blob: 27f3e787facaff34b29f28ed6830148a843ffbdb [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * fs/direct-io.c
3 *
4 * Copyright (C) 2002, Linus Torvalds.
5 *
6 * O_DIRECT
7 *
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.
20 */
21
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>
38
39/*
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.
42 */
43#define DIO_PAGES 64
44
45/*
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.
52 *
53 * If blkfactor is zero then the user's request was aligned to the filesystem's
54 * blocksize.
55 *
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
Jes Sorensen1b1dcc12006-01-09 15:59:24 -080059 * (blockdev) this locking is not done, and if it is DIO_OWN_LOCKING i_mutex is
Linus Torvalds1da177e2005-04-16 15:20:36 -070060 * 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).
62 */
63
64struct dio {
65 /* BIO submission state */
66 struct bio *bio; /* bio under assembly */
67 struct inode *inode;
68 int rw;
Daniel McNeil29504ff2005-04-16 15:25:50 -070069 loff_t i_size; /* i_size when submitted */
Linus Torvalds1da177e2005-04-16 15:20:36 -070070 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_blocks_t *get_blocks; /* 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_blocks() result */
95
96 /*
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().
100 */
101 struct page *cur_page; /* The page */
102 unsigned cur_page_offset; /* Offset into it, in bytes */
103 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
104 sector_t cur_page_block; /* Where it starts */
105
106 /*
107 * Page fetching state. These variables belong to dio_refill_pages().
108 */
109 int curr_page; /* changes */
110 int total_pages; /* doesn't change */
111 unsigned long curr_user_address;/* changes */
112
113 /*
114 * Page queue. These variables belong to dio_refill_pages() and
115 * dio_get_page().
116 */
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() */
121
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) */
128
129 /* AIO related stuff */
130 struct kiocb *iocb; /* kiocb */
131 int is_async; /* is IO async ? */
132 ssize_t result; /* IO result */
133};
134
135/*
136 * How many pages are in the queue?
137 */
138static inline unsigned dio_pages_present(struct dio *dio)
139{
140 return dio->tail - dio->head;
141}
142
143/*
144 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
145 */
146static int dio_refill_pages(struct dio *dio)
147{
148 int ret;
149 int nr_pages;
150
151 nr_pages = min(dio->total_pages - dio->curr_page, DIO_PAGES);
152 down_read(&current->mm->mmap_sem);
153 ret = get_user_pages(
154 current, /* Task for fault acounting */
155 current->mm, /* whose pages? */
156 dio->curr_user_address, /* Where from? */
157 nr_pages, /* How many pages? */
158 dio->rw == READ, /* Write to memory? */
159 0, /* force (?) */
160 &dio->pages[0],
161 NULL); /* vmas */
162 up_read(&current->mm->mmap_sem);
163
164 if (ret < 0 && dio->blocks_available && (dio->rw == WRITE)) {
Nick Pigginb5810032005-10-29 18:16:12 -0700165 struct page *page = ZERO_PAGE(dio->curr_user_address);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166 /*
167 * A memory fault, but the filesystem has some outstanding
168 * mapped blocks. We need to use those blocks up to avoid
169 * leaking stale data in the file.
170 */
171 if (dio->page_errors == 0)
172 dio->page_errors = ret;
Nick Pigginb5810032005-10-29 18:16:12 -0700173 page_cache_get(page);
174 dio->pages[0] = page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700175 dio->head = 0;
176 dio->tail = 1;
177 ret = 0;
178 goto out;
179 }
180
181 if (ret >= 0) {
182 dio->curr_user_address += ret * PAGE_SIZE;
183 dio->curr_page += ret;
184 dio->head = 0;
185 dio->tail = ret;
186 ret = 0;
187 }
188out:
189 return ret;
190}
191
192/*
193 * Get another userspace page. Returns an ERR_PTR on error. Pages are
194 * buffered inside the dio so that we can call get_user_pages() against a
195 * decent number of pages, less frequently. To provide nicer use of the
196 * L1 cache.
197 */
198static struct page *dio_get_page(struct dio *dio)
199{
200 if (dio_pages_present(dio) == 0) {
201 int ret;
202
203 ret = dio_refill_pages(dio);
204 if (ret)
205 return ERR_PTR(ret);
206 BUG_ON(dio_pages_present(dio) == 0);
207 }
208 return dio->pages[dio->head++];
209}
210
211/*
212 * Called when all DIO BIO I/O has been completed - let the filesystem
213 * know, if it registered an interest earlier via get_blocks. Pass the
214 * private field of the map buffer_head so that filesystems can use it
215 * to hold additional state between get_blocks calls and dio_complete.
216 */
217static void dio_complete(struct dio *dio, loff_t offset, ssize_t bytes)
218{
219 if (dio->end_io && dio->result)
Christoph Hellwig92198f72005-06-23 22:00:59 -0700220 dio->end_io(dio->iocb, offset, bytes, dio->map_bh.b_private);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700221 if (dio->lock_type == DIO_LOCKING)
222 up_read(&dio->inode->i_alloc_sem);
223}
224
225/*
226 * Called when a BIO has been processed. If the count goes to zero then IO is
227 * complete and we can signal this to the AIO layer.
228 */
229static void finished_one_bio(struct dio *dio)
230{
231 unsigned long flags;
232
233 spin_lock_irqsave(&dio->bio_lock, flags);
234 if (dio->bio_count == 1) {
235 if (dio->is_async) {
Daniel McNeil29504ff2005-04-16 15:25:50 -0700236 ssize_t transferred;
237 loff_t offset;
238
Linus Torvalds1da177e2005-04-16 15:20:36 -0700239 /*
240 * Last reference to the dio is going away.
241 * Drop spinlock and complete the DIO.
242 */
243 spin_unlock_irqrestore(&dio->bio_lock, flags);
Daniel McNeil29504ff2005-04-16 15:25:50 -0700244
245 /* Check for short read case */
246 transferred = dio->result;
247 offset = dio->iocb->ki_pos;
248
249 if ((dio->rw == READ) &&
250 ((offset + transferred) > dio->i_size))
251 transferred = dio->i_size - offset;
252
253 dio_complete(dio, offset, transferred);
254
Linus Torvalds1da177e2005-04-16 15:20:36 -0700255 /* Complete AIO later if falling back to buffered i/o */
256 if (dio->result == dio->size ||
257 ((dio->rw == READ) && dio->result)) {
Daniel McNeil29504ff2005-04-16 15:25:50 -0700258 aio_complete(dio->iocb, transferred, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259 kfree(dio);
260 return;
261 } else {
262 /*
263 * Falling back to buffered
264 */
265 spin_lock_irqsave(&dio->bio_lock, flags);
266 dio->bio_count--;
267 if (dio->waiter)
268 wake_up_process(dio->waiter);
269 spin_unlock_irqrestore(&dio->bio_lock, flags);
270 return;
271 }
272 }
273 }
274 dio->bio_count--;
275 spin_unlock_irqrestore(&dio->bio_lock, flags);
276}
277
278static int dio_bio_complete(struct dio *dio, struct bio *bio);
279/*
280 * Asynchronous IO callback.
281 */
282static int dio_bio_end_aio(struct bio *bio, unsigned int bytes_done, int error)
283{
284 struct dio *dio = bio->bi_private;
285
286 if (bio->bi_size)
287 return 1;
288
289 /* cleanup the bio */
290 dio_bio_complete(dio, bio);
291 return 0;
292}
293
294/*
295 * The BIO completion handler simply queues the BIO up for the process-context
296 * handler.
297 *
298 * During I/O bi_private points at the dio. After I/O, bi_private is used to
299 * implement a singly-linked list of completed BIOs, at dio->bio_list.
300 */
301static int dio_bio_end_io(struct bio *bio, unsigned int bytes_done, int error)
302{
303 struct dio *dio = bio->bi_private;
304 unsigned long flags;
305
306 if (bio->bi_size)
307 return 1;
308
309 spin_lock_irqsave(&dio->bio_lock, flags);
310 bio->bi_private = dio->bio_list;
311 dio->bio_list = bio;
312 dio->bios_in_flight--;
313 if (dio->waiter && dio->bios_in_flight == 0)
314 wake_up_process(dio->waiter);
315 spin_unlock_irqrestore(&dio->bio_lock, flags);
316 return 0;
317}
318
319static int
320dio_bio_alloc(struct dio *dio, struct block_device *bdev,
321 sector_t first_sector, int nr_vecs)
322{
323 struct bio *bio;
324
325 bio = bio_alloc(GFP_KERNEL, nr_vecs);
326 if (bio == NULL)
327 return -ENOMEM;
328
329 bio->bi_bdev = bdev;
330 bio->bi_sector = first_sector;
331 if (dio->is_async)
332 bio->bi_end_io = dio_bio_end_aio;
333 else
334 bio->bi_end_io = dio_bio_end_io;
335
336 dio->bio = bio;
337 return 0;
338}
339
340/*
341 * In the AIO read case we speculatively dirty the pages before starting IO.
342 * During IO completion, any of these pages which happen to have been written
343 * back will be redirtied by bio_check_pages_dirty().
344 */
345static void dio_bio_submit(struct dio *dio)
346{
347 struct bio *bio = dio->bio;
348 unsigned long flags;
349
350 bio->bi_private = dio;
351 spin_lock_irqsave(&dio->bio_lock, flags);
352 dio->bio_count++;
353 dio->bios_in_flight++;
354 spin_unlock_irqrestore(&dio->bio_lock, flags);
355 if (dio->is_async && dio->rw == READ)
356 bio_set_pages_dirty(bio);
357 submit_bio(dio->rw, bio);
358
359 dio->bio = NULL;
360 dio->boundary = 0;
361}
362
363/*
364 * Release any resources in case of a failure
365 */
366static void dio_cleanup(struct dio *dio)
367{
368 while (dio_pages_present(dio))
369 page_cache_release(dio_get_page(dio));
370}
371
372/*
373 * Wait for the next BIO to complete. Remove it and return it.
374 */
375static struct bio *dio_await_one(struct dio *dio)
376{
377 unsigned long flags;
378 struct bio *bio;
379
380 spin_lock_irqsave(&dio->bio_lock, flags);
381 while (dio->bio_list == NULL) {
382 set_current_state(TASK_UNINTERRUPTIBLE);
383 if (dio->bio_list == NULL) {
384 dio->waiter = current;
385 spin_unlock_irqrestore(&dio->bio_lock, flags);
386 blk_run_address_space(dio->inode->i_mapping);
387 io_schedule();
388 spin_lock_irqsave(&dio->bio_lock, flags);
389 dio->waiter = NULL;
390 }
391 set_current_state(TASK_RUNNING);
392 }
393 bio = dio->bio_list;
394 dio->bio_list = bio->bi_private;
395 spin_unlock_irqrestore(&dio->bio_lock, flags);
396 return bio;
397}
398
399/*
400 * Process one completed BIO. No locks are held.
401 */
402static int dio_bio_complete(struct dio *dio, struct bio *bio)
403{
404 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
405 struct bio_vec *bvec = bio->bi_io_vec;
406 int page_no;
407
408 if (!uptodate)
409 dio->result = -EIO;
410
411 if (dio->is_async && dio->rw == READ) {
412 bio_check_pages_dirty(bio); /* transfers ownership */
413 } else {
414 for (page_no = 0; page_no < bio->bi_vcnt; page_no++) {
415 struct page *page = bvec[page_no].bv_page;
416
417 if (dio->rw == READ && !PageCompound(page))
418 set_page_dirty_lock(page);
419 page_cache_release(page);
420 }
421 bio_put(bio);
422 }
423 finished_one_bio(dio);
424 return uptodate ? 0 : -EIO;
425}
426
427/*
428 * Wait on and process all in-flight BIOs.
429 */
430static int dio_await_completion(struct dio *dio)
431{
432 int ret = 0;
433
434 if (dio->bio)
435 dio_bio_submit(dio);
436
437 /*
438 * The bio_lock is not held for the read of bio_count.
439 * This is ok since it is the dio_bio_complete() that changes
440 * bio_count.
441 */
442 while (dio->bio_count) {
443 struct bio *bio = dio_await_one(dio);
444 int ret2;
445
446 ret2 = dio_bio_complete(dio, bio);
447 if (ret == 0)
448 ret = ret2;
449 }
450 return ret;
451}
452
453/*
454 * A really large O_DIRECT read or write can generate a lot of BIOs. So
455 * to keep the memory consumption sane we periodically reap any completed BIOs
456 * during the BIO generation phase.
457 *
458 * This also helps to limit the peak amount of pinned userspace memory.
459 */
460static int dio_bio_reap(struct dio *dio)
461{
462 int ret = 0;
463
464 if (dio->reap_counter++ >= 64) {
465 while (dio->bio_list) {
466 unsigned long flags;
467 struct bio *bio;
468 int ret2;
469
470 spin_lock_irqsave(&dio->bio_lock, flags);
471 bio = dio->bio_list;
472 dio->bio_list = bio->bi_private;
473 spin_unlock_irqrestore(&dio->bio_lock, flags);
474 ret2 = dio_bio_complete(dio, bio);
475 if (ret == 0)
476 ret = ret2;
477 }
478 dio->reap_counter = 0;
479 }
480 return ret;
481}
482
483/*
484 * Call into the fs to map some more disk blocks. We record the current number
485 * of available blocks at dio->blocks_available. These are in units of the
486 * fs blocksize, (1 << inode->i_blkbits).
487 *
488 * The fs is allowed to map lots of blocks at once. If it wants to do that,
489 * it uses the passed inode-relative block number as the file offset, as usual.
490 *
491 * get_blocks() is passed the number of i_blkbits-sized blocks which direct_io
492 * has remaining to do. The fs should not map more than this number of blocks.
493 *
494 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
495 * indicate how much contiguous disk space has been made available at
496 * bh->b_blocknr.
497 *
498 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
499 * This isn't very efficient...
500 *
501 * In the case of filesystem holes: the fs may return an arbitrarily-large
502 * hole by returning an appropriate value in b_size and by clearing
503 * buffer_mapped(). However the direct-io code will only process holes one
504 * block at a time - it will repeatedly call get_blocks() as it walks the hole.
505 */
506static int get_more_blocks(struct dio *dio)
507{
508 int ret;
509 struct buffer_head *map_bh = &dio->map_bh;
510 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
511 unsigned long fs_count; /* Number of filesystem-sized blocks */
512 unsigned long dio_count;/* Number of dio_block-sized blocks */
513 unsigned long blkmask;
514 int create;
515
516 /*
517 * If there was a memory error and we've overwritten all the
518 * mapped blocks then we can now return that memory error
519 */
520 ret = dio->page_errors;
521 if (ret == 0) {
522 map_bh->b_state = 0;
523 map_bh->b_size = 0;
524 BUG_ON(dio->block_in_file >= dio->final_block_in_request);
525 fs_startblk = dio->block_in_file >> dio->blkfactor;
526 dio_count = dio->final_block_in_request - dio->block_in_file;
527 fs_count = dio_count >> dio->blkfactor;
528 blkmask = (1 << dio->blkfactor) - 1;
529 if (dio_count & blkmask)
530 fs_count++;
531
532 create = dio->rw == WRITE;
533 if (dio->lock_type == DIO_LOCKING) {
534 if (dio->block_in_file < (i_size_read(dio->inode) >>
535 dio->blkbits))
536 create = 0;
537 } else if (dio->lock_type == DIO_NO_LOCKING) {
538 create = 0;
539 }
540 /*
541 * For writes inside i_size we forbid block creations: only
542 * overwrites are permitted. We fall back to buffered writes
543 * at a higher level for inside-i_size block-instantiating
544 * writes.
545 */
546 ret = (*dio->get_blocks)(dio->inode, fs_startblk, fs_count,
547 map_bh, create);
548 }
549 return ret;
550}
551
552/*
553 * There is no bio. Make one now.
554 */
555static int dio_new_bio(struct dio *dio, sector_t start_sector)
556{
557 sector_t sector;
558 int ret, nr_pages;
559
560 ret = dio_bio_reap(dio);
561 if (ret)
562 goto out;
563 sector = start_sector << (dio->blkbits - 9);
564 nr_pages = min(dio->pages_in_io, bio_get_nr_vecs(dio->map_bh.b_bdev));
565 BUG_ON(nr_pages <= 0);
566 ret = dio_bio_alloc(dio, dio->map_bh.b_bdev, sector, nr_pages);
567 dio->boundary = 0;
568out:
569 return ret;
570}
571
572/*
573 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
574 * that was successful then update final_block_in_bio and take a ref against
575 * the just-added page.
576 *
577 * Return zero on success. Non-zero means the caller needs to start a new BIO.
578 */
579static int dio_bio_add_page(struct dio *dio)
580{
581 int ret;
582
583 ret = bio_add_page(dio->bio, dio->cur_page,
584 dio->cur_page_len, dio->cur_page_offset);
585 if (ret == dio->cur_page_len) {
586 /*
587 * Decrement count only, if we are done with this page
588 */
589 if ((dio->cur_page_len + dio->cur_page_offset) == PAGE_SIZE)
590 dio->pages_in_io--;
591 page_cache_get(dio->cur_page);
592 dio->final_block_in_bio = dio->cur_page_block +
593 (dio->cur_page_len >> dio->blkbits);
594 ret = 0;
595 } else {
596 ret = 1;
597 }
598 return ret;
599}
600
601/*
602 * Put cur_page under IO. The section of cur_page which is described by
603 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
604 * starts on-disk at cur_page_block.
605 *
606 * We take a ref against the page here (on behalf of its presence in the bio).
607 *
608 * The caller of this function is responsible for removing cur_page from the
609 * dio, and for dropping the refcount which came from that presence.
610 */
611static int dio_send_cur_page(struct dio *dio)
612{
613 int ret = 0;
614
615 if (dio->bio) {
616 /*
617 * See whether this new request is contiguous with the old
618 */
619 if (dio->final_block_in_bio != dio->cur_page_block)
620 dio_bio_submit(dio);
621 /*
622 * Submit now if the underlying fs is about to perform a
623 * metadata read
624 */
625 if (dio->boundary)
626 dio_bio_submit(dio);
627 }
628
629 if (dio->bio == NULL) {
630 ret = dio_new_bio(dio, dio->cur_page_block);
631 if (ret)
632 goto out;
633 }
634
635 if (dio_bio_add_page(dio) != 0) {
636 dio_bio_submit(dio);
637 ret = dio_new_bio(dio, dio->cur_page_block);
638 if (ret == 0) {
639 ret = dio_bio_add_page(dio);
640 BUG_ON(ret != 0);
641 }
642 }
643out:
644 return ret;
645}
646
647/*
648 * An autonomous function to put a chunk of a page under deferred IO.
649 *
650 * The caller doesn't actually know (or care) whether this piece of page is in
651 * a BIO, or is under IO or whatever. We just take care of all possible
652 * situations here. The separation between the logic of do_direct_IO() and
653 * that of submit_page_section() is important for clarity. Please don't break.
654 *
655 * The chunk of page starts on-disk at blocknr.
656 *
657 * We perform deferred IO, by recording the last-submitted page inside our
658 * private part of the dio structure. If possible, we just expand the IO
659 * across that page here.
660 *
661 * If that doesn't work out then we put the old page into the bio and add this
662 * page to the dio instead.
663 */
664static int
665submit_page_section(struct dio *dio, struct page *page,
666 unsigned offset, unsigned len, sector_t blocknr)
667{
668 int ret = 0;
669
670 /*
671 * Can we just grow the current page's presence in the dio?
672 */
673 if ( (dio->cur_page == page) &&
674 (dio->cur_page_offset + dio->cur_page_len == offset) &&
675 (dio->cur_page_block +
676 (dio->cur_page_len >> dio->blkbits) == blocknr)) {
677 dio->cur_page_len += len;
678
679 /*
680 * If dio->boundary then we want to schedule the IO now to
681 * avoid metadata seeks.
682 */
683 if (dio->boundary) {
684 ret = dio_send_cur_page(dio);
685 page_cache_release(dio->cur_page);
686 dio->cur_page = NULL;
687 }
688 goto out;
689 }
690
691 /*
692 * If there's a deferred page already there then send it.
693 */
694 if (dio->cur_page) {
695 ret = dio_send_cur_page(dio);
696 page_cache_release(dio->cur_page);
697 dio->cur_page = NULL;
698 if (ret)
699 goto out;
700 }
701
702 page_cache_get(page); /* It is in dio */
703 dio->cur_page = page;
704 dio->cur_page_offset = offset;
705 dio->cur_page_len = len;
706 dio->cur_page_block = blocknr;
707out:
708 return ret;
709}
710
711/*
712 * Clean any dirty buffers in the blockdev mapping which alias newly-created
713 * file blocks. Only called for S_ISREG files - blockdevs do not set
714 * buffer_new
715 */
716static void clean_blockdev_aliases(struct dio *dio)
717{
718 unsigned i;
719 unsigned nblocks;
720
721 nblocks = dio->map_bh.b_size >> dio->inode->i_blkbits;
722
723 for (i = 0; i < nblocks; i++) {
724 unmap_underlying_metadata(dio->map_bh.b_bdev,
725 dio->map_bh.b_blocknr + i);
726 }
727}
728
729/*
730 * If we are not writing the entire block and get_block() allocated
731 * the block for us, we need to fill-in the unused portion of the
732 * block with zeros. This happens only if user-buffer, fileoffset or
733 * io length is not filesystem block-size multiple.
734 *
735 * `end' is zero if we're doing the start of the IO, 1 at the end of the
736 * IO.
737 */
738static void dio_zero_block(struct dio *dio, int end)
739{
740 unsigned dio_blocks_per_fs_block;
741 unsigned this_chunk_blocks; /* In dio_blocks */
742 unsigned this_chunk_bytes;
743 struct page *page;
744
745 dio->start_zero_done = 1;
746 if (!dio->blkfactor || !buffer_new(&dio->map_bh))
747 return;
748
749 dio_blocks_per_fs_block = 1 << dio->blkfactor;
750 this_chunk_blocks = dio->block_in_file & (dio_blocks_per_fs_block - 1);
751
752 if (!this_chunk_blocks)
753 return;
754
755 /*
756 * We need to zero out part of an fs block. It is either at the
757 * beginning or the end of the fs block.
758 */
759 if (end)
760 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
761
762 this_chunk_bytes = this_chunk_blocks << dio->blkbits;
763
764 page = ZERO_PAGE(dio->curr_user_address);
765 if (submit_page_section(dio, page, 0, this_chunk_bytes,
766 dio->next_block_for_io))
767 return;
768
769 dio->next_block_for_io += this_chunk_blocks;
770}
771
772/*
773 * Walk the user pages, and the file, mapping blocks to disk and generating
774 * a sequence of (page,offset,len,block) mappings. These mappings are injected
775 * into submit_page_section(), which takes care of the next stage of submission
776 *
777 * Direct IO against a blockdev is different from a file. Because we can
778 * happily perform page-sized but 512-byte aligned IOs. It is important that
779 * blockdev IO be able to have fine alignment and large sizes.
780 *
781 * So what we do is to permit the ->get_blocks function to populate bh.b_size
782 * with the size of IO which is permitted at this offset and this i_blkbits.
783 *
784 * For best results, the blockdev should be set up with 512-byte i_blkbits and
785 * it should set b_size to PAGE_SIZE or more inside get_blocks(). This gives
786 * fine alignment but still allows this function to work in PAGE_SIZE units.
787 */
788static int do_direct_IO(struct dio *dio)
789{
790 const unsigned blkbits = dio->blkbits;
791 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
792 struct page *page;
793 unsigned block_in_page;
794 struct buffer_head *map_bh = &dio->map_bh;
795 int ret = 0;
796
797 /* The I/O can start at any block offset within the first page */
798 block_in_page = dio->first_block_in_page;
799
800 while (dio->block_in_file < dio->final_block_in_request) {
801 page = dio_get_page(dio);
802 if (IS_ERR(page)) {
803 ret = PTR_ERR(page);
804 goto out;
805 }
806
807 while (block_in_page < blocks_per_page) {
808 unsigned offset_in_page = block_in_page << blkbits;
809 unsigned this_chunk_bytes; /* # of bytes mapped */
810 unsigned this_chunk_blocks; /* # of blocks */
811 unsigned u;
812
813 if (dio->blocks_available == 0) {
814 /*
815 * Need to go and map some more disk
816 */
817 unsigned long blkmask;
818 unsigned long dio_remainder;
819
820 ret = get_more_blocks(dio);
821 if (ret) {
822 page_cache_release(page);
823 goto out;
824 }
825 if (!buffer_mapped(map_bh))
826 goto do_holes;
827
828 dio->blocks_available =
829 map_bh->b_size >> dio->blkbits;
830 dio->next_block_for_io =
831 map_bh->b_blocknr << dio->blkfactor;
832 if (buffer_new(map_bh))
833 clean_blockdev_aliases(dio);
834
835 if (!dio->blkfactor)
836 goto do_holes;
837
838 blkmask = (1 << dio->blkfactor) - 1;
839 dio_remainder = (dio->block_in_file & blkmask);
840
841 /*
842 * If we are at the start of IO and that IO
843 * starts partway into a fs-block,
844 * dio_remainder will be non-zero. If the IO
845 * is a read then we can simply advance the IO
846 * cursor to the first block which is to be
847 * read. But if the IO is a write and the
848 * block was newly allocated we cannot do that;
849 * the start of the fs block must be zeroed out
850 * on-disk
851 */
852 if (!buffer_new(map_bh))
853 dio->next_block_for_io += dio_remainder;
854 dio->blocks_available -= dio_remainder;
855 }
856do_holes:
857 /* Handle holes */
858 if (!buffer_mapped(map_bh)) {
859 char *kaddr;
Jeff Moyer35dc8162006-02-03 03:04:27 -0800860 loff_t i_size_aligned;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861
862 /* AKPM: eargh, -ENOTBLK is a hack */
863 if (dio->rw == WRITE) {
864 page_cache_release(page);
865 return -ENOTBLK;
866 }
867
Jeff Moyer35dc8162006-02-03 03:04:27 -0800868 /*
869 * Be sure to account for a partial block as the
870 * last block in the file
871 */
872 i_size_aligned = ALIGN(i_size_read(dio->inode),
873 1 << blkbits);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700874 if (dio->block_in_file >=
Jeff Moyer35dc8162006-02-03 03:04:27 -0800875 i_size_aligned >> blkbits) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876 /* We hit eof */
877 page_cache_release(page);
878 goto out;
879 }
880 kaddr = kmap_atomic(page, KM_USER0);
881 memset(kaddr + (block_in_page << blkbits),
882 0, 1 << blkbits);
883 flush_dcache_page(page);
884 kunmap_atomic(kaddr, KM_USER0);
885 dio->block_in_file++;
886 block_in_page++;
887 goto next_block;
888 }
889
890 /*
891 * If we're performing IO which has an alignment which
892 * is finer than the underlying fs, go check to see if
893 * we must zero out the start of this block.
894 */
895 if (unlikely(dio->blkfactor && !dio->start_zero_done))
896 dio_zero_block(dio, 0);
897
898 /*
899 * Work out, in this_chunk_blocks, how much disk we
900 * can add to this page
901 */
902 this_chunk_blocks = dio->blocks_available;
903 u = (PAGE_SIZE - offset_in_page) >> blkbits;
904 if (this_chunk_blocks > u)
905 this_chunk_blocks = u;
906 u = dio->final_block_in_request - dio->block_in_file;
907 if (this_chunk_blocks > u)
908 this_chunk_blocks = u;
909 this_chunk_bytes = this_chunk_blocks << blkbits;
910 BUG_ON(this_chunk_bytes == 0);
911
912 dio->boundary = buffer_boundary(map_bh);
913 ret = submit_page_section(dio, page, offset_in_page,
914 this_chunk_bytes, dio->next_block_for_io);
915 if (ret) {
916 page_cache_release(page);
917 goto out;
918 }
919 dio->next_block_for_io += this_chunk_blocks;
920
921 dio->block_in_file += this_chunk_blocks;
922 block_in_page += this_chunk_blocks;
923 dio->blocks_available -= this_chunk_blocks;
924next_block:
925 if (dio->block_in_file > dio->final_block_in_request)
926 BUG();
927 if (dio->block_in_file == dio->final_block_in_request)
928 break;
929 }
930
931 /* Drop the ref which was taken in get_user_pages() */
932 page_cache_release(page);
933 block_in_page = 0;
934 }
935out:
936 return ret;
937}
938
939/*
Jes Sorensen1b1dcc12006-01-09 15:59:24 -0800940 * Releases both i_mutex and i_alloc_sem
Linus Torvalds1da177e2005-04-16 15:20:36 -0700941 */
942static ssize_t
943direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
944 const struct iovec *iov, loff_t offset, unsigned long nr_segs,
945 unsigned blkbits, get_blocks_t get_blocks, dio_iodone_t end_io,
946 struct dio *dio)
947{
948 unsigned long user_addr;
949 int seg;
950 ssize_t ret = 0;
951 ssize_t ret2;
952 size_t bytes;
953
954 dio->bio = NULL;
955 dio->inode = inode;
956 dio->rw = rw;
957 dio->blkbits = blkbits;
958 dio->blkfactor = inode->i_blkbits - blkbits;
959 dio->start_zero_done = 0;
960 dio->size = 0;
961 dio->block_in_file = offset >> blkbits;
962 dio->blocks_available = 0;
963 dio->cur_page = NULL;
964
965 dio->boundary = 0;
966 dio->reap_counter = 0;
967 dio->get_blocks = get_blocks;
968 dio->end_io = end_io;
969 dio->map_bh.b_private = NULL;
970 dio->final_block_in_bio = -1;
971 dio->next_block_for_io = -1;
972
973 dio->page_errors = 0;
974 dio->result = 0;
975 dio->iocb = iocb;
Daniel McNeil29504ff2005-04-16 15:25:50 -0700976 dio->i_size = i_size_read(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977
978 /*
979 * BIO completion state.
980 *
981 * ->bio_count starts out at one, and we decrement it to zero after all
982 * BIOs are submitted. This to avoid the situation where a really fast
983 * (or synchronous) device could take the count to zero while we're
984 * still submitting BIOs.
985 */
986 dio->bio_count = 1;
987 dio->bios_in_flight = 0;
988 spin_lock_init(&dio->bio_lock);
989 dio->bio_list = NULL;
990 dio->waiter = NULL;
991
992 /*
993 * In case of non-aligned buffers, we may need 2 more
994 * pages since we need to zero out first and last block.
995 */
996 if (unlikely(dio->blkfactor))
997 dio->pages_in_io = 2;
998 else
999 dio->pages_in_io = 0;
1000
1001 for (seg = 0; seg < nr_segs; seg++) {
1002 user_addr = (unsigned long)iov[seg].iov_base;
1003 dio->pages_in_io +=
1004 ((user_addr+iov[seg].iov_len +PAGE_SIZE-1)/PAGE_SIZE
1005 - user_addr/PAGE_SIZE);
1006 }
1007
1008 for (seg = 0; seg < nr_segs; seg++) {
1009 user_addr = (unsigned long)iov[seg].iov_base;
1010 dio->size += bytes = iov[seg].iov_len;
1011
1012 /* Index into the first page of the first block */
1013 dio->first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
1014 dio->final_block_in_request = dio->block_in_file +
1015 (bytes >> blkbits);
1016 /* Page fetching state */
1017 dio->head = 0;
1018 dio->tail = 0;
1019 dio->curr_page = 0;
1020
1021 dio->total_pages = 0;
1022 if (user_addr & (PAGE_SIZE-1)) {
1023 dio->total_pages++;
1024 bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
1025 }
1026 dio->total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
1027 dio->curr_user_address = user_addr;
1028
1029 ret = do_direct_IO(dio);
1030
1031 dio->result += iov[seg].iov_len -
1032 ((dio->final_block_in_request - dio->block_in_file) <<
1033 blkbits);
1034
1035 if (ret) {
1036 dio_cleanup(dio);
1037 break;
1038 }
1039 } /* end iovec loop */
1040
1041 if (ret == -ENOTBLK && rw == WRITE) {
1042 /*
1043 * The remaining part of the request will be
1044 * be handled by buffered I/O when we return
1045 */
1046 ret = 0;
1047 }
1048 /*
1049 * There may be some unwritten disk at the end of a part-written
1050 * fs-block-sized block. Go zero that now.
1051 */
1052 dio_zero_block(dio, 1);
1053
1054 if (dio->cur_page) {
1055 ret2 = dio_send_cur_page(dio);
1056 if (ret == 0)
1057 ret = ret2;
1058 page_cache_release(dio->cur_page);
1059 dio->cur_page = NULL;
1060 }
1061 if (dio->bio)
1062 dio_bio_submit(dio);
1063
1064 /*
1065 * It is possible that, we return short IO due to end of file.
1066 * In that case, we need to release all the pages we got hold on.
1067 */
1068 dio_cleanup(dio);
1069
1070 /*
1071 * All block lookups have been performed. For READ requests
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001072 * we can let i_mutex go now that its achieved its purpose
Linus Torvalds1da177e2005-04-16 15:20:36 -07001073 * of protecting us from looking up uninitialized blocks.
1074 */
1075 if ((rw == READ) && (dio->lock_type == DIO_LOCKING))
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001076 mutex_unlock(&dio->inode->i_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001077
1078 /*
1079 * OK, all BIOs are submitted, so we can decrement bio_count to truly
1080 * reflect the number of to-be-processed BIOs.
1081 */
1082 if (dio->is_async) {
1083 int should_wait = 0;
1084
1085 if (dio->result < dio->size && rw == WRITE) {
1086 dio->waiter = current;
1087 should_wait = 1;
1088 }
1089 if (ret == 0)
1090 ret = dio->result;
1091 finished_one_bio(dio); /* This can free the dio */
1092 blk_run_address_space(inode->i_mapping);
1093 if (should_wait) {
1094 unsigned long flags;
1095 /*
1096 * Wait for already issued I/O to drain out and
1097 * release its references to user-space pages
1098 * before returning to fallback on buffered I/O
1099 */
1100
1101 spin_lock_irqsave(&dio->bio_lock, flags);
1102 set_current_state(TASK_UNINTERRUPTIBLE);
1103 while (dio->bio_count) {
1104 spin_unlock_irqrestore(&dio->bio_lock, flags);
1105 io_schedule();
1106 spin_lock_irqsave(&dio->bio_lock, flags);
1107 set_current_state(TASK_UNINTERRUPTIBLE);
1108 }
1109 spin_unlock_irqrestore(&dio->bio_lock, flags);
1110 set_current_state(TASK_RUNNING);
1111 kfree(dio);
1112 }
1113 } else {
1114 ssize_t transferred = 0;
1115
1116 finished_one_bio(dio);
1117 ret2 = dio_await_completion(dio);
1118 if (ret == 0)
1119 ret = ret2;
1120 if (ret == 0)
1121 ret = dio->page_errors;
1122 if (dio->result) {
1123 loff_t i_size = i_size_read(inode);
1124
1125 transferred = dio->result;
1126 /*
1127 * Adjust the return value if the read crossed a
1128 * non-block-aligned EOF.
1129 */
1130 if (rw == READ && (offset + transferred > i_size))
1131 transferred = i_size - offset;
1132 }
1133 dio_complete(dio, offset, transferred);
1134 if (ret == 0)
1135 ret = transferred;
1136
1137 /* We could have also come here on an AIO file extend */
1138 if (!is_sync_kiocb(iocb) && rw == WRITE &&
1139 ret >= 0 && dio->result == dio->size)
1140 /*
1141 * For AIO writes where we have completed the
1142 * i/o, we have to mark the the aio complete.
1143 */
1144 aio_complete(iocb, ret, 0);
1145 kfree(dio);
1146 }
1147 return ret;
1148}
1149
1150/*
1151 * This is a library function for use by filesystem drivers.
1152 * The locking rules are governed by the dio_lock_type parameter.
1153 *
1154 * DIO_NO_LOCKING (no locking, for raw block device access)
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001155 * For writes, i_mutex is not held on entry; it is never taken.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001156 *
1157 * DIO_LOCKING (simple locking for regular files)
Nathan Scott3fb962b2006-03-15 15:14:45 +11001158 * For writes we are called under i_mutex and return with i_mutex held, even
1159 * though it is internally dropped.
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001160 * For reads, i_mutex is not held on entry, but it is taken and dropped before
Linus Torvalds1da177e2005-04-16 15:20:36 -07001161 * returning.
1162 *
1163 * DIO_OWN_LOCKING (filesystem provides synchronisation and handling of
1164 * uninitialised data, allowing parallel direct readers and writers)
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001165 * For writes we are called without i_mutex, return without it, never touch it.
Nathan Scott3fb962b2006-03-15 15:14:45 +11001166 * For reads we are called under i_mutex and return with i_mutex held, even
1167 * though it may be internally dropped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168 *
1169 * Additional i_alloc_sem locking requirements described inline below.
1170 */
1171ssize_t
1172__blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1173 struct block_device *bdev, const struct iovec *iov, loff_t offset,
1174 unsigned long nr_segs, get_blocks_t get_blocks, dio_iodone_t end_io,
1175 int dio_lock_type)
1176{
1177 int seg;
1178 size_t size;
1179 unsigned long addr;
1180 unsigned blkbits = inode->i_blkbits;
1181 unsigned bdev_blkbits = 0;
1182 unsigned blocksize_mask = (1 << blkbits) - 1;
1183 ssize_t retval = -EINVAL;
1184 loff_t end = offset;
1185 struct dio *dio;
Nathan Scott3fb962b2006-03-15 15:14:45 +11001186 int release_i_mutex = 0;
1187 int acquire_i_mutex = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001188
1189 if (rw & WRITE)
1190 current->flags |= PF_SYNCWRITE;
1191
1192 if (bdev)
1193 bdev_blkbits = blksize_bits(bdev_hardsect_size(bdev));
1194
1195 if (offset & blocksize_mask) {
1196 if (bdev)
1197 blkbits = bdev_blkbits;
1198 blocksize_mask = (1 << blkbits) - 1;
1199 if (offset & blocksize_mask)
1200 goto out;
1201 }
1202
1203 /* Check the memory alignment. Blocks cannot straddle pages */
1204 for (seg = 0; seg < nr_segs; seg++) {
1205 addr = (unsigned long)iov[seg].iov_base;
1206 size = iov[seg].iov_len;
1207 end += size;
1208 if ((addr & blocksize_mask) || (size & blocksize_mask)) {
1209 if (bdev)
1210 blkbits = bdev_blkbits;
1211 blocksize_mask = (1 << blkbits) - 1;
1212 if ((addr & blocksize_mask) || (size & blocksize_mask))
1213 goto out;
1214 }
1215 }
1216
1217 dio = kmalloc(sizeof(*dio), GFP_KERNEL);
1218 retval = -ENOMEM;
1219 if (!dio)
1220 goto out;
1221
1222 /*
1223 * For block device access DIO_NO_LOCKING is used,
1224 * neither readers nor writers do any locking at all
1225 * For regular files using DIO_LOCKING,
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001226 * readers need to grab i_mutex and i_alloc_sem
1227 * writers need to grab i_alloc_sem only (i_mutex is already held)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001228 * For regular files using DIO_OWN_LOCKING,
1229 * neither readers nor writers take any locks here
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230 */
1231 dio->lock_type = dio_lock_type;
1232 if (dio_lock_type != DIO_NO_LOCKING) {
1233 /* watch out for a 0 len io from a tricksy fs */
1234 if (rw == READ && end > offset) {
1235 struct address_space *mapping;
1236
1237 mapping = iocb->ki_filp->f_mapping;
1238 if (dio_lock_type != DIO_OWN_LOCKING) {
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001239 mutex_lock(&inode->i_mutex);
Nathan Scott3fb962b2006-03-15 15:14:45 +11001240 release_i_mutex = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001241 }
1242
1243 retval = filemap_write_and_wait_range(mapping, offset,
1244 end - 1);
1245 if (retval) {
1246 kfree(dio);
1247 goto out;
1248 }
1249
1250 if (dio_lock_type == DIO_OWN_LOCKING) {
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001251 mutex_unlock(&inode->i_mutex);
Nathan Scott3fb962b2006-03-15 15:14:45 +11001252 acquire_i_mutex = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001253 }
1254 }
1255
1256 if (dio_lock_type == DIO_LOCKING)
1257 down_read(&inode->i_alloc_sem);
1258 }
1259
1260 /*
1261 * For file extending writes updating i_size before data
1262 * writeouts complete can expose uninitialized blocks. So
1263 * even for AIO, we need to wait for i/o to complete before
1264 * returning in this case.
1265 */
1266 dio->is_async = !is_sync_kiocb(iocb) && !((rw == WRITE) &&
1267 (end > i_size_read(inode)));
1268
1269 retval = direct_io_worker(rw, iocb, inode, iov, offset,
1270 nr_segs, blkbits, get_blocks, end_io, dio);
1271
1272 if (rw == READ && dio_lock_type == DIO_LOCKING)
Nathan Scott3fb962b2006-03-15 15:14:45 +11001273 release_i_mutex = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001274
1275out:
Nathan Scott3fb962b2006-03-15 15:14:45 +11001276 if (release_i_mutex)
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001277 mutex_unlock(&inode->i_mutex);
Nathan Scott3fb962b2006-03-15 15:14:45 +11001278 else if (acquire_i_mutex)
1279 mutex_lock(&inode->i_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280 if (rw & WRITE)
1281 current->flags &= ~PF_SYNCWRITE;
1282 return retval;
1283}
1284EXPORT_SYMBOL(__blockdev_direct_IO);