blob: 1d55e7e6734247e8464899f308d94f8d0a398f4b [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
59 * (blockdev) this locking is not done, and if it is DIO_OWN_LOCKING i_sem 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).
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)) {
165 /*
166 * A memory fault, but the filesystem has some outstanding
167 * mapped blocks. We need to use those blocks up to avoid
168 * leaking stale data in the file.
169 */
170 if (dio->page_errors == 0)
171 dio->page_errors = ret;
172 dio->pages[0] = ZERO_PAGE(dio->curr_user_address);
173 dio->head = 0;
174 dio->tail = 1;
175 ret = 0;
176 goto out;
177 }
178
179 if (ret >= 0) {
180 dio->curr_user_address += ret * PAGE_SIZE;
181 dio->curr_page += ret;
182 dio->head = 0;
183 dio->tail = ret;
184 ret = 0;
185 }
186out:
187 return ret;
188}
189
190/*
191 * Get another userspace page. Returns an ERR_PTR on error. Pages are
192 * buffered inside the dio so that we can call get_user_pages() against a
193 * decent number of pages, less frequently. To provide nicer use of the
194 * L1 cache.
195 */
196static struct page *dio_get_page(struct dio *dio)
197{
198 if (dio_pages_present(dio) == 0) {
199 int ret;
200
201 ret = dio_refill_pages(dio);
202 if (ret)
203 return ERR_PTR(ret);
204 BUG_ON(dio_pages_present(dio) == 0);
205 }
206 return dio->pages[dio->head++];
207}
208
209/*
210 * Called when all DIO BIO I/O has been completed - let the filesystem
211 * know, if it registered an interest earlier via get_blocks. Pass the
212 * private field of the map buffer_head so that filesystems can use it
213 * to hold additional state between get_blocks calls and dio_complete.
214 */
215static void dio_complete(struct dio *dio, loff_t offset, ssize_t bytes)
216{
217 if (dio->end_io && dio->result)
218 dio->end_io(dio->inode, offset, bytes, dio->map_bh.b_private);
219 if (dio->lock_type == DIO_LOCKING)
220 up_read(&dio->inode->i_alloc_sem);
221}
222
223/*
224 * Called when a BIO has been processed. If the count goes to zero then IO is
225 * complete and we can signal this to the AIO layer.
226 */
227static void finished_one_bio(struct dio *dio)
228{
229 unsigned long flags;
230
231 spin_lock_irqsave(&dio->bio_lock, flags);
232 if (dio->bio_count == 1) {
233 if (dio->is_async) {
Daniel McNeil29504ff2005-04-16 15:25:50 -0700234 ssize_t transferred;
235 loff_t offset;
236
Linus Torvalds1da177e2005-04-16 15:20:36 -0700237 /*
238 * Last reference to the dio is going away.
239 * Drop spinlock and complete the DIO.
240 */
241 spin_unlock_irqrestore(&dio->bio_lock, flags);
Daniel McNeil29504ff2005-04-16 15:25:50 -0700242
243 /* Check for short read case */
244 transferred = dio->result;
245 offset = dio->iocb->ki_pos;
246
247 if ((dio->rw == READ) &&
248 ((offset + transferred) > dio->i_size))
249 transferred = dio->i_size - offset;
250
251 dio_complete(dio, offset, transferred);
252
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253 /* Complete AIO later if falling back to buffered i/o */
254 if (dio->result == dio->size ||
255 ((dio->rw == READ) && dio->result)) {
Daniel McNeil29504ff2005-04-16 15:25:50 -0700256 aio_complete(dio->iocb, transferred, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257 kfree(dio);
258 return;
259 } else {
260 /*
261 * Falling back to buffered
262 */
263 spin_lock_irqsave(&dio->bio_lock, flags);
264 dio->bio_count--;
265 if (dio->waiter)
266 wake_up_process(dio->waiter);
267 spin_unlock_irqrestore(&dio->bio_lock, flags);
268 return;
269 }
270 }
271 }
272 dio->bio_count--;
273 spin_unlock_irqrestore(&dio->bio_lock, flags);
274}
275
276static int dio_bio_complete(struct dio *dio, struct bio *bio);
277/*
278 * Asynchronous IO callback.
279 */
280static int dio_bio_end_aio(struct bio *bio, unsigned int bytes_done, int error)
281{
282 struct dio *dio = bio->bi_private;
283
284 if (bio->bi_size)
285 return 1;
286
287 /* cleanup the bio */
288 dio_bio_complete(dio, bio);
289 return 0;
290}
291
292/*
293 * The BIO completion handler simply queues the BIO up for the process-context
294 * handler.
295 *
296 * During I/O bi_private points at the dio. After I/O, bi_private is used to
297 * implement a singly-linked list of completed BIOs, at dio->bio_list.
298 */
299static int dio_bio_end_io(struct bio *bio, unsigned int bytes_done, int error)
300{
301 struct dio *dio = bio->bi_private;
302 unsigned long flags;
303
304 if (bio->bi_size)
305 return 1;
306
307 spin_lock_irqsave(&dio->bio_lock, flags);
308 bio->bi_private = dio->bio_list;
309 dio->bio_list = bio;
310 dio->bios_in_flight--;
311 if (dio->waiter && dio->bios_in_flight == 0)
312 wake_up_process(dio->waiter);
313 spin_unlock_irqrestore(&dio->bio_lock, flags);
314 return 0;
315}
316
317static int
318dio_bio_alloc(struct dio *dio, struct block_device *bdev,
319 sector_t first_sector, int nr_vecs)
320{
321 struct bio *bio;
322
323 bio = bio_alloc(GFP_KERNEL, nr_vecs);
324 if (bio == NULL)
325 return -ENOMEM;
326
327 bio->bi_bdev = bdev;
328 bio->bi_sector = first_sector;
329 if (dio->is_async)
330 bio->bi_end_io = dio_bio_end_aio;
331 else
332 bio->bi_end_io = dio_bio_end_io;
333
334 dio->bio = bio;
335 return 0;
336}
337
338/*
339 * In the AIO read case we speculatively dirty the pages before starting IO.
340 * During IO completion, any of these pages which happen to have been written
341 * back will be redirtied by bio_check_pages_dirty().
342 */
343static void dio_bio_submit(struct dio *dio)
344{
345 struct bio *bio = dio->bio;
346 unsigned long flags;
347
348 bio->bi_private = dio;
349 spin_lock_irqsave(&dio->bio_lock, flags);
350 dio->bio_count++;
351 dio->bios_in_flight++;
352 spin_unlock_irqrestore(&dio->bio_lock, flags);
353 if (dio->is_async && dio->rw == READ)
354 bio_set_pages_dirty(bio);
355 submit_bio(dio->rw, bio);
356
357 dio->bio = NULL;
358 dio->boundary = 0;
359}
360
361/*
362 * Release any resources in case of a failure
363 */
364static void dio_cleanup(struct dio *dio)
365{
366 while (dio_pages_present(dio))
367 page_cache_release(dio_get_page(dio));
368}
369
370/*
371 * Wait for the next BIO to complete. Remove it and return it.
372 */
373static struct bio *dio_await_one(struct dio *dio)
374{
375 unsigned long flags;
376 struct bio *bio;
377
378 spin_lock_irqsave(&dio->bio_lock, flags);
379 while (dio->bio_list == NULL) {
380 set_current_state(TASK_UNINTERRUPTIBLE);
381 if (dio->bio_list == NULL) {
382 dio->waiter = current;
383 spin_unlock_irqrestore(&dio->bio_lock, flags);
384 blk_run_address_space(dio->inode->i_mapping);
385 io_schedule();
386 spin_lock_irqsave(&dio->bio_lock, flags);
387 dio->waiter = NULL;
388 }
389 set_current_state(TASK_RUNNING);
390 }
391 bio = dio->bio_list;
392 dio->bio_list = bio->bi_private;
393 spin_unlock_irqrestore(&dio->bio_lock, flags);
394 return bio;
395}
396
397/*
398 * Process one completed BIO. No locks are held.
399 */
400static int dio_bio_complete(struct dio *dio, struct bio *bio)
401{
402 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
403 struct bio_vec *bvec = bio->bi_io_vec;
404 int page_no;
405
406 if (!uptodate)
407 dio->result = -EIO;
408
409 if (dio->is_async && dio->rw == READ) {
410 bio_check_pages_dirty(bio); /* transfers ownership */
411 } else {
412 for (page_no = 0; page_no < bio->bi_vcnt; page_no++) {
413 struct page *page = bvec[page_no].bv_page;
414
415 if (dio->rw == READ && !PageCompound(page))
416 set_page_dirty_lock(page);
417 page_cache_release(page);
418 }
419 bio_put(bio);
420 }
421 finished_one_bio(dio);
422 return uptodate ? 0 : -EIO;
423}
424
425/*
426 * Wait on and process all in-flight BIOs.
427 */
428static int dio_await_completion(struct dio *dio)
429{
430 int ret = 0;
431
432 if (dio->bio)
433 dio_bio_submit(dio);
434
435 /*
436 * The bio_lock is not held for the read of bio_count.
437 * This is ok since it is the dio_bio_complete() that changes
438 * bio_count.
439 */
440 while (dio->bio_count) {
441 struct bio *bio = dio_await_one(dio);
442 int ret2;
443
444 ret2 = dio_bio_complete(dio, bio);
445 if (ret == 0)
446 ret = ret2;
447 }
448 return ret;
449}
450
451/*
452 * A really large O_DIRECT read or write can generate a lot of BIOs. So
453 * to keep the memory consumption sane we periodically reap any completed BIOs
454 * during the BIO generation phase.
455 *
456 * This also helps to limit the peak amount of pinned userspace memory.
457 */
458static int dio_bio_reap(struct dio *dio)
459{
460 int ret = 0;
461
462 if (dio->reap_counter++ >= 64) {
463 while (dio->bio_list) {
464 unsigned long flags;
465 struct bio *bio;
466 int ret2;
467
468 spin_lock_irqsave(&dio->bio_lock, flags);
469 bio = dio->bio_list;
470 dio->bio_list = bio->bi_private;
471 spin_unlock_irqrestore(&dio->bio_lock, flags);
472 ret2 = dio_bio_complete(dio, bio);
473 if (ret == 0)
474 ret = ret2;
475 }
476 dio->reap_counter = 0;
477 }
478 return ret;
479}
480
481/*
482 * Call into the fs to map some more disk blocks. We record the current number
483 * of available blocks at dio->blocks_available. These are in units of the
484 * fs blocksize, (1 << inode->i_blkbits).
485 *
486 * The fs is allowed to map lots of blocks at once. If it wants to do that,
487 * it uses the passed inode-relative block number as the file offset, as usual.
488 *
489 * get_blocks() is passed the number of i_blkbits-sized blocks which direct_io
490 * has remaining to do. The fs should not map more than this number of blocks.
491 *
492 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
493 * indicate how much contiguous disk space has been made available at
494 * bh->b_blocknr.
495 *
496 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
497 * This isn't very efficient...
498 *
499 * In the case of filesystem holes: the fs may return an arbitrarily-large
500 * hole by returning an appropriate value in b_size and by clearing
501 * buffer_mapped(). However the direct-io code will only process holes one
502 * block at a time - it will repeatedly call get_blocks() as it walks the hole.
503 */
504static int get_more_blocks(struct dio *dio)
505{
506 int ret;
507 struct buffer_head *map_bh = &dio->map_bh;
508 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
509 unsigned long fs_count; /* Number of filesystem-sized blocks */
510 unsigned long dio_count;/* Number of dio_block-sized blocks */
511 unsigned long blkmask;
512 int create;
513
514 /*
515 * If there was a memory error and we've overwritten all the
516 * mapped blocks then we can now return that memory error
517 */
518 ret = dio->page_errors;
519 if (ret == 0) {
520 map_bh->b_state = 0;
521 map_bh->b_size = 0;
522 BUG_ON(dio->block_in_file >= dio->final_block_in_request);
523 fs_startblk = dio->block_in_file >> dio->blkfactor;
524 dio_count = dio->final_block_in_request - dio->block_in_file;
525 fs_count = dio_count >> dio->blkfactor;
526 blkmask = (1 << dio->blkfactor) - 1;
527 if (dio_count & blkmask)
528 fs_count++;
529
530 create = dio->rw == WRITE;
531 if (dio->lock_type == DIO_LOCKING) {
532 if (dio->block_in_file < (i_size_read(dio->inode) >>
533 dio->blkbits))
534 create = 0;
535 } else if (dio->lock_type == DIO_NO_LOCKING) {
536 create = 0;
537 }
538 /*
539 * For writes inside i_size we forbid block creations: only
540 * overwrites are permitted. We fall back to buffered writes
541 * at a higher level for inside-i_size block-instantiating
542 * writes.
543 */
544 ret = (*dio->get_blocks)(dio->inode, fs_startblk, fs_count,
545 map_bh, create);
546 }
547 return ret;
548}
549
550/*
551 * There is no bio. Make one now.
552 */
553static int dio_new_bio(struct dio *dio, sector_t start_sector)
554{
555 sector_t sector;
556 int ret, nr_pages;
557
558 ret = dio_bio_reap(dio);
559 if (ret)
560 goto out;
561 sector = start_sector << (dio->blkbits - 9);
562 nr_pages = min(dio->pages_in_io, bio_get_nr_vecs(dio->map_bh.b_bdev));
563 BUG_ON(nr_pages <= 0);
564 ret = dio_bio_alloc(dio, dio->map_bh.b_bdev, sector, nr_pages);
565 dio->boundary = 0;
566out:
567 return ret;
568}
569
570/*
571 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
572 * that was successful then update final_block_in_bio and take a ref against
573 * the just-added page.
574 *
575 * Return zero on success. Non-zero means the caller needs to start a new BIO.
576 */
577static int dio_bio_add_page(struct dio *dio)
578{
579 int ret;
580
581 ret = bio_add_page(dio->bio, dio->cur_page,
582 dio->cur_page_len, dio->cur_page_offset);
583 if (ret == dio->cur_page_len) {
584 /*
585 * Decrement count only, if we are done with this page
586 */
587 if ((dio->cur_page_len + dio->cur_page_offset) == PAGE_SIZE)
588 dio->pages_in_io--;
589 page_cache_get(dio->cur_page);
590 dio->final_block_in_bio = dio->cur_page_block +
591 (dio->cur_page_len >> dio->blkbits);
592 ret = 0;
593 } else {
594 ret = 1;
595 }
596 return ret;
597}
598
599/*
600 * Put cur_page under IO. The section of cur_page which is described by
601 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
602 * starts on-disk at cur_page_block.
603 *
604 * We take a ref against the page here (on behalf of its presence in the bio).
605 *
606 * The caller of this function is responsible for removing cur_page from the
607 * dio, and for dropping the refcount which came from that presence.
608 */
609static int dio_send_cur_page(struct dio *dio)
610{
611 int ret = 0;
612
613 if (dio->bio) {
614 /*
615 * See whether this new request is contiguous with the old
616 */
617 if (dio->final_block_in_bio != dio->cur_page_block)
618 dio_bio_submit(dio);
619 /*
620 * Submit now if the underlying fs is about to perform a
621 * metadata read
622 */
623 if (dio->boundary)
624 dio_bio_submit(dio);
625 }
626
627 if (dio->bio == NULL) {
628 ret = dio_new_bio(dio, dio->cur_page_block);
629 if (ret)
630 goto out;
631 }
632
633 if (dio_bio_add_page(dio) != 0) {
634 dio_bio_submit(dio);
635 ret = dio_new_bio(dio, dio->cur_page_block);
636 if (ret == 0) {
637 ret = dio_bio_add_page(dio);
638 BUG_ON(ret != 0);
639 }
640 }
641out:
642 return ret;
643}
644
645/*
646 * An autonomous function to put a chunk of a page under deferred IO.
647 *
648 * The caller doesn't actually know (or care) whether this piece of page is in
649 * a BIO, or is under IO or whatever. We just take care of all possible
650 * situations here. The separation between the logic of do_direct_IO() and
651 * that of submit_page_section() is important for clarity. Please don't break.
652 *
653 * The chunk of page starts on-disk at blocknr.
654 *
655 * We perform deferred IO, by recording the last-submitted page inside our
656 * private part of the dio structure. If possible, we just expand the IO
657 * across that page here.
658 *
659 * If that doesn't work out then we put the old page into the bio and add this
660 * page to the dio instead.
661 */
662static int
663submit_page_section(struct dio *dio, struct page *page,
664 unsigned offset, unsigned len, sector_t blocknr)
665{
666 int ret = 0;
667
668 /*
669 * Can we just grow the current page's presence in the dio?
670 */
671 if ( (dio->cur_page == page) &&
672 (dio->cur_page_offset + dio->cur_page_len == offset) &&
673 (dio->cur_page_block +
674 (dio->cur_page_len >> dio->blkbits) == blocknr)) {
675 dio->cur_page_len += len;
676
677 /*
678 * If dio->boundary then we want to schedule the IO now to
679 * avoid metadata seeks.
680 */
681 if (dio->boundary) {
682 ret = dio_send_cur_page(dio);
683 page_cache_release(dio->cur_page);
684 dio->cur_page = NULL;
685 }
686 goto out;
687 }
688
689 /*
690 * If there's a deferred page already there then send it.
691 */
692 if (dio->cur_page) {
693 ret = dio_send_cur_page(dio);
694 page_cache_release(dio->cur_page);
695 dio->cur_page = NULL;
696 if (ret)
697 goto out;
698 }
699
700 page_cache_get(page); /* It is in dio */
701 dio->cur_page = page;
702 dio->cur_page_offset = offset;
703 dio->cur_page_len = len;
704 dio->cur_page_block = blocknr;
705out:
706 return ret;
707}
708
709/*
710 * Clean any dirty buffers in the blockdev mapping which alias newly-created
711 * file blocks. Only called for S_ISREG files - blockdevs do not set
712 * buffer_new
713 */
714static void clean_blockdev_aliases(struct dio *dio)
715{
716 unsigned i;
717 unsigned nblocks;
718
719 nblocks = dio->map_bh.b_size >> dio->inode->i_blkbits;
720
721 for (i = 0; i < nblocks; i++) {
722 unmap_underlying_metadata(dio->map_bh.b_bdev,
723 dio->map_bh.b_blocknr + i);
724 }
725}
726
727/*
728 * If we are not writing the entire block and get_block() allocated
729 * the block for us, we need to fill-in the unused portion of the
730 * block with zeros. This happens only if user-buffer, fileoffset or
731 * io length is not filesystem block-size multiple.
732 *
733 * `end' is zero if we're doing the start of the IO, 1 at the end of the
734 * IO.
735 */
736static void dio_zero_block(struct dio *dio, int end)
737{
738 unsigned dio_blocks_per_fs_block;
739 unsigned this_chunk_blocks; /* In dio_blocks */
740 unsigned this_chunk_bytes;
741 struct page *page;
742
743 dio->start_zero_done = 1;
744 if (!dio->blkfactor || !buffer_new(&dio->map_bh))
745 return;
746
747 dio_blocks_per_fs_block = 1 << dio->blkfactor;
748 this_chunk_blocks = dio->block_in_file & (dio_blocks_per_fs_block - 1);
749
750 if (!this_chunk_blocks)
751 return;
752
753 /*
754 * We need to zero out part of an fs block. It is either at the
755 * beginning or the end of the fs block.
756 */
757 if (end)
758 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
759
760 this_chunk_bytes = this_chunk_blocks << dio->blkbits;
761
762 page = ZERO_PAGE(dio->curr_user_address);
763 if (submit_page_section(dio, page, 0, this_chunk_bytes,
764 dio->next_block_for_io))
765 return;
766
767 dio->next_block_for_io += this_chunk_blocks;
768}
769
770/*
771 * Walk the user pages, and the file, mapping blocks to disk and generating
772 * a sequence of (page,offset,len,block) mappings. These mappings are injected
773 * into submit_page_section(), which takes care of the next stage of submission
774 *
775 * Direct IO against a blockdev is different from a file. Because we can
776 * happily perform page-sized but 512-byte aligned IOs. It is important that
777 * blockdev IO be able to have fine alignment and large sizes.
778 *
779 * So what we do is to permit the ->get_blocks function to populate bh.b_size
780 * with the size of IO which is permitted at this offset and this i_blkbits.
781 *
782 * For best results, the blockdev should be set up with 512-byte i_blkbits and
783 * it should set b_size to PAGE_SIZE or more inside get_blocks(). This gives
784 * fine alignment but still allows this function to work in PAGE_SIZE units.
785 */
786static int do_direct_IO(struct dio *dio)
787{
788 const unsigned blkbits = dio->blkbits;
789 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
790 struct page *page;
791 unsigned block_in_page;
792 struct buffer_head *map_bh = &dio->map_bh;
793 int ret = 0;
794
795 /* The I/O can start at any block offset within the first page */
796 block_in_page = dio->first_block_in_page;
797
798 while (dio->block_in_file < dio->final_block_in_request) {
799 page = dio_get_page(dio);
800 if (IS_ERR(page)) {
801 ret = PTR_ERR(page);
802 goto out;
803 }
804
805 while (block_in_page < blocks_per_page) {
806 unsigned offset_in_page = block_in_page << blkbits;
807 unsigned this_chunk_bytes; /* # of bytes mapped */
808 unsigned this_chunk_blocks; /* # of blocks */
809 unsigned u;
810
811 if (dio->blocks_available == 0) {
812 /*
813 * Need to go and map some more disk
814 */
815 unsigned long blkmask;
816 unsigned long dio_remainder;
817
818 ret = get_more_blocks(dio);
819 if (ret) {
820 page_cache_release(page);
821 goto out;
822 }
823 if (!buffer_mapped(map_bh))
824 goto do_holes;
825
826 dio->blocks_available =
827 map_bh->b_size >> dio->blkbits;
828 dio->next_block_for_io =
829 map_bh->b_blocknr << dio->blkfactor;
830 if (buffer_new(map_bh))
831 clean_blockdev_aliases(dio);
832
833 if (!dio->blkfactor)
834 goto do_holes;
835
836 blkmask = (1 << dio->blkfactor) - 1;
837 dio_remainder = (dio->block_in_file & blkmask);
838
839 /*
840 * If we are at the start of IO and that IO
841 * starts partway into a fs-block,
842 * dio_remainder will be non-zero. If the IO
843 * is a read then we can simply advance the IO
844 * cursor to the first block which is to be
845 * read. But if the IO is a write and the
846 * block was newly allocated we cannot do that;
847 * the start of the fs block must be zeroed out
848 * on-disk
849 */
850 if (!buffer_new(map_bh))
851 dio->next_block_for_io += dio_remainder;
852 dio->blocks_available -= dio_remainder;
853 }
854do_holes:
855 /* Handle holes */
856 if (!buffer_mapped(map_bh)) {
857 char *kaddr;
858
859 /* AKPM: eargh, -ENOTBLK is a hack */
860 if (dio->rw == WRITE) {
861 page_cache_release(page);
862 return -ENOTBLK;
863 }
864
865 if (dio->block_in_file >=
866 i_size_read(dio->inode)>>blkbits) {
867 /* We hit eof */
868 page_cache_release(page);
869 goto out;
870 }
871 kaddr = kmap_atomic(page, KM_USER0);
872 memset(kaddr + (block_in_page << blkbits),
873 0, 1 << blkbits);
874 flush_dcache_page(page);
875 kunmap_atomic(kaddr, KM_USER0);
876 dio->block_in_file++;
877 block_in_page++;
878 goto next_block;
879 }
880
881 /*
882 * If we're performing IO which has an alignment which
883 * is finer than the underlying fs, go check to see if
884 * we must zero out the start of this block.
885 */
886 if (unlikely(dio->blkfactor && !dio->start_zero_done))
887 dio_zero_block(dio, 0);
888
889 /*
890 * Work out, in this_chunk_blocks, how much disk we
891 * can add to this page
892 */
893 this_chunk_blocks = dio->blocks_available;
894 u = (PAGE_SIZE - offset_in_page) >> blkbits;
895 if (this_chunk_blocks > u)
896 this_chunk_blocks = u;
897 u = dio->final_block_in_request - dio->block_in_file;
898 if (this_chunk_blocks > u)
899 this_chunk_blocks = u;
900 this_chunk_bytes = this_chunk_blocks << blkbits;
901 BUG_ON(this_chunk_bytes == 0);
902
903 dio->boundary = buffer_boundary(map_bh);
904 ret = submit_page_section(dio, page, offset_in_page,
905 this_chunk_bytes, dio->next_block_for_io);
906 if (ret) {
907 page_cache_release(page);
908 goto out;
909 }
910 dio->next_block_for_io += this_chunk_blocks;
911
912 dio->block_in_file += this_chunk_blocks;
913 block_in_page += this_chunk_blocks;
914 dio->blocks_available -= this_chunk_blocks;
915next_block:
916 if (dio->block_in_file > dio->final_block_in_request)
917 BUG();
918 if (dio->block_in_file == dio->final_block_in_request)
919 break;
920 }
921
922 /* Drop the ref which was taken in get_user_pages() */
923 page_cache_release(page);
924 block_in_page = 0;
925 }
926out:
927 return ret;
928}
929
930/*
931 * Releases both i_sem and i_alloc_sem
932 */
933static ssize_t
934direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
935 const struct iovec *iov, loff_t offset, unsigned long nr_segs,
936 unsigned blkbits, get_blocks_t get_blocks, dio_iodone_t end_io,
937 struct dio *dio)
938{
939 unsigned long user_addr;
940 int seg;
941 ssize_t ret = 0;
942 ssize_t ret2;
943 size_t bytes;
944
945 dio->bio = NULL;
946 dio->inode = inode;
947 dio->rw = rw;
948 dio->blkbits = blkbits;
949 dio->blkfactor = inode->i_blkbits - blkbits;
950 dio->start_zero_done = 0;
951 dio->size = 0;
952 dio->block_in_file = offset >> blkbits;
953 dio->blocks_available = 0;
954 dio->cur_page = NULL;
955
956 dio->boundary = 0;
957 dio->reap_counter = 0;
958 dio->get_blocks = get_blocks;
959 dio->end_io = end_io;
960 dio->map_bh.b_private = NULL;
961 dio->final_block_in_bio = -1;
962 dio->next_block_for_io = -1;
963
964 dio->page_errors = 0;
965 dio->result = 0;
966 dio->iocb = iocb;
Daniel McNeil29504ff2005-04-16 15:25:50 -0700967 dio->i_size = i_size_read(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700968
969 /*
970 * BIO completion state.
971 *
972 * ->bio_count starts out at one, and we decrement it to zero after all
973 * BIOs are submitted. This to avoid the situation where a really fast
974 * (or synchronous) device could take the count to zero while we're
975 * still submitting BIOs.
976 */
977 dio->bio_count = 1;
978 dio->bios_in_flight = 0;
979 spin_lock_init(&dio->bio_lock);
980 dio->bio_list = NULL;
981 dio->waiter = NULL;
982
983 /*
984 * In case of non-aligned buffers, we may need 2 more
985 * pages since we need to zero out first and last block.
986 */
987 if (unlikely(dio->blkfactor))
988 dio->pages_in_io = 2;
989 else
990 dio->pages_in_io = 0;
991
992 for (seg = 0; seg < nr_segs; seg++) {
993 user_addr = (unsigned long)iov[seg].iov_base;
994 dio->pages_in_io +=
995 ((user_addr+iov[seg].iov_len +PAGE_SIZE-1)/PAGE_SIZE
996 - user_addr/PAGE_SIZE);
997 }
998
999 for (seg = 0; seg < nr_segs; seg++) {
1000 user_addr = (unsigned long)iov[seg].iov_base;
1001 dio->size += bytes = iov[seg].iov_len;
1002
1003 /* Index into the first page of the first block */
1004 dio->first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
1005 dio->final_block_in_request = dio->block_in_file +
1006 (bytes >> blkbits);
1007 /* Page fetching state */
1008 dio->head = 0;
1009 dio->tail = 0;
1010 dio->curr_page = 0;
1011
1012 dio->total_pages = 0;
1013 if (user_addr & (PAGE_SIZE-1)) {
1014 dio->total_pages++;
1015 bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
1016 }
1017 dio->total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
1018 dio->curr_user_address = user_addr;
1019
1020 ret = do_direct_IO(dio);
1021
1022 dio->result += iov[seg].iov_len -
1023 ((dio->final_block_in_request - dio->block_in_file) <<
1024 blkbits);
1025
1026 if (ret) {
1027 dio_cleanup(dio);
1028 break;
1029 }
1030 } /* end iovec loop */
1031
1032 if (ret == -ENOTBLK && rw == WRITE) {
1033 /*
1034 * The remaining part of the request will be
1035 * be handled by buffered I/O when we return
1036 */
1037 ret = 0;
1038 }
1039 /*
1040 * There may be some unwritten disk at the end of a part-written
1041 * fs-block-sized block. Go zero that now.
1042 */
1043 dio_zero_block(dio, 1);
1044
1045 if (dio->cur_page) {
1046 ret2 = dio_send_cur_page(dio);
1047 if (ret == 0)
1048 ret = ret2;
1049 page_cache_release(dio->cur_page);
1050 dio->cur_page = NULL;
1051 }
1052 if (dio->bio)
1053 dio_bio_submit(dio);
1054
1055 /*
1056 * It is possible that, we return short IO due to end of file.
1057 * In that case, we need to release all the pages we got hold on.
1058 */
1059 dio_cleanup(dio);
1060
1061 /*
1062 * All block lookups have been performed. For READ requests
1063 * we can let i_sem go now that its achieved its purpose
1064 * of protecting us from looking up uninitialized blocks.
1065 */
1066 if ((rw == READ) && (dio->lock_type == DIO_LOCKING))
1067 up(&dio->inode->i_sem);
1068
1069 /*
1070 * OK, all BIOs are submitted, so we can decrement bio_count to truly
1071 * reflect the number of to-be-processed BIOs.
1072 */
1073 if (dio->is_async) {
1074 int should_wait = 0;
1075
1076 if (dio->result < dio->size && rw == WRITE) {
1077 dio->waiter = current;
1078 should_wait = 1;
1079 }
1080 if (ret == 0)
1081 ret = dio->result;
1082 finished_one_bio(dio); /* This can free the dio */
1083 blk_run_address_space(inode->i_mapping);
1084 if (should_wait) {
1085 unsigned long flags;
1086 /*
1087 * Wait for already issued I/O to drain out and
1088 * release its references to user-space pages
1089 * before returning to fallback on buffered I/O
1090 */
1091
1092 spin_lock_irqsave(&dio->bio_lock, flags);
1093 set_current_state(TASK_UNINTERRUPTIBLE);
1094 while (dio->bio_count) {
1095 spin_unlock_irqrestore(&dio->bio_lock, flags);
1096 io_schedule();
1097 spin_lock_irqsave(&dio->bio_lock, flags);
1098 set_current_state(TASK_UNINTERRUPTIBLE);
1099 }
1100 spin_unlock_irqrestore(&dio->bio_lock, flags);
1101 set_current_state(TASK_RUNNING);
1102 kfree(dio);
1103 }
1104 } else {
1105 ssize_t transferred = 0;
1106
1107 finished_one_bio(dio);
1108 ret2 = dio_await_completion(dio);
1109 if (ret == 0)
1110 ret = ret2;
1111 if (ret == 0)
1112 ret = dio->page_errors;
1113 if (dio->result) {
1114 loff_t i_size = i_size_read(inode);
1115
1116 transferred = dio->result;
1117 /*
1118 * Adjust the return value if the read crossed a
1119 * non-block-aligned EOF.
1120 */
1121 if (rw == READ && (offset + transferred > i_size))
1122 transferred = i_size - offset;
1123 }
1124 dio_complete(dio, offset, transferred);
1125 if (ret == 0)
1126 ret = transferred;
1127
1128 /* We could have also come here on an AIO file extend */
1129 if (!is_sync_kiocb(iocb) && rw == WRITE &&
1130 ret >= 0 && dio->result == dio->size)
1131 /*
1132 * For AIO writes where we have completed the
1133 * i/o, we have to mark the the aio complete.
1134 */
1135 aio_complete(iocb, ret, 0);
1136 kfree(dio);
1137 }
1138 return ret;
1139}
1140
1141/*
1142 * This is a library function for use by filesystem drivers.
1143 * The locking rules are governed by the dio_lock_type parameter.
1144 *
1145 * DIO_NO_LOCKING (no locking, for raw block device access)
1146 * For writes, i_sem is not held on entry; it is never taken.
1147 *
1148 * DIO_LOCKING (simple locking for regular files)
1149 * For writes we are called under i_sem and return with i_sem held, even though
1150 * it is internally dropped.
1151 * For reads, i_sem is not held on entry, but it is taken and dropped before
1152 * returning.
1153 *
1154 * DIO_OWN_LOCKING (filesystem provides synchronisation and handling of
1155 * uninitialised data, allowing parallel direct readers and writers)
1156 * For writes we are called without i_sem, return without it, never touch it.
1157 * For reads, i_sem is held on entry and will be released before returning.
1158 *
1159 * Additional i_alloc_sem locking requirements described inline below.
1160 */
1161ssize_t
1162__blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1163 struct block_device *bdev, const struct iovec *iov, loff_t offset,
1164 unsigned long nr_segs, get_blocks_t get_blocks, dio_iodone_t end_io,
1165 int dio_lock_type)
1166{
1167 int seg;
1168 size_t size;
1169 unsigned long addr;
1170 unsigned blkbits = inode->i_blkbits;
1171 unsigned bdev_blkbits = 0;
1172 unsigned blocksize_mask = (1 << blkbits) - 1;
1173 ssize_t retval = -EINVAL;
1174 loff_t end = offset;
1175 struct dio *dio;
1176 int reader_with_isem = (rw == READ && dio_lock_type == DIO_OWN_LOCKING);
1177
1178 if (rw & WRITE)
1179 current->flags |= PF_SYNCWRITE;
1180
1181 if (bdev)
1182 bdev_blkbits = blksize_bits(bdev_hardsect_size(bdev));
1183
1184 if (offset & blocksize_mask) {
1185 if (bdev)
1186 blkbits = bdev_blkbits;
1187 blocksize_mask = (1 << blkbits) - 1;
1188 if (offset & blocksize_mask)
1189 goto out;
1190 }
1191
1192 /* Check the memory alignment. Blocks cannot straddle pages */
1193 for (seg = 0; seg < nr_segs; seg++) {
1194 addr = (unsigned long)iov[seg].iov_base;
1195 size = iov[seg].iov_len;
1196 end += size;
1197 if ((addr & blocksize_mask) || (size & blocksize_mask)) {
1198 if (bdev)
1199 blkbits = bdev_blkbits;
1200 blocksize_mask = (1 << blkbits) - 1;
1201 if ((addr & blocksize_mask) || (size & blocksize_mask))
1202 goto out;
1203 }
1204 }
1205
1206 dio = kmalloc(sizeof(*dio), GFP_KERNEL);
1207 retval = -ENOMEM;
1208 if (!dio)
1209 goto out;
1210
1211 /*
1212 * For block device access DIO_NO_LOCKING is used,
1213 * neither readers nor writers do any locking at all
1214 * For regular files using DIO_LOCKING,
1215 * readers need to grab i_sem and i_alloc_sem
1216 * writers need to grab i_alloc_sem only (i_sem is already held)
1217 * For regular files using DIO_OWN_LOCKING,
1218 * neither readers nor writers take any locks here
1219 * (i_sem is already held and release for writers here)
1220 */
1221 dio->lock_type = dio_lock_type;
1222 if (dio_lock_type != DIO_NO_LOCKING) {
1223 /* watch out for a 0 len io from a tricksy fs */
1224 if (rw == READ && end > offset) {
1225 struct address_space *mapping;
1226
1227 mapping = iocb->ki_filp->f_mapping;
1228 if (dio_lock_type != DIO_OWN_LOCKING) {
1229 down(&inode->i_sem);
1230 reader_with_isem = 1;
1231 }
1232
1233 retval = filemap_write_and_wait_range(mapping, offset,
1234 end - 1);
1235 if (retval) {
1236 kfree(dio);
1237 goto out;
1238 }
1239
1240 if (dio_lock_type == DIO_OWN_LOCKING) {
1241 up(&inode->i_sem);
1242 reader_with_isem = 0;
1243 }
1244 }
1245
1246 if (dio_lock_type == DIO_LOCKING)
1247 down_read(&inode->i_alloc_sem);
1248 }
1249
1250 /*
1251 * For file extending writes updating i_size before data
1252 * writeouts complete can expose uninitialized blocks. So
1253 * even for AIO, we need to wait for i/o to complete before
1254 * returning in this case.
1255 */
1256 dio->is_async = !is_sync_kiocb(iocb) && !((rw == WRITE) &&
1257 (end > i_size_read(inode)));
1258
1259 retval = direct_io_worker(rw, iocb, inode, iov, offset,
1260 nr_segs, blkbits, get_blocks, end_io, dio);
1261
1262 if (rw == READ && dio_lock_type == DIO_LOCKING)
1263 reader_with_isem = 0;
1264
1265out:
1266 if (reader_with_isem)
1267 up(&inode->i_sem);
1268 if (rw & WRITE)
1269 current->flags &= ~PF_SYNCWRITE;
1270 return retval;
1271}
1272EXPORT_SYMBOL(__blockdev_direct_IO);