Loading...
1/*
2 * fs/direct-io.c
3 *
4 * Copyright (C) 2002, Linus Torvalds.
5 *
6 * O_DIRECT
7 *
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.
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/task_io_accounting_ops.h>
31#include <linux/bio.h>
32#include <linux/wait.h>
33#include <linux/err.h>
34#include <linux/blkdev.h>
35#include <linux/buffer_head.h>
36#include <linux/rwsem.h>
37#include <linux/uio.h>
38#include <linux/atomic.h>
39#include <linux/prefetch.h>
40
41/*
42 * How many user pages to map in one call to get_user_pages(). This determines
43 * the size of a structure in the slab cache
44 */
45#define DIO_PAGES 64
46
47/*
48 * This code generally works in units of "dio_blocks". A dio_block is
49 * somewhere between the hard sector size and the filesystem block size. it
50 * is determined on a per-invocation basis. When talking to the filesystem
51 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
52 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
53 * to bio_block quantities by shifting left by blkfactor.
54 *
55 * If blkfactor is zero then the user's request was aligned to the filesystem's
56 * blocksize.
57 */
58
59/* dio_state only used in the submission path */
60
61struct dio_submit {
62 struct bio *bio; /* bio under assembly */
63 unsigned blkbits; /* doesn't change */
64 unsigned blkfactor; /* When we're using an alignment which
65 is finer than the filesystem's soft
66 blocksize, this specifies how much
67 finer. blkfactor=2 means 1/4-block
68 alignment. Does not change */
69 unsigned start_zero_done; /* flag: sub-blocksize zeroing has
70 been performed at the start of a
71 write */
72 int pages_in_io; /* approximate total IO pages */
73 size_t size; /* total request size (doesn't change)*/
74 sector_t block_in_file; /* Current offset into the underlying
75 file in dio_block units. */
76 unsigned blocks_available; /* At block_in_file. changes */
77 int reap_counter; /* rate limit reaping */
78 sector_t final_block_in_request;/* doesn't change */
79 unsigned first_block_in_page; /* doesn't change, Used only once */
80 int boundary; /* prev block is at a boundary */
81 get_block_t *get_block; /* block mapping function */
82 dio_submit_t *submit_io; /* IO submition function */
83
84 loff_t logical_offset_in_bio; /* current first logical block in bio */
85 sector_t final_block_in_bio; /* current final block in bio + 1 */
86 sector_t next_block_for_io; /* next block to be put under IO,
87 in dio_blocks units */
88
89 /*
90 * Deferred addition of a page to the dio. These variables are
91 * private to dio_send_cur_page(), submit_page_section() and
92 * dio_bio_add_page().
93 */
94 struct page *cur_page; /* The page */
95 unsigned cur_page_offset; /* Offset into it, in bytes */
96 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
97 sector_t cur_page_block; /* Where it starts */
98 loff_t cur_page_fs_offset; /* Offset in file */
99
100 /*
101 * Page fetching state. These variables belong to dio_refill_pages().
102 */
103 int curr_page; /* changes */
104 int total_pages; /* doesn't change */
105 unsigned long curr_user_address;/* changes */
106
107 /*
108 * Page queue. These variables belong to dio_refill_pages() and
109 * dio_get_page().
110 */
111 unsigned head; /* next page to process */
112 unsigned tail; /* last valid page + 1 */
113};
114
115/* dio_state communicated between submission path and end_io */
116struct dio {
117 int flags; /* doesn't change */
118 int rw;
119 struct inode *inode;
120 loff_t i_size; /* i_size when submitted */
121 dio_iodone_t *end_io; /* IO completion function */
122
123 void *private; /* copy from map_bh.b_private */
124
125 /* BIO completion state */
126 spinlock_t bio_lock; /* protects BIO fields below */
127 int page_errors; /* errno from get_user_pages() */
128 int is_async; /* is IO async ? */
129 int io_error; /* IO error in completion path */
130 unsigned long refcount; /* direct_io_worker() and bios */
131 struct bio *bio_list; /* singly linked via bi_private */
132 struct task_struct *waiter; /* waiting task (NULL if none) */
133
134 /* AIO related stuff */
135 struct kiocb *iocb; /* kiocb */
136 ssize_t result; /* IO result */
137
138 /*
139 * pages[] (and any fields placed after it) are not zeroed out at
140 * allocation time. Don't add new fields after pages[] unless you
141 * wish that they not be zeroed.
142 */
143 struct page *pages[DIO_PAGES]; /* page buffer */
144} ____cacheline_aligned_in_smp;
145
146static struct kmem_cache *dio_cache __read_mostly;
147
148/*
149 * How many pages are in the queue?
150 */
151static inline unsigned dio_pages_present(struct dio_submit *sdio)
152{
153 return sdio->tail - sdio->head;
154}
155
156/*
157 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
158 */
159static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
160{
161 int ret;
162 int nr_pages;
163
164 nr_pages = min(sdio->total_pages - sdio->curr_page, DIO_PAGES);
165 ret = get_user_pages_fast(
166 sdio->curr_user_address, /* Where from? */
167 nr_pages, /* How many pages? */
168 dio->rw == READ, /* Write to memory? */
169 &dio->pages[0]); /* Put results here */
170
171 if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) {
172 struct page *page = ZERO_PAGE(0);
173 /*
174 * A memory fault, but the filesystem has some outstanding
175 * mapped blocks. We need to use those blocks up to avoid
176 * leaking stale data in the file.
177 */
178 if (dio->page_errors == 0)
179 dio->page_errors = ret;
180 page_cache_get(page);
181 dio->pages[0] = page;
182 sdio->head = 0;
183 sdio->tail = 1;
184 ret = 0;
185 goto out;
186 }
187
188 if (ret >= 0) {
189 sdio->curr_user_address += ret * PAGE_SIZE;
190 sdio->curr_page += ret;
191 sdio->head = 0;
192 sdio->tail = ret;
193 ret = 0;
194 }
195out:
196 return ret;
197}
198
199/*
200 * Get another userspace page. Returns an ERR_PTR on error. Pages are
201 * buffered inside the dio so that we can call get_user_pages() against a
202 * decent number of pages, less frequently. To provide nicer use of the
203 * L1 cache.
204 */
205static inline struct page *dio_get_page(struct dio *dio,
206 struct dio_submit *sdio)
207{
208 if (dio_pages_present(sdio) == 0) {
209 int ret;
210
211 ret = dio_refill_pages(dio, sdio);
212 if (ret)
213 return ERR_PTR(ret);
214 BUG_ON(dio_pages_present(sdio) == 0);
215 }
216 return dio->pages[sdio->head++];
217}
218
219/**
220 * dio_complete() - called when all DIO BIO I/O has been completed
221 * @offset: the byte offset in the file of the completed operation
222 *
223 * This releases locks as dictated by the locking type, lets interested parties
224 * know that a DIO operation has completed, and calculates the resulting return
225 * code for the operation.
226 *
227 * It lets the filesystem know if it registered an interest earlier via
228 * get_block. Pass the private field of the map buffer_head so that
229 * filesystems can use it to hold additional state between get_block calls and
230 * dio_complete.
231 */
232static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret, bool is_async)
233{
234 ssize_t transferred = 0;
235
236 /*
237 * AIO submission can race with bio completion to get here while
238 * expecting to have the last io completed by bio completion.
239 * In that case -EIOCBQUEUED is in fact not an error we want
240 * to preserve through this call.
241 */
242 if (ret == -EIOCBQUEUED)
243 ret = 0;
244
245 if (dio->result) {
246 transferred = dio->result;
247
248 /* Check for short read case */
249 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
250 transferred = dio->i_size - offset;
251 }
252
253 if (ret == 0)
254 ret = dio->page_errors;
255 if (ret == 0)
256 ret = dio->io_error;
257 if (ret == 0)
258 ret = transferred;
259
260 if (dio->end_io && dio->result) {
261 dio->end_io(dio->iocb, offset, transferred,
262 dio->private, ret, is_async);
263 } else {
264 if (is_async)
265 aio_complete(dio->iocb, ret, 0);
266 inode_dio_done(dio->inode);
267 }
268
269 return ret;
270}
271
272static int dio_bio_complete(struct dio *dio, struct bio *bio);
273/*
274 * Asynchronous IO callback.
275 */
276static void dio_bio_end_aio(struct bio *bio, int error)
277{
278 struct dio *dio = bio->bi_private;
279 unsigned long remaining;
280 unsigned long flags;
281
282 /* cleanup the bio */
283 dio_bio_complete(dio, bio);
284
285 spin_lock_irqsave(&dio->bio_lock, flags);
286 remaining = --dio->refcount;
287 if (remaining == 1 && dio->waiter)
288 wake_up_process(dio->waiter);
289 spin_unlock_irqrestore(&dio->bio_lock, flags);
290
291 if (remaining == 0) {
292 dio_complete(dio, dio->iocb->ki_pos, 0, true);
293 kmem_cache_free(dio_cache, dio);
294 }
295}
296
297/*
298 * The BIO completion handler simply queues the BIO up for the process-context
299 * handler.
300 *
301 * During I/O bi_private points at the dio. After I/O, bi_private is used to
302 * implement a singly-linked list of completed BIOs, at dio->bio_list.
303 */
304static void dio_bio_end_io(struct bio *bio, int error)
305{
306 struct dio *dio = bio->bi_private;
307 unsigned long flags;
308
309 spin_lock_irqsave(&dio->bio_lock, flags);
310 bio->bi_private = dio->bio_list;
311 dio->bio_list = bio;
312 if (--dio->refcount == 1 && dio->waiter)
313 wake_up_process(dio->waiter);
314 spin_unlock_irqrestore(&dio->bio_lock, flags);
315}
316
317/**
318 * dio_end_io - handle the end io action for the given bio
319 * @bio: The direct io bio thats being completed
320 * @error: Error if there was one
321 *
322 * This is meant to be called by any filesystem that uses their own dio_submit_t
323 * so that the DIO specific endio actions are dealt with after the filesystem
324 * has done it's completion work.
325 */
326void dio_end_io(struct bio *bio, int error)
327{
328 struct dio *dio = bio->bi_private;
329
330 if (dio->is_async)
331 dio_bio_end_aio(bio, error);
332 else
333 dio_bio_end_io(bio, error);
334}
335EXPORT_SYMBOL_GPL(dio_end_io);
336
337static inline void
338dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
339 struct block_device *bdev,
340 sector_t first_sector, int nr_vecs)
341{
342 struct bio *bio;
343
344 /*
345 * bio_alloc() is guaranteed to return a bio when called with
346 * __GFP_WAIT and we request a valid number of vectors.
347 */
348 bio = bio_alloc(GFP_KERNEL, nr_vecs);
349
350 bio->bi_bdev = bdev;
351 bio->bi_sector = first_sector;
352 if (dio->is_async)
353 bio->bi_end_io = dio_bio_end_aio;
354 else
355 bio->bi_end_io = dio_bio_end_io;
356
357 sdio->bio = bio;
358 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
359}
360
361/*
362 * In the AIO read case we speculatively dirty the pages before starting IO.
363 * During IO completion, any of these pages which happen to have been written
364 * back will be redirtied by bio_check_pages_dirty().
365 *
366 * bios hold a dio reference between submit_bio and ->end_io.
367 */
368static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
369{
370 struct bio *bio = sdio->bio;
371 unsigned long flags;
372
373 bio->bi_private = dio;
374
375 spin_lock_irqsave(&dio->bio_lock, flags);
376 dio->refcount++;
377 spin_unlock_irqrestore(&dio->bio_lock, flags);
378
379 if (dio->is_async && dio->rw == READ)
380 bio_set_pages_dirty(bio);
381
382 if (sdio->submit_io)
383 sdio->submit_io(dio->rw, bio, dio->inode,
384 sdio->logical_offset_in_bio);
385 else
386 submit_bio(dio->rw, bio);
387
388 sdio->bio = NULL;
389 sdio->boundary = 0;
390 sdio->logical_offset_in_bio = 0;
391}
392
393/*
394 * Release any resources in case of a failure
395 */
396static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
397{
398 while (dio_pages_present(sdio))
399 page_cache_release(dio_get_page(dio, sdio));
400}
401
402/*
403 * Wait for the next BIO to complete. Remove it and return it. NULL is
404 * returned once all BIOs have been completed. This must only be called once
405 * all bios have been issued so that dio->refcount can only decrease. This
406 * requires that that the caller hold a reference on the dio.
407 */
408static struct bio *dio_await_one(struct dio *dio)
409{
410 unsigned long flags;
411 struct bio *bio = NULL;
412
413 spin_lock_irqsave(&dio->bio_lock, flags);
414
415 /*
416 * Wait as long as the list is empty and there are bios in flight. bio
417 * completion drops the count, maybe adds to the list, and wakes while
418 * holding the bio_lock so we don't need set_current_state()'s barrier
419 * and can call it after testing our condition.
420 */
421 while (dio->refcount > 1 && dio->bio_list == NULL) {
422 __set_current_state(TASK_UNINTERRUPTIBLE);
423 dio->waiter = current;
424 spin_unlock_irqrestore(&dio->bio_lock, flags);
425 io_schedule();
426 /* wake up sets us TASK_RUNNING */
427 spin_lock_irqsave(&dio->bio_lock, flags);
428 dio->waiter = NULL;
429 }
430 if (dio->bio_list) {
431 bio = dio->bio_list;
432 dio->bio_list = bio->bi_private;
433 }
434 spin_unlock_irqrestore(&dio->bio_lock, flags);
435 return bio;
436}
437
438/*
439 * Process one completed BIO. No locks are held.
440 */
441static int dio_bio_complete(struct dio *dio, struct bio *bio)
442{
443 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
444 struct bio_vec *bvec = bio->bi_io_vec;
445 int page_no;
446
447 if (!uptodate)
448 dio->io_error = -EIO;
449
450 if (dio->is_async && dio->rw == READ) {
451 bio_check_pages_dirty(bio); /* transfers ownership */
452 } else {
453 for (page_no = 0; page_no < bio->bi_vcnt; page_no++) {
454 struct page *page = bvec[page_no].bv_page;
455
456 if (dio->rw == READ && !PageCompound(page))
457 set_page_dirty_lock(page);
458 page_cache_release(page);
459 }
460 bio_put(bio);
461 }
462 return uptodate ? 0 : -EIO;
463}
464
465/*
466 * Wait on and process all in-flight BIOs. This must only be called once
467 * all bios have been issued so that the refcount can only decrease.
468 * This just waits for all bios to make it through dio_bio_complete. IO
469 * errors are propagated through dio->io_error and should be propagated via
470 * dio_complete().
471 */
472static void dio_await_completion(struct dio *dio)
473{
474 struct bio *bio;
475 do {
476 bio = dio_await_one(dio);
477 if (bio)
478 dio_bio_complete(dio, bio);
479 } while (bio);
480}
481
482/*
483 * A really large O_DIRECT read or write can generate a lot of BIOs. So
484 * to keep the memory consumption sane we periodically reap any completed BIOs
485 * during the BIO generation phase.
486 *
487 * This also helps to limit the peak amount of pinned userspace memory.
488 */
489static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
490{
491 int ret = 0;
492
493 if (sdio->reap_counter++ >= 64) {
494 while (dio->bio_list) {
495 unsigned long flags;
496 struct bio *bio;
497 int ret2;
498
499 spin_lock_irqsave(&dio->bio_lock, flags);
500 bio = dio->bio_list;
501 dio->bio_list = bio->bi_private;
502 spin_unlock_irqrestore(&dio->bio_lock, flags);
503 ret2 = dio_bio_complete(dio, bio);
504 if (ret == 0)
505 ret = ret2;
506 }
507 sdio->reap_counter = 0;
508 }
509 return ret;
510}
511
512/*
513 * Call into the fs to map some more disk blocks. We record the current number
514 * of available blocks at sdio->blocks_available. These are in units of the
515 * fs blocksize, (1 << inode->i_blkbits).
516 *
517 * The fs is allowed to map lots of blocks at once. If it wants to do that,
518 * it uses the passed inode-relative block number as the file offset, as usual.
519 *
520 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
521 * has remaining to do. The fs should not map more than this number of blocks.
522 *
523 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
524 * indicate how much contiguous disk space has been made available at
525 * bh->b_blocknr.
526 *
527 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
528 * This isn't very efficient...
529 *
530 * In the case of filesystem holes: the fs may return an arbitrarily-large
531 * hole by returning an appropriate value in b_size and by clearing
532 * buffer_mapped(). However the direct-io code will only process holes one
533 * block at a time - it will repeatedly call get_block() as it walks the hole.
534 */
535static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
536 struct buffer_head *map_bh)
537{
538 int ret;
539 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
540 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
541 unsigned long fs_count; /* Number of filesystem-sized blocks */
542 int create;
543
544 /*
545 * If there was a memory error and we've overwritten all the
546 * mapped blocks then we can now return that memory error
547 */
548 ret = dio->page_errors;
549 if (ret == 0) {
550 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
551 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
552 fs_endblk = (sdio->final_block_in_request - 1) >>
553 sdio->blkfactor;
554 fs_count = fs_endblk - fs_startblk + 1;
555
556 map_bh->b_state = 0;
557 map_bh->b_size = fs_count << dio->inode->i_blkbits;
558
559 /*
560 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we
561 * forbid block creations: only overwrites are permitted.
562 * We will return early to the caller once we see an
563 * unmapped buffer head returned, and the caller will fall
564 * back to buffered I/O.
565 *
566 * Otherwise the decision is left to the get_blocks method,
567 * which may decide to handle it or also return an unmapped
568 * buffer head.
569 */
570 create = dio->rw & WRITE;
571 if (dio->flags & DIO_SKIP_HOLES) {
572 if (sdio->block_in_file < (i_size_read(dio->inode) >>
573 sdio->blkbits))
574 create = 0;
575 }
576
577 ret = (*sdio->get_block)(dio->inode, fs_startblk,
578 map_bh, create);
579
580 /* Store for completion */
581 dio->private = map_bh->b_private;
582 }
583 return ret;
584}
585
586/*
587 * There is no bio. Make one now.
588 */
589static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
590 sector_t start_sector, struct buffer_head *map_bh)
591{
592 sector_t sector;
593 int ret, nr_pages;
594
595 ret = dio_bio_reap(dio, sdio);
596 if (ret)
597 goto out;
598 sector = start_sector << (sdio->blkbits - 9);
599 nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(map_bh->b_bdev));
600 nr_pages = min(nr_pages, BIO_MAX_PAGES);
601 BUG_ON(nr_pages <= 0);
602 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
603 sdio->boundary = 0;
604out:
605 return ret;
606}
607
608/*
609 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
610 * that was successful then update final_block_in_bio and take a ref against
611 * the just-added page.
612 *
613 * Return zero on success. Non-zero means the caller needs to start a new BIO.
614 */
615static inline int dio_bio_add_page(struct dio_submit *sdio)
616{
617 int ret;
618
619 ret = bio_add_page(sdio->bio, sdio->cur_page,
620 sdio->cur_page_len, sdio->cur_page_offset);
621 if (ret == sdio->cur_page_len) {
622 /*
623 * Decrement count only, if we are done with this page
624 */
625 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
626 sdio->pages_in_io--;
627 page_cache_get(sdio->cur_page);
628 sdio->final_block_in_bio = sdio->cur_page_block +
629 (sdio->cur_page_len >> sdio->blkbits);
630 ret = 0;
631 } else {
632 ret = 1;
633 }
634 return ret;
635}
636
637/*
638 * Put cur_page under IO. The section of cur_page which is described by
639 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
640 * starts on-disk at cur_page_block.
641 *
642 * We take a ref against the page here (on behalf of its presence in the bio).
643 *
644 * The caller of this function is responsible for removing cur_page from the
645 * dio, and for dropping the refcount which came from that presence.
646 */
647static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
648 struct buffer_head *map_bh)
649{
650 int ret = 0;
651
652 if (sdio->bio) {
653 loff_t cur_offset = sdio->cur_page_fs_offset;
654 loff_t bio_next_offset = sdio->logical_offset_in_bio +
655 sdio->bio->bi_size;
656
657 /*
658 * See whether this new request is contiguous with the old.
659 *
660 * Btrfs cannot handle having logically non-contiguous requests
661 * submitted. For example if you have
662 *
663 * Logical: [0-4095][HOLE][8192-12287]
664 * Physical: [0-4095] [4096-8191]
665 *
666 * We cannot submit those pages together as one BIO. So if our
667 * current logical offset in the file does not equal what would
668 * be the next logical offset in the bio, submit the bio we
669 * have.
670 */
671 if (sdio->final_block_in_bio != sdio->cur_page_block ||
672 cur_offset != bio_next_offset)
673 dio_bio_submit(dio, sdio);
674 /*
675 * Submit now if the underlying fs is about to perform a
676 * metadata read
677 */
678 else if (sdio->boundary)
679 dio_bio_submit(dio, sdio);
680 }
681
682 if (sdio->bio == NULL) {
683 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
684 if (ret)
685 goto out;
686 }
687
688 if (dio_bio_add_page(sdio) != 0) {
689 dio_bio_submit(dio, sdio);
690 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
691 if (ret == 0) {
692 ret = dio_bio_add_page(sdio);
693 BUG_ON(ret != 0);
694 }
695 }
696out:
697 return ret;
698}
699
700/*
701 * An autonomous function to put a chunk of a page under deferred IO.
702 *
703 * The caller doesn't actually know (or care) whether this piece of page is in
704 * a BIO, or is under IO or whatever. We just take care of all possible
705 * situations here. The separation between the logic of do_direct_IO() and
706 * that of submit_page_section() is important for clarity. Please don't break.
707 *
708 * The chunk of page starts on-disk at blocknr.
709 *
710 * We perform deferred IO, by recording the last-submitted page inside our
711 * private part of the dio structure. If possible, we just expand the IO
712 * across that page here.
713 *
714 * If that doesn't work out then we put the old page into the bio and add this
715 * page to the dio instead.
716 */
717static inline int
718submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
719 unsigned offset, unsigned len, sector_t blocknr,
720 struct buffer_head *map_bh)
721{
722 int ret = 0;
723
724 if (dio->rw & WRITE) {
725 /*
726 * Read accounting is performed in submit_bio()
727 */
728 task_io_account_write(len);
729 }
730
731 /*
732 * Can we just grow the current page's presence in the dio?
733 */
734 if (sdio->cur_page == page &&
735 sdio->cur_page_offset + sdio->cur_page_len == offset &&
736 sdio->cur_page_block +
737 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
738 sdio->cur_page_len += len;
739
740 /*
741 * If sdio->boundary then we want to schedule the IO now to
742 * avoid metadata seeks.
743 */
744 if (sdio->boundary) {
745 ret = dio_send_cur_page(dio, sdio, map_bh);
746 page_cache_release(sdio->cur_page);
747 sdio->cur_page = NULL;
748 }
749 goto out;
750 }
751
752 /*
753 * If there's a deferred page already there then send it.
754 */
755 if (sdio->cur_page) {
756 ret = dio_send_cur_page(dio, sdio, map_bh);
757 page_cache_release(sdio->cur_page);
758 sdio->cur_page = NULL;
759 if (ret)
760 goto out;
761 }
762
763 page_cache_get(page); /* It is in dio */
764 sdio->cur_page = page;
765 sdio->cur_page_offset = offset;
766 sdio->cur_page_len = len;
767 sdio->cur_page_block = blocknr;
768 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
769out:
770 return ret;
771}
772
773/*
774 * Clean any dirty buffers in the blockdev mapping which alias newly-created
775 * file blocks. Only called for S_ISREG files - blockdevs do not set
776 * buffer_new
777 */
778static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
779{
780 unsigned i;
781 unsigned nblocks;
782
783 nblocks = map_bh->b_size >> dio->inode->i_blkbits;
784
785 for (i = 0; i < nblocks; i++) {
786 unmap_underlying_metadata(map_bh->b_bdev,
787 map_bh->b_blocknr + i);
788 }
789}
790
791/*
792 * If we are not writing the entire block and get_block() allocated
793 * the block for us, we need to fill-in the unused portion of the
794 * block with zeros. This happens only if user-buffer, fileoffset or
795 * io length is not filesystem block-size multiple.
796 *
797 * `end' is zero if we're doing the start of the IO, 1 at the end of the
798 * IO.
799 */
800static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
801 int end, struct buffer_head *map_bh)
802{
803 unsigned dio_blocks_per_fs_block;
804 unsigned this_chunk_blocks; /* In dio_blocks */
805 unsigned this_chunk_bytes;
806 struct page *page;
807
808 sdio->start_zero_done = 1;
809 if (!sdio->blkfactor || !buffer_new(map_bh))
810 return;
811
812 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
813 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
814
815 if (!this_chunk_blocks)
816 return;
817
818 /*
819 * We need to zero out part of an fs block. It is either at the
820 * beginning or the end of the fs block.
821 */
822 if (end)
823 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
824
825 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
826
827 page = ZERO_PAGE(0);
828 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
829 sdio->next_block_for_io, map_bh))
830 return;
831
832 sdio->next_block_for_io += this_chunk_blocks;
833}
834
835/*
836 * Walk the user pages, and the file, mapping blocks to disk and generating
837 * a sequence of (page,offset,len,block) mappings. These mappings are injected
838 * into submit_page_section(), which takes care of the next stage of submission
839 *
840 * Direct IO against a blockdev is different from a file. Because we can
841 * happily perform page-sized but 512-byte aligned IOs. It is important that
842 * blockdev IO be able to have fine alignment and large sizes.
843 *
844 * So what we do is to permit the ->get_block function to populate bh.b_size
845 * with the size of IO which is permitted at this offset and this i_blkbits.
846 *
847 * For best results, the blockdev should be set up with 512-byte i_blkbits and
848 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
849 * fine alignment but still allows this function to work in PAGE_SIZE units.
850 */
851static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
852 struct buffer_head *map_bh)
853{
854 const unsigned blkbits = sdio->blkbits;
855 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
856 struct page *page;
857 unsigned block_in_page;
858 int ret = 0;
859
860 /* The I/O can start at any block offset within the first page */
861 block_in_page = sdio->first_block_in_page;
862
863 while (sdio->block_in_file < sdio->final_block_in_request) {
864 page = dio_get_page(dio, sdio);
865 if (IS_ERR(page)) {
866 ret = PTR_ERR(page);
867 goto out;
868 }
869
870 while (block_in_page < blocks_per_page) {
871 unsigned offset_in_page = block_in_page << blkbits;
872 unsigned this_chunk_bytes; /* # of bytes mapped */
873 unsigned this_chunk_blocks; /* # of blocks */
874 unsigned u;
875
876 if (sdio->blocks_available == 0) {
877 /*
878 * Need to go and map some more disk
879 */
880 unsigned long blkmask;
881 unsigned long dio_remainder;
882
883 ret = get_more_blocks(dio, sdio, map_bh);
884 if (ret) {
885 page_cache_release(page);
886 goto out;
887 }
888 if (!buffer_mapped(map_bh))
889 goto do_holes;
890
891 sdio->blocks_available =
892 map_bh->b_size >> sdio->blkbits;
893 sdio->next_block_for_io =
894 map_bh->b_blocknr << sdio->blkfactor;
895 if (buffer_new(map_bh))
896 clean_blockdev_aliases(dio, map_bh);
897
898 if (!sdio->blkfactor)
899 goto do_holes;
900
901 blkmask = (1 << sdio->blkfactor) - 1;
902 dio_remainder = (sdio->block_in_file & blkmask);
903
904 /*
905 * If we are at the start of IO and that IO
906 * starts partway into a fs-block,
907 * dio_remainder will be non-zero. If the IO
908 * is a read then we can simply advance the IO
909 * cursor to the first block which is to be
910 * read. But if the IO is a write and the
911 * block was newly allocated we cannot do that;
912 * the start of the fs block must be zeroed out
913 * on-disk
914 */
915 if (!buffer_new(map_bh))
916 sdio->next_block_for_io += dio_remainder;
917 sdio->blocks_available -= dio_remainder;
918 }
919do_holes:
920 /* Handle holes */
921 if (!buffer_mapped(map_bh)) {
922 loff_t i_size_aligned;
923
924 /* AKPM: eargh, -ENOTBLK is a hack */
925 if (dio->rw & WRITE) {
926 page_cache_release(page);
927 return -ENOTBLK;
928 }
929
930 /*
931 * Be sure to account for a partial block as the
932 * last block in the file
933 */
934 i_size_aligned = ALIGN(i_size_read(dio->inode),
935 1 << blkbits);
936 if (sdio->block_in_file >=
937 i_size_aligned >> blkbits) {
938 /* We hit eof */
939 page_cache_release(page);
940 goto out;
941 }
942 zero_user(page, block_in_page << blkbits,
943 1 << blkbits);
944 sdio->block_in_file++;
945 block_in_page++;
946 goto next_block;
947 }
948
949 /*
950 * If we're performing IO which has an alignment which
951 * is finer than the underlying fs, go check to see if
952 * we must zero out the start of this block.
953 */
954 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
955 dio_zero_block(dio, sdio, 0, map_bh);
956
957 /*
958 * Work out, in this_chunk_blocks, how much disk we
959 * can add to this page
960 */
961 this_chunk_blocks = sdio->blocks_available;
962 u = (PAGE_SIZE - offset_in_page) >> blkbits;
963 if (this_chunk_blocks > u)
964 this_chunk_blocks = u;
965 u = sdio->final_block_in_request - sdio->block_in_file;
966 if (this_chunk_blocks > u)
967 this_chunk_blocks = u;
968 this_chunk_bytes = this_chunk_blocks << blkbits;
969 BUG_ON(this_chunk_bytes == 0);
970
971 sdio->boundary = buffer_boundary(map_bh);
972 ret = submit_page_section(dio, sdio, page,
973 offset_in_page,
974 this_chunk_bytes,
975 sdio->next_block_for_io,
976 map_bh);
977 if (ret) {
978 page_cache_release(page);
979 goto out;
980 }
981 sdio->next_block_for_io += this_chunk_blocks;
982
983 sdio->block_in_file += this_chunk_blocks;
984 block_in_page += this_chunk_blocks;
985 sdio->blocks_available -= this_chunk_blocks;
986next_block:
987 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
988 if (sdio->block_in_file == sdio->final_block_in_request)
989 break;
990 }
991
992 /* Drop the ref which was taken in get_user_pages() */
993 page_cache_release(page);
994 block_in_page = 0;
995 }
996out:
997 return ret;
998}
999
1000static inline int drop_refcount(struct dio *dio)
1001{
1002 int ret2;
1003 unsigned long flags;
1004
1005 /*
1006 * Sync will always be dropping the final ref and completing the
1007 * operation. AIO can if it was a broken operation described above or
1008 * in fact if all the bios race to complete before we get here. In
1009 * that case dio_complete() translates the EIOCBQUEUED into the proper
1010 * return code that the caller will hand to aio_complete().
1011 *
1012 * This is managed by the bio_lock instead of being an atomic_t so that
1013 * completion paths can drop their ref and use the remaining count to
1014 * decide to wake the submission path atomically.
1015 */
1016 spin_lock_irqsave(&dio->bio_lock, flags);
1017 ret2 = --dio->refcount;
1018 spin_unlock_irqrestore(&dio->bio_lock, flags);
1019 return ret2;
1020}
1021
1022/*
1023 * This is a library function for use by filesystem drivers.
1024 *
1025 * The locking rules are governed by the flags parameter:
1026 * - if the flags value contains DIO_LOCKING we use a fancy locking
1027 * scheme for dumb filesystems.
1028 * For writes this function is called under i_mutex and returns with
1029 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1030 * taken and dropped again before returning.
1031 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1032 * internal locking but rather rely on the filesystem to synchronize
1033 * direct I/O reads/writes versus each other and truncate.
1034 *
1035 * To help with locking against truncate we incremented the i_dio_count
1036 * counter before starting direct I/O, and decrement it once we are done.
1037 * Truncate can wait for it to reach zero to provide exclusion. It is
1038 * expected that filesystem provide exclusion between new direct I/O
1039 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1040 * but other filesystems need to take care of this on their own.
1041 *
1042 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1043 * is always inlined. Otherwise gcc is unable to split the structure into
1044 * individual fields and will generate much worse code. This is important
1045 * for the whole file.
1046 */
1047static inline ssize_t
1048do_blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1049 struct block_device *bdev, const struct iovec *iov, loff_t offset,
1050 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1051 dio_submit_t submit_io, int flags)
1052{
1053 int seg;
1054 size_t size;
1055 unsigned long addr;
1056 unsigned blkbits = inode->i_blkbits;
1057 unsigned blocksize_mask = (1 << blkbits) - 1;
1058 ssize_t retval = -EINVAL;
1059 loff_t end = offset;
1060 struct dio *dio;
1061 struct dio_submit sdio = { 0, };
1062 unsigned long user_addr;
1063 size_t bytes;
1064 struct buffer_head map_bh = { 0, };
1065
1066 if (rw & WRITE)
1067 rw = WRITE_ODIRECT;
1068
1069 /*
1070 * Avoid references to bdev if not absolutely needed to give
1071 * the early prefetch in the caller enough time.
1072 */
1073
1074 if (offset & blocksize_mask) {
1075 if (bdev)
1076 blkbits = blksize_bits(bdev_logical_block_size(bdev));
1077 blocksize_mask = (1 << blkbits) - 1;
1078 if (offset & blocksize_mask)
1079 goto out;
1080 }
1081
1082 /* Check the memory alignment. Blocks cannot straddle pages */
1083 for (seg = 0; seg < nr_segs; seg++) {
1084 addr = (unsigned long)iov[seg].iov_base;
1085 size = iov[seg].iov_len;
1086 end += size;
1087 if (unlikely((addr & blocksize_mask) ||
1088 (size & blocksize_mask))) {
1089 if (bdev)
1090 blkbits = blksize_bits(
1091 bdev_logical_block_size(bdev));
1092 blocksize_mask = (1 << blkbits) - 1;
1093 if ((addr & blocksize_mask) || (size & blocksize_mask))
1094 goto out;
1095 }
1096 }
1097
1098 /* watch out for a 0 len io from a tricksy fs */
1099 if (rw == READ && end == offset)
1100 return 0;
1101
1102 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
1103 retval = -ENOMEM;
1104 if (!dio)
1105 goto out;
1106 /*
1107 * Believe it or not, zeroing out the page array caused a .5%
1108 * performance regression in a database benchmark. So, we take
1109 * care to only zero out what's needed.
1110 */
1111 memset(dio, 0, offsetof(struct dio, pages));
1112
1113 dio->flags = flags;
1114 if (dio->flags & DIO_LOCKING) {
1115 if (rw == READ) {
1116 struct address_space *mapping =
1117 iocb->ki_filp->f_mapping;
1118
1119 /* will be released by direct_io_worker */
1120 mutex_lock(&inode->i_mutex);
1121
1122 retval = filemap_write_and_wait_range(mapping, offset,
1123 end - 1);
1124 if (retval) {
1125 mutex_unlock(&inode->i_mutex);
1126 kmem_cache_free(dio_cache, dio);
1127 goto out;
1128 }
1129 }
1130 }
1131
1132 /*
1133 * Will be decremented at I/O completion time.
1134 */
1135 atomic_inc(&inode->i_dio_count);
1136
1137 /*
1138 * For file extending writes updating i_size before data
1139 * writeouts complete can expose uninitialized blocks. So
1140 * even for AIO, we need to wait for i/o to complete before
1141 * returning in this case.
1142 */
1143 dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
1144 (end > i_size_read(inode)));
1145
1146 retval = 0;
1147
1148 dio->inode = inode;
1149 dio->rw = rw;
1150 sdio.blkbits = blkbits;
1151 sdio.blkfactor = inode->i_blkbits - blkbits;
1152 sdio.block_in_file = offset >> blkbits;
1153
1154 sdio.get_block = get_block;
1155 dio->end_io = end_io;
1156 sdio.submit_io = submit_io;
1157 sdio.final_block_in_bio = -1;
1158 sdio.next_block_for_io = -1;
1159
1160 dio->iocb = iocb;
1161 dio->i_size = i_size_read(inode);
1162
1163 spin_lock_init(&dio->bio_lock);
1164 dio->refcount = 1;
1165
1166 /*
1167 * In case of non-aligned buffers, we may need 2 more
1168 * pages since we need to zero out first and last block.
1169 */
1170 if (unlikely(sdio.blkfactor))
1171 sdio.pages_in_io = 2;
1172
1173 for (seg = 0; seg < nr_segs; seg++) {
1174 user_addr = (unsigned long)iov[seg].iov_base;
1175 sdio.pages_in_io +=
1176 ((user_addr + iov[seg].iov_len + PAGE_SIZE-1) /
1177 PAGE_SIZE - user_addr / PAGE_SIZE);
1178 }
1179
1180 for (seg = 0; seg < nr_segs; seg++) {
1181 user_addr = (unsigned long)iov[seg].iov_base;
1182 sdio.size += bytes = iov[seg].iov_len;
1183
1184 /* Index into the first page of the first block */
1185 sdio.first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
1186 sdio.final_block_in_request = sdio.block_in_file +
1187 (bytes >> blkbits);
1188 /* Page fetching state */
1189 sdio.head = 0;
1190 sdio.tail = 0;
1191 sdio.curr_page = 0;
1192
1193 sdio.total_pages = 0;
1194 if (user_addr & (PAGE_SIZE-1)) {
1195 sdio.total_pages++;
1196 bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
1197 }
1198 sdio.total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
1199 sdio.curr_user_address = user_addr;
1200
1201 retval = do_direct_IO(dio, &sdio, &map_bh);
1202
1203 dio->result += iov[seg].iov_len -
1204 ((sdio.final_block_in_request - sdio.block_in_file) <<
1205 blkbits);
1206
1207 if (retval) {
1208 dio_cleanup(dio, &sdio);
1209 break;
1210 }
1211 } /* end iovec loop */
1212
1213 if (retval == -ENOTBLK) {
1214 /*
1215 * The remaining part of the request will be
1216 * be handled by buffered I/O when we return
1217 */
1218 retval = 0;
1219 }
1220 /*
1221 * There may be some unwritten disk at the end of a part-written
1222 * fs-block-sized block. Go zero that now.
1223 */
1224 dio_zero_block(dio, &sdio, 1, &map_bh);
1225
1226 if (sdio.cur_page) {
1227 ssize_t ret2;
1228
1229 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1230 if (retval == 0)
1231 retval = ret2;
1232 page_cache_release(sdio.cur_page);
1233 sdio.cur_page = NULL;
1234 }
1235 if (sdio.bio)
1236 dio_bio_submit(dio, &sdio);
1237
1238 /*
1239 * It is possible that, we return short IO due to end of file.
1240 * In that case, we need to release all the pages we got hold on.
1241 */
1242 dio_cleanup(dio, &sdio);
1243
1244 /*
1245 * All block lookups have been performed. For READ requests
1246 * we can let i_mutex go now that its achieved its purpose
1247 * of protecting us from looking up uninitialized blocks.
1248 */
1249 if (rw == READ && (dio->flags & DIO_LOCKING))
1250 mutex_unlock(&dio->inode->i_mutex);
1251
1252 /*
1253 * The only time we want to leave bios in flight is when a successful
1254 * partial aio read or full aio write have been setup. In that case
1255 * bio completion will call aio_complete. The only time it's safe to
1256 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1257 * This had *better* be the only place that raises -EIOCBQUEUED.
1258 */
1259 BUG_ON(retval == -EIOCBQUEUED);
1260 if (dio->is_async && retval == 0 && dio->result &&
1261 ((rw & READ) || (dio->result == sdio.size)))
1262 retval = -EIOCBQUEUED;
1263
1264 if (retval != -EIOCBQUEUED)
1265 dio_await_completion(dio);
1266
1267 if (drop_refcount(dio) == 0) {
1268 retval = dio_complete(dio, offset, retval, false);
1269 kmem_cache_free(dio_cache, dio);
1270 } else
1271 BUG_ON(retval != -EIOCBQUEUED);
1272
1273out:
1274 return retval;
1275}
1276
1277ssize_t
1278__blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1279 struct block_device *bdev, const struct iovec *iov, loff_t offset,
1280 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1281 dio_submit_t submit_io, int flags)
1282{
1283 /*
1284 * The block device state is needed in the end to finally
1285 * submit everything. Since it's likely to be cache cold
1286 * prefetch it here as first thing to hide some of the
1287 * latency.
1288 *
1289 * Attempt to prefetch the pieces we likely need later.
1290 */
1291 prefetch(&bdev->bd_disk->part_tbl);
1292 prefetch(bdev->bd_queue);
1293 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1294
1295 return do_blockdev_direct_IO(rw, iocb, inode, bdev, iov, offset,
1296 nr_segs, get_block, end_io,
1297 submit_io, flags);
1298}
1299
1300EXPORT_SYMBOL(__blockdev_direct_IO);
1301
1302static __init int dio_init(void)
1303{
1304 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1305 return 0;
1306}
1307module_init(dio_init)