For block device access DIO_NO_LOCKING is used,
1161 * neither readers nor writers do any locking at all
1162 * For regular files using DIO_LOCKING,
1163 * readers need to grab i_mutex and i_alloc_sem
1164 * writers need to grab i_alloc_sem only (i_mutex is already held)
1165 * For regular files using DIO_OWN_LOCKING,
1166 * neither readers nor writers take any locks here
1168 dio->lock_type = dio_lock_type;
1169 if (dio_lock_type != DIO_NO_LOCKING) {
1170 /* watch out for a 0 len io from a tricksy fs */
1171 if (rw == READ && end > offset) {
1172 struct address_space *mapping;
1174 mapping = iocb->ki_filp->f_mapping;
1175 if (dio_lock_type != DIO_OWN_LOCKING) {
1176 mutex_lock(&inode->i_mutex);
1177 release_i_mutex = 1;
1180 retval = filemap_write_and_wait_range(mapping, offset,
1181 end - 1);
1182 if (retval) {
1183 kfree(dio);
1184 goto out;
1187 if (dio_lock_type == DIO_OWN_LOCKING) {
1188 mutex_unlock(&inode->i_mutex);
1189 acquire_i_mutex = 1;
1193 if (dio_lock_type == DIO_LOCKING)
1194 /* lockdep: not the owner will release it */
1195 down_read_non_owner(&inode->i_alloc_sem);
1199 * For file extending writes updating i_size before data
1200 * writeouts complete can expose uninitialized blocks. So
1201 * even for AIO, we need to wait for i/o to complete before
1202 * returning in this case.
1204 dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
1205 (end > i_size_read(inode)));
1207 retval = direct_io_worker(rw, iocb, inode, iov, offset,
1208 nr_segs, blkbits, get_block, end_io, dio);
1211 * In case of error extending write may have instantiated a few
1212 * blocks outside i_size. Trim these off again for DIO_LOCKING.
1213 * NOTE: DIO_NO_LOCK/DIO_OWN_LOCK callers have to handle this by
1214 * it's own meaner.
1216 if (unlikely(retval < 0 && (rw & WRITE))) {
1217 loff_t isize = i_size_read(inode);
1219 if (end > isize && dio_lock_type == DIO_LOCKING)
1220 vmtruncate(inode, isize);
1223 if (rw == READ && dio_lock_type == DIO_LOCKING)
1224 release_i_mutex = 0;
1226 out:
1227 if (release_i_mutex)
1228 mutex_unlock(&inode->i_mutex);
1229 else if (acquire_i_mutex)
1230 mutex_lock(&inode->i_mutex);
1231 return retval;
1233 EXPORT_SYMBOL(__blockdev_direct_IO);