]> git.ipfire.org Git - people/arne_f/kernel.git/blame - fs/direct-io.c
fs: use helper bio_add_page() instead of open coding on bi_io_vec
[people/arne_f/kernel.git] / fs / direct-io.c
CommitLineData
1da177e4
LT
1/*
2 * fs/direct-io.c
3 *
4 * Copyright (C) 2002, Linus Torvalds.
5 *
6 * O_DIRECT
7 *
e1f8e874 8 * 04Jul2002 Andrew Morton
1da177e4
LT
9 * Initial version
10 * 11Sep2002 janetinc@us.ibm.com
11 * added readv/writev support.
e1f8e874 12 * 29Oct2002 Andrew Morton
1da177e4
LT
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>
98c4d57d 30#include <linux/task_io_accounting_ops.h>
1da177e4
LT
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>
60063497 38#include <linux/atomic.h>
65dd2aa9 39#include <linux/prefetch.h>
1da177e4
LT
40
41/*
42 * How many user pages to map in one call to get_user_pages(). This determines
cde1ecb3 43 * the size of a structure in the slab cache
1da177e4
LT
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.
1da177e4
LT
57 */
58
eb28be2b
AK
59/* dio_state only used in the submission path */
60
61struct dio_submit {
1da177e4 62 struct bio *bio; /* bio under assembly */
1da177e4
LT
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 */
1da177e4
LT
73 sector_t block_in_file; /* Current offset into the underlying
74 file in dio_block units. */
75 unsigned blocks_available; /* At block_in_file. changes */
0dc2bc49 76 int reap_counter; /* rate limit reaping */
1da177e4 77 sector_t final_block_in_request;/* doesn't change */
1da177e4 78 int boundary; /* prev block is at a boundary */
1d8fa7a2 79 get_block_t *get_block; /* block mapping function */
facd07b0 80 dio_submit_t *submit_io; /* IO submition function */
eb28be2b 81
facd07b0 82 loff_t logical_offset_in_bio; /* current first logical block in bio */
1da177e4
LT
83 sector_t final_block_in_bio; /* current final block in bio + 1 */
84 sector_t next_block_for_io; /* next block to be put under IO,
85 in dio_blocks units */
1da177e4
LT
86
87 /*
88 * Deferred addition of a page to the dio. These variables are
89 * private to dio_send_cur_page(), submit_page_section() and
90 * dio_bio_add_page().
91 */
92 struct page *cur_page; /* The page */
93 unsigned cur_page_offset; /* Offset into it, in bytes */
94 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
95 sector_t cur_page_block; /* Where it starts */
facd07b0 96 loff_t cur_page_fs_offset; /* Offset in file */
1da177e4 97
7b2c99d1 98 struct iov_iter *iter;
1da177e4
LT
99 /*
100 * Page queue. These variables belong to dio_refill_pages() and
101 * dio_get_page().
102 */
1da177e4
LT
103 unsigned head; /* next page to process */
104 unsigned tail; /* last valid page + 1 */
7b2c99d1 105 size_t from, to;
eb28be2b
AK
106};
107
108/* dio_state communicated between submission path and end_io */
109struct dio {
110 int flags; /* doesn't change */
eb28be2b 111 int rw;
0dc2bc49 112 struct inode *inode;
eb28be2b
AK
113 loff_t i_size; /* i_size when submitted */
114 dio_iodone_t *end_io; /* IO completion function */
eb28be2b 115
18772641 116 void *private; /* copy from map_bh.b_private */
eb28be2b
AK
117
118 /* BIO completion state */
119 spinlock_t bio_lock; /* protects BIO fields below */
0dc2bc49
AK
120 int page_errors; /* errno from get_user_pages() */
121 int is_async; /* is IO async ? */
7b7a8665 122 bool defer_completion; /* defer AIO completion to workqueue? */
0dc2bc49 123 int io_error; /* IO error in completion path */
eb28be2b
AK
124 unsigned long refcount; /* direct_io_worker() and bios */
125 struct bio *bio_list; /* singly linked via bi_private */
126 struct task_struct *waiter; /* waiting task (NULL if none) */
127
128 /* AIO related stuff */
129 struct kiocb *iocb; /* kiocb */
eb28be2b
AK
130 ssize_t result; /* IO result */
131
23aee091
JM
132 /*
133 * pages[] (and any fields placed after it) are not zeroed out at
134 * allocation time. Don't add new fields after pages[] unless you
135 * wish that they not be zeroed.
136 */
7b7a8665
CH
137 union {
138 struct page *pages[DIO_PAGES]; /* page buffer */
139 struct work_struct complete_work;/* deferred AIO completion */
140 };
6e8267f5
AK
141} ____cacheline_aligned_in_smp;
142
143static struct kmem_cache *dio_cache __read_mostly;
1da177e4
LT
144
145/*
146 * How many pages are in the queue?
147 */
eb28be2b 148static inline unsigned dio_pages_present(struct dio_submit *sdio)
1da177e4 149{
eb28be2b 150 return sdio->tail - sdio->head;
1da177e4
LT
151}
152
153/*
154 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
155 */
ba253fbf 156static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
1da177e4 157{
7b2c99d1 158 ssize_t ret;
1da177e4 159
2c80929c 160 ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES,
7b2c99d1 161 &sdio->from);
1da177e4 162
eb28be2b 163 if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) {
557ed1fa 164 struct page *page = ZERO_PAGE(0);
1da177e4
LT
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;
b5810039
NP
172 page_cache_get(page);
173 dio->pages[0] = page;
eb28be2b
AK
174 sdio->head = 0;
175 sdio->tail = 1;
7b2c99d1
AV
176 sdio->from = 0;
177 sdio->to = PAGE_SIZE;
178 return 0;
1da177e4
LT
179 }
180
181 if (ret >= 0) {
7b2c99d1
AV
182 iov_iter_advance(sdio->iter, ret);
183 ret += sdio->from;
eb28be2b 184 sdio->head = 0;
7b2c99d1
AV
185 sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
186 sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1;
187 return 0;
1da177e4 188 }
1da177e4
LT
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 */
ba253fbf 198static inline struct page *dio_get_page(struct dio *dio,
6fcc5420 199 struct dio_submit *sdio)
1da177e4 200{
eb28be2b 201 if (dio_pages_present(sdio) == 0) {
1da177e4
LT
202 int ret;
203
eb28be2b 204 ret = dio_refill_pages(dio, sdio);
1da177e4
LT
205 if (ret)
206 return ERR_PTR(ret);
eb28be2b 207 BUG_ON(dio_pages_present(sdio) == 0);
1da177e4 208 }
6fcc5420 209 return dio->pages[sdio->head];
1da177e4
LT
210}
211
6d544bb4
ZB
212/**
213 * dio_complete() - called when all DIO BIO I/O has been completed
214 * @offset: the byte offset in the file of the completed operation
215 *
7b7a8665
CH
216 * This drops i_dio_count, lets interested parties know that a DIO operation
217 * has completed, and calculates the resulting return code for the operation.
6d544bb4
ZB
218 *
219 * It lets the filesystem know if it registered an interest earlier via
220 * get_block. Pass the private field of the map buffer_head so that
221 * filesystems can use it to hold additional state between get_block calls and
222 * dio_complete.
1da177e4 223 */
7b7a8665
CH
224static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret,
225 bool is_async)
1da177e4 226{
6d544bb4
ZB
227 ssize_t transferred = 0;
228
8459d86a
ZB
229 /*
230 * AIO submission can race with bio completion to get here while
231 * expecting to have the last io completed by bio completion.
232 * In that case -EIOCBQUEUED is in fact not an error we want
233 * to preserve through this call.
234 */
235 if (ret == -EIOCBQUEUED)
236 ret = 0;
237
6d544bb4
ZB
238 if (dio->result) {
239 transferred = dio->result;
240
241 /* Check for short read case */
242 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
243 transferred = dio->i_size - offset;
244 }
245
6d544bb4
ZB
246 if (ret == 0)
247 ret = dio->page_errors;
248 if (ret == 0)
249 ret = dio->io_error;
250 if (ret == 0)
251 ret = transferred;
252
7b7a8665
CH
253 if (dio->end_io && dio->result)
254 dio->end_io(dio->iocb, offset, transferred, dio->private);
255
fe0f07d0
JA
256 if (!(dio->flags & DIO_SKIP_DIO_COUNT))
257 inode_dio_end(dio->inode);
258
02afc27f
CH
259 if (is_async) {
260 if (dio->rw & WRITE) {
261 int err;
262
263 err = generic_write_sync(dio->iocb->ki_filp, offset,
264 transferred);
265 if (err < 0 && ret > 0)
266 ret = err;
267 }
268
04b2fa9f 269 dio->iocb->ki_complete(dio->iocb, ret, 0);
02afc27f 270 }
40e2e973 271
7b7a8665 272 kmem_cache_free(dio_cache, dio);
6d544bb4 273 return ret;
1da177e4
LT
274}
275
7b7a8665
CH
276static void dio_aio_complete_work(struct work_struct *work)
277{
278 struct dio *dio = container_of(work, struct dio, complete_work);
279
280 dio_complete(dio, dio->iocb->ki_pos, 0, true);
281}
282
1da177e4 283static int dio_bio_complete(struct dio *dio, struct bio *bio);
7b7a8665 284
1da177e4
LT
285/*
286 * Asynchronous IO callback.
287 */
4246a0b6 288static void dio_bio_end_aio(struct bio *bio)
1da177e4
LT
289{
290 struct dio *dio = bio->bi_private;
5eb6c7a2
ZB
291 unsigned long remaining;
292 unsigned long flags;
1da177e4 293
1da177e4
LT
294 /* cleanup the bio */
295 dio_bio_complete(dio, bio);
0273201e 296
5eb6c7a2
ZB
297 spin_lock_irqsave(&dio->bio_lock, flags);
298 remaining = --dio->refcount;
299 if (remaining == 1 && dio->waiter)
20258b2b 300 wake_up_process(dio->waiter);
5eb6c7a2 301 spin_unlock_irqrestore(&dio->bio_lock, flags);
20258b2b 302
8459d86a 303 if (remaining == 0) {
7b7a8665
CH
304 if (dio->result && dio->defer_completion) {
305 INIT_WORK(&dio->complete_work, dio_aio_complete_work);
306 queue_work(dio->inode->i_sb->s_dio_done_wq,
307 &dio->complete_work);
308 } else {
309 dio_complete(dio, dio->iocb->ki_pos, 0, true);
310 }
8459d86a 311 }
1da177e4
LT
312}
313
314/*
315 * The BIO completion handler simply queues the BIO up for the process-context
316 * handler.
317 *
318 * During I/O bi_private points at the dio. After I/O, bi_private is used to
319 * implement a singly-linked list of completed BIOs, at dio->bio_list.
320 */
4246a0b6 321static void dio_bio_end_io(struct bio *bio)
1da177e4
LT
322{
323 struct dio *dio = bio->bi_private;
324 unsigned long flags;
325
1da177e4
LT
326 spin_lock_irqsave(&dio->bio_lock, flags);
327 bio->bi_private = dio->bio_list;
328 dio->bio_list = bio;
5eb6c7a2 329 if (--dio->refcount == 1 && dio->waiter)
1da177e4
LT
330 wake_up_process(dio->waiter);
331 spin_unlock_irqrestore(&dio->bio_lock, flags);
1da177e4
LT
332}
333
facd07b0
JB
334/**
335 * dio_end_io - handle the end io action for the given bio
336 * @bio: The direct io bio thats being completed
337 * @error: Error if there was one
338 *
339 * This is meant to be called by any filesystem that uses their own dio_submit_t
340 * so that the DIO specific endio actions are dealt with after the filesystem
341 * has done it's completion work.
342 */
343void dio_end_io(struct bio *bio, int error)
344{
345 struct dio *dio = bio->bi_private;
346
347 if (dio->is_async)
4246a0b6 348 dio_bio_end_aio(bio);
facd07b0 349 else
4246a0b6 350 dio_bio_end_io(bio);
facd07b0
JB
351}
352EXPORT_SYMBOL_GPL(dio_end_io);
353
ba253fbf 354static inline void
eb28be2b
AK
355dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
356 struct block_device *bdev,
357 sector_t first_sector, int nr_vecs)
1da177e4
LT
358{
359 struct bio *bio;
360
20d9600c
DD
361 /*
362 * bio_alloc() is guaranteed to return a bio when called with
363 * __GFP_WAIT and we request a valid number of vectors.
364 */
1da177e4 365 bio = bio_alloc(GFP_KERNEL, nr_vecs);
1da177e4
LT
366
367 bio->bi_bdev = bdev;
4f024f37 368 bio->bi_iter.bi_sector = first_sector;
1da177e4
LT
369 if (dio->is_async)
370 bio->bi_end_io = dio_bio_end_aio;
371 else
372 bio->bi_end_io = dio_bio_end_io;
373
eb28be2b
AK
374 sdio->bio = bio;
375 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
1da177e4
LT
376}
377
378/*
379 * In the AIO read case we speculatively dirty the pages before starting IO.
380 * During IO completion, any of these pages which happen to have been written
381 * back will be redirtied by bio_check_pages_dirty().
0273201e
ZB
382 *
383 * bios hold a dio reference between submit_bio and ->end_io.
1da177e4 384 */
ba253fbf 385static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
1da177e4 386{
eb28be2b 387 struct bio *bio = sdio->bio;
5eb6c7a2 388 unsigned long flags;
1da177e4
LT
389
390 bio->bi_private = dio;
5eb6c7a2
ZB
391
392 spin_lock_irqsave(&dio->bio_lock, flags);
393 dio->refcount++;
394 spin_unlock_irqrestore(&dio->bio_lock, flags);
395
1da177e4
LT
396 if (dio->is_async && dio->rw == READ)
397 bio_set_pages_dirty(bio);
5eb6c7a2 398
eb28be2b
AK
399 if (sdio->submit_io)
400 sdio->submit_io(dio->rw, bio, dio->inode,
401 sdio->logical_offset_in_bio);
facd07b0
JB
402 else
403 submit_bio(dio->rw, bio);
1da177e4 404
eb28be2b
AK
405 sdio->bio = NULL;
406 sdio->boundary = 0;
407 sdio->logical_offset_in_bio = 0;
1da177e4
LT
408}
409
410/*
411 * Release any resources in case of a failure
412 */
ba253fbf 413static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
1da177e4 414{
7b2c99d1
AV
415 while (sdio->head < sdio->tail)
416 page_cache_release(dio->pages[sdio->head++]);
1da177e4
LT
417}
418
419/*
0273201e
ZB
420 * Wait for the next BIO to complete. Remove it and return it. NULL is
421 * returned once all BIOs have been completed. This must only be called once
422 * all bios have been issued so that dio->refcount can only decrease. This
423 * requires that that the caller hold a reference on the dio.
1da177e4
LT
424 */
425static struct bio *dio_await_one(struct dio *dio)
426{
427 unsigned long flags;
0273201e 428 struct bio *bio = NULL;
1da177e4
LT
429
430 spin_lock_irqsave(&dio->bio_lock, flags);
5eb6c7a2
ZB
431
432 /*
433 * Wait as long as the list is empty and there are bios in flight. bio
434 * completion drops the count, maybe adds to the list, and wakes while
435 * holding the bio_lock so we don't need set_current_state()'s barrier
436 * and can call it after testing our condition.
437 */
438 while (dio->refcount > 1 && dio->bio_list == NULL) {
439 __set_current_state(TASK_UNINTERRUPTIBLE);
440 dio->waiter = current;
441 spin_unlock_irqrestore(&dio->bio_lock, flags);
442 io_schedule();
443 /* wake up sets us TASK_RUNNING */
444 spin_lock_irqsave(&dio->bio_lock, flags);
445 dio->waiter = NULL;
1da177e4 446 }
0273201e
ZB
447 if (dio->bio_list) {
448 bio = dio->bio_list;
449 dio->bio_list = bio->bi_private;
450 }
1da177e4
LT
451 spin_unlock_irqrestore(&dio->bio_lock, flags);
452 return bio;
453}
454
455/*
456 * Process one completed BIO. No locks are held.
457 */
458static int dio_bio_complete(struct dio *dio, struct bio *bio)
459{
cb34e057
KO
460 struct bio_vec *bvec;
461 unsigned i;
9b81c842 462 int err;
1da177e4 463
4246a0b6 464 if (bio->bi_error)
174e27c6 465 dio->io_error = -EIO;
1da177e4
LT
466
467 if (dio->is_async && dio->rw == READ) {
468 bio_check_pages_dirty(bio); /* transfers ownership */
9b81c842 469 err = bio->bi_error;
1da177e4 470 } else {
cb34e057
KO
471 bio_for_each_segment_all(bvec, bio, i) {
472 struct page *page = bvec->bv_page;
1da177e4
LT
473
474 if (dio->rw == READ && !PageCompound(page))
475 set_page_dirty_lock(page);
476 page_cache_release(page);
477 }
9b81c842 478 err = bio->bi_error;
1da177e4
LT
479 bio_put(bio);
480 }
9b81c842 481 return err;
1da177e4
LT
482}
483
484/*
0273201e
ZB
485 * Wait on and process all in-flight BIOs. This must only be called once
486 * all bios have been issued so that the refcount can only decrease.
487 * This just waits for all bios to make it through dio_bio_complete. IO
beb7dd86 488 * errors are propagated through dio->io_error and should be propagated via
0273201e 489 * dio_complete().
1da177e4 490 */
6d544bb4 491static void dio_await_completion(struct dio *dio)
1da177e4 492{
0273201e
ZB
493 struct bio *bio;
494 do {
495 bio = dio_await_one(dio);
496 if (bio)
497 dio_bio_complete(dio, bio);
498 } while (bio);
1da177e4
LT
499}
500
501/*
502 * A really large O_DIRECT read or write can generate a lot of BIOs. So
503 * to keep the memory consumption sane we periodically reap any completed BIOs
504 * during the BIO generation phase.
505 *
506 * This also helps to limit the peak amount of pinned userspace memory.
507 */
ba253fbf 508static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
1da177e4
LT
509{
510 int ret = 0;
511
eb28be2b 512 if (sdio->reap_counter++ >= 64) {
1da177e4
LT
513 while (dio->bio_list) {
514 unsigned long flags;
515 struct bio *bio;
516 int ret2;
517
518 spin_lock_irqsave(&dio->bio_lock, flags);
519 bio = dio->bio_list;
520 dio->bio_list = bio->bi_private;
521 spin_unlock_irqrestore(&dio->bio_lock, flags);
522 ret2 = dio_bio_complete(dio, bio);
523 if (ret == 0)
524 ret = ret2;
525 }
eb28be2b 526 sdio->reap_counter = 0;
1da177e4
LT
527 }
528 return ret;
529}
530
7b7a8665
CH
531/*
532 * Create workqueue for deferred direct IO completions. We allocate the
533 * workqueue when it's first needed. This avoids creating workqueue for
534 * filesystems that don't need it and also allows us to create the workqueue
535 * late enough so the we can include s_id in the name of the workqueue.
536 */
537static int sb_init_dio_done_wq(struct super_block *sb)
538{
45150c43 539 struct workqueue_struct *old;
7b7a8665
CH
540 struct workqueue_struct *wq = alloc_workqueue("dio/%s",
541 WQ_MEM_RECLAIM, 0,
542 sb->s_id);
543 if (!wq)
544 return -ENOMEM;
545 /*
546 * This has to be atomic as more DIOs can race to create the workqueue
547 */
45150c43 548 old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
7b7a8665 549 /* Someone created workqueue before us? Free ours... */
45150c43 550 if (old)
7b7a8665
CH
551 destroy_workqueue(wq);
552 return 0;
553}
554
555static int dio_set_defer_completion(struct dio *dio)
556{
557 struct super_block *sb = dio->inode->i_sb;
558
559 if (dio->defer_completion)
560 return 0;
561 dio->defer_completion = true;
562 if (!sb->s_dio_done_wq)
563 return sb_init_dio_done_wq(sb);
564 return 0;
565}
566
1da177e4
LT
567/*
568 * Call into the fs to map some more disk blocks. We record the current number
eb28be2b 569 * of available blocks at sdio->blocks_available. These are in units of the
1da177e4
LT
570 * fs blocksize, (1 << inode->i_blkbits).
571 *
572 * The fs is allowed to map lots of blocks at once. If it wants to do that,
573 * it uses the passed inode-relative block number as the file offset, as usual.
574 *
1d8fa7a2 575 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
1da177e4
LT
576 * has remaining to do. The fs should not map more than this number of blocks.
577 *
578 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
579 * indicate how much contiguous disk space has been made available at
580 * bh->b_blocknr.
581 *
582 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
583 * This isn't very efficient...
584 *
585 * In the case of filesystem holes: the fs may return an arbitrarily-large
586 * hole by returning an appropriate value in b_size and by clearing
587 * buffer_mapped(). However the direct-io code will only process holes one
1d8fa7a2 588 * block at a time - it will repeatedly call get_block() as it walks the hole.
1da177e4 589 */
18772641
AK
590static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
591 struct buffer_head *map_bh)
1da177e4
LT
592{
593 int ret;
1da177e4 594 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
ae55e1aa 595 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
1da177e4 596 unsigned long fs_count; /* Number of filesystem-sized blocks */
1da177e4 597 int create;
ab73857e 598 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
1da177e4
LT
599
600 /*
601 * If there was a memory error and we've overwritten all the
602 * mapped blocks then we can now return that memory error
603 */
604 ret = dio->page_errors;
605 if (ret == 0) {
eb28be2b
AK
606 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
607 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
ae55e1aa
TM
608 fs_endblk = (sdio->final_block_in_request - 1) >>
609 sdio->blkfactor;
610 fs_count = fs_endblk - fs_startblk + 1;
1da177e4 611
3c674e74 612 map_bh->b_state = 0;
ab73857e 613 map_bh->b_size = fs_count << i_blkbits;
3c674e74 614
5fe878ae
CH
615 /*
616 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we
617 * forbid block creations: only overwrites are permitted.
618 * We will return early to the caller once we see an
619 * unmapped buffer head returned, and the caller will fall
620 * back to buffered I/O.
621 *
622 * Otherwise the decision is left to the get_blocks method,
623 * which may decide to handle it or also return an unmapped
624 * buffer head.
625 */
b31dc66a 626 create = dio->rw & WRITE;
5fe878ae 627 if (dio->flags & DIO_SKIP_HOLES) {
eb28be2b
AK
628 if (sdio->block_in_file < (i_size_read(dio->inode) >>
629 sdio->blkbits))
1da177e4 630 create = 0;
1da177e4 631 }
3c674e74 632
eb28be2b 633 ret = (*sdio->get_block)(dio->inode, fs_startblk,
1da177e4 634 map_bh, create);
18772641
AK
635
636 /* Store for completion */
637 dio->private = map_bh->b_private;
7b7a8665
CH
638
639 if (ret == 0 && buffer_defer_completion(map_bh))
640 ret = dio_set_defer_completion(dio);
1da177e4
LT
641 }
642 return ret;
643}
644
645/*
646 * There is no bio. Make one now.
647 */
ba253fbf
AK
648static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
649 sector_t start_sector, struct buffer_head *map_bh)
1da177e4
LT
650{
651 sector_t sector;
652 int ret, nr_pages;
653
eb28be2b 654 ret = dio_bio_reap(dio, sdio);
1da177e4
LT
655 if (ret)
656 goto out;
eb28be2b 657 sector = start_sector << (sdio->blkbits - 9);
18772641 658 nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(map_bh->b_bdev));
1da177e4 659 BUG_ON(nr_pages <= 0);
18772641 660 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
eb28be2b 661 sdio->boundary = 0;
1da177e4
LT
662out:
663 return ret;
664}
665
666/*
667 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
668 * that was successful then update final_block_in_bio and take a ref against
669 * the just-added page.
670 *
671 * Return zero on success. Non-zero means the caller needs to start a new BIO.
672 */
ba253fbf 673static inline int dio_bio_add_page(struct dio_submit *sdio)
1da177e4
LT
674{
675 int ret;
676
eb28be2b
AK
677 ret = bio_add_page(sdio->bio, sdio->cur_page,
678 sdio->cur_page_len, sdio->cur_page_offset);
679 if (ret == sdio->cur_page_len) {
1da177e4
LT
680 /*
681 * Decrement count only, if we are done with this page
682 */
eb28be2b
AK
683 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
684 sdio->pages_in_io--;
685 page_cache_get(sdio->cur_page);
686 sdio->final_block_in_bio = sdio->cur_page_block +
687 (sdio->cur_page_len >> sdio->blkbits);
1da177e4
LT
688 ret = 0;
689 } else {
690 ret = 1;
691 }
692 return ret;
693}
694
695/*
696 * Put cur_page under IO. The section of cur_page which is described by
697 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
698 * starts on-disk at cur_page_block.
699 *
700 * We take a ref against the page here (on behalf of its presence in the bio).
701 *
702 * The caller of this function is responsible for removing cur_page from the
703 * dio, and for dropping the refcount which came from that presence.
704 */
ba253fbf
AK
705static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
706 struct buffer_head *map_bh)
1da177e4
LT
707{
708 int ret = 0;
709
eb28be2b
AK
710 if (sdio->bio) {
711 loff_t cur_offset = sdio->cur_page_fs_offset;
712 loff_t bio_next_offset = sdio->logical_offset_in_bio +
4f024f37 713 sdio->bio->bi_iter.bi_size;
c2c6ca41 714
1da177e4 715 /*
c2c6ca41
JB
716 * See whether this new request is contiguous with the old.
717 *
f0940cee
NK
718 * Btrfs cannot handle having logically non-contiguous requests
719 * submitted. For example if you have
c2c6ca41
JB
720 *
721 * Logical: [0-4095][HOLE][8192-12287]
f0940cee 722 * Physical: [0-4095] [4096-8191]
c2c6ca41
JB
723 *
724 * We cannot submit those pages together as one BIO. So if our
725 * current logical offset in the file does not equal what would
726 * be the next logical offset in the bio, submit the bio we
727 * have.
1da177e4 728 */
eb28be2b 729 if (sdio->final_block_in_bio != sdio->cur_page_block ||
c2c6ca41 730 cur_offset != bio_next_offset)
eb28be2b 731 dio_bio_submit(dio, sdio);
1da177e4
LT
732 }
733
eb28be2b 734 if (sdio->bio == NULL) {
18772641 735 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
1da177e4
LT
736 if (ret)
737 goto out;
738 }
739
eb28be2b
AK
740 if (dio_bio_add_page(sdio) != 0) {
741 dio_bio_submit(dio, sdio);
18772641 742 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
1da177e4 743 if (ret == 0) {
eb28be2b 744 ret = dio_bio_add_page(sdio);
1da177e4
LT
745 BUG_ON(ret != 0);
746 }
747 }
748out:
749 return ret;
750}
751
752/*
753 * An autonomous function to put a chunk of a page under deferred IO.
754 *
755 * The caller doesn't actually know (or care) whether this piece of page is in
756 * a BIO, or is under IO or whatever. We just take care of all possible
757 * situations here. The separation between the logic of do_direct_IO() and
758 * that of submit_page_section() is important for clarity. Please don't break.
759 *
760 * The chunk of page starts on-disk at blocknr.
761 *
762 * We perform deferred IO, by recording the last-submitted page inside our
763 * private part of the dio structure. If possible, we just expand the IO
764 * across that page here.
765 *
766 * If that doesn't work out then we put the old page into the bio and add this
767 * page to the dio instead.
768 */
ba253fbf 769static inline int
eb28be2b 770submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
18772641
AK
771 unsigned offset, unsigned len, sector_t blocknr,
772 struct buffer_head *map_bh)
1da177e4
LT
773{
774 int ret = 0;
775
98c4d57d
AM
776 if (dio->rw & WRITE) {
777 /*
778 * Read accounting is performed in submit_bio()
779 */
780 task_io_account_write(len);
781 }
782
1da177e4
LT
783 /*
784 * Can we just grow the current page's presence in the dio?
785 */
eb28be2b
AK
786 if (sdio->cur_page == page &&
787 sdio->cur_page_offset + sdio->cur_page_len == offset &&
788 sdio->cur_page_block +
789 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
790 sdio->cur_page_len += len;
1da177e4
LT
791 goto out;
792 }
793
794 /*
795 * If there's a deferred page already there then send it.
796 */
eb28be2b 797 if (sdio->cur_page) {
18772641 798 ret = dio_send_cur_page(dio, sdio, map_bh);
eb28be2b
AK
799 page_cache_release(sdio->cur_page);
800 sdio->cur_page = NULL;
1da177e4 801 if (ret)
b1058b98 802 return ret;
1da177e4
LT
803 }
804
805 page_cache_get(page); /* It is in dio */
eb28be2b
AK
806 sdio->cur_page = page;
807 sdio->cur_page_offset = offset;
808 sdio->cur_page_len = len;
809 sdio->cur_page_block = blocknr;
810 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
1da177e4 811out:
b1058b98
JK
812 /*
813 * If sdio->boundary then we want to schedule the IO now to
814 * avoid metadata seeks.
815 */
816 if (sdio->boundary) {
817 ret = dio_send_cur_page(dio, sdio, map_bh);
818 dio_bio_submit(dio, sdio);
819 page_cache_release(sdio->cur_page);
820 sdio->cur_page = NULL;
821 }
1da177e4
LT
822 return ret;
823}
824
825/*
826 * Clean any dirty buffers in the blockdev mapping which alias newly-created
827 * file blocks. Only called for S_ISREG files - blockdevs do not set
828 * buffer_new
829 */
18772641 830static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
1da177e4
LT
831{
832 unsigned i;
833 unsigned nblocks;
834
18772641 835 nblocks = map_bh->b_size >> dio->inode->i_blkbits;
1da177e4
LT
836
837 for (i = 0; i < nblocks; i++) {
18772641
AK
838 unmap_underlying_metadata(map_bh->b_bdev,
839 map_bh->b_blocknr + i);
1da177e4
LT
840 }
841}
842
843/*
844 * If we are not writing the entire block and get_block() allocated
845 * the block for us, we need to fill-in the unused portion of the
846 * block with zeros. This happens only if user-buffer, fileoffset or
847 * io length is not filesystem block-size multiple.
848 *
849 * `end' is zero if we're doing the start of the IO, 1 at the end of the
850 * IO.
851 */
ba253fbf
AK
852static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
853 int end, struct buffer_head *map_bh)
1da177e4
LT
854{
855 unsigned dio_blocks_per_fs_block;
856 unsigned this_chunk_blocks; /* In dio_blocks */
857 unsigned this_chunk_bytes;
858 struct page *page;
859
eb28be2b 860 sdio->start_zero_done = 1;
18772641 861 if (!sdio->blkfactor || !buffer_new(map_bh))
1da177e4
LT
862 return;
863
eb28be2b
AK
864 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
865 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
1da177e4
LT
866
867 if (!this_chunk_blocks)
868 return;
869
870 /*
871 * We need to zero out part of an fs block. It is either at the
872 * beginning or the end of the fs block.
873 */
874 if (end)
875 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
876
eb28be2b 877 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
1da177e4 878
557ed1fa 879 page = ZERO_PAGE(0);
eb28be2b 880 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
18772641 881 sdio->next_block_for_io, map_bh))
1da177e4
LT
882 return;
883
eb28be2b 884 sdio->next_block_for_io += this_chunk_blocks;
1da177e4
LT
885}
886
887/*
888 * Walk the user pages, and the file, mapping blocks to disk and generating
889 * a sequence of (page,offset,len,block) mappings. These mappings are injected
890 * into submit_page_section(), which takes care of the next stage of submission
891 *
892 * Direct IO against a blockdev is different from a file. Because we can
893 * happily perform page-sized but 512-byte aligned IOs. It is important that
894 * blockdev IO be able to have fine alignment and large sizes.
895 *
1d8fa7a2 896 * So what we do is to permit the ->get_block function to populate bh.b_size
1da177e4
LT
897 * with the size of IO which is permitted at this offset and this i_blkbits.
898 *
899 * For best results, the blockdev should be set up with 512-byte i_blkbits and
1d8fa7a2 900 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
1da177e4
LT
901 * fine alignment but still allows this function to work in PAGE_SIZE units.
902 */
18772641
AK
903static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
904 struct buffer_head *map_bh)
1da177e4 905{
eb28be2b 906 const unsigned blkbits = sdio->blkbits;
1da177e4
LT
907 int ret = 0;
908
eb28be2b 909 while (sdio->block_in_file < sdio->final_block_in_request) {
7b2c99d1
AV
910 struct page *page;
911 size_t from, to;
6fcc5420
BH
912
913 page = dio_get_page(dio, sdio);
1da177e4
LT
914 if (IS_ERR(page)) {
915 ret = PTR_ERR(page);
916 goto out;
917 }
6fcc5420
BH
918 from = sdio->head ? 0 : sdio->from;
919 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
920 sdio->head++;
1da177e4 921
7b2c99d1 922 while (from < to) {
1da177e4
LT
923 unsigned this_chunk_bytes; /* # of bytes mapped */
924 unsigned this_chunk_blocks; /* # of blocks */
925 unsigned u;
926
eb28be2b 927 if (sdio->blocks_available == 0) {
1da177e4
LT
928 /*
929 * Need to go and map some more disk
930 */
931 unsigned long blkmask;
932 unsigned long dio_remainder;
933
18772641 934 ret = get_more_blocks(dio, sdio, map_bh);
1da177e4
LT
935 if (ret) {
936 page_cache_release(page);
937 goto out;
938 }
939 if (!buffer_mapped(map_bh))
940 goto do_holes;
941
eb28be2b
AK
942 sdio->blocks_available =
943 map_bh->b_size >> sdio->blkbits;
944 sdio->next_block_for_io =
945 map_bh->b_blocknr << sdio->blkfactor;
1da177e4 946 if (buffer_new(map_bh))
18772641 947 clean_blockdev_aliases(dio, map_bh);
1da177e4 948
eb28be2b 949 if (!sdio->blkfactor)
1da177e4
LT
950 goto do_holes;
951
eb28be2b
AK
952 blkmask = (1 << sdio->blkfactor) - 1;
953 dio_remainder = (sdio->block_in_file & blkmask);
1da177e4
LT
954
955 /*
956 * If we are at the start of IO and that IO
957 * starts partway into a fs-block,
958 * dio_remainder will be non-zero. If the IO
959 * is a read then we can simply advance the IO
960 * cursor to the first block which is to be
961 * read. But if the IO is a write and the
962 * block was newly allocated we cannot do that;
963 * the start of the fs block must be zeroed out
964 * on-disk
965 */
966 if (!buffer_new(map_bh))
eb28be2b
AK
967 sdio->next_block_for_io += dio_remainder;
968 sdio->blocks_available -= dio_remainder;
1da177e4
LT
969 }
970do_holes:
971 /* Handle holes */
972 if (!buffer_mapped(map_bh)) {
35dc8161 973 loff_t i_size_aligned;
1da177e4
LT
974
975 /* AKPM: eargh, -ENOTBLK is a hack */
b31dc66a 976 if (dio->rw & WRITE) {
1da177e4
LT
977 page_cache_release(page);
978 return -ENOTBLK;
979 }
980
35dc8161
JM
981 /*
982 * Be sure to account for a partial block as the
983 * last block in the file
984 */
985 i_size_aligned = ALIGN(i_size_read(dio->inode),
986 1 << blkbits);
eb28be2b 987 if (sdio->block_in_file >=
35dc8161 988 i_size_aligned >> blkbits) {
1da177e4
LT
989 /* We hit eof */
990 page_cache_release(page);
991 goto out;
992 }
7b2c99d1 993 zero_user(page, from, 1 << blkbits);
eb28be2b 994 sdio->block_in_file++;
7b2c99d1 995 from += 1 << blkbits;
3320c60b 996 dio->result += 1 << blkbits;
1da177e4
LT
997 goto next_block;
998 }
999
1000 /*
1001 * If we're performing IO which has an alignment which
1002 * is finer than the underlying fs, go check to see if
1003 * we must zero out the start of this block.
1004 */
eb28be2b 1005 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
18772641 1006 dio_zero_block(dio, sdio, 0, map_bh);
1da177e4
LT
1007
1008 /*
1009 * Work out, in this_chunk_blocks, how much disk we
1010 * can add to this page
1011 */
eb28be2b 1012 this_chunk_blocks = sdio->blocks_available;
7b2c99d1 1013 u = (to - from) >> blkbits;
1da177e4
LT
1014 if (this_chunk_blocks > u)
1015 this_chunk_blocks = u;
eb28be2b 1016 u = sdio->final_block_in_request - sdio->block_in_file;
1da177e4
LT
1017 if (this_chunk_blocks > u)
1018 this_chunk_blocks = u;
1019 this_chunk_bytes = this_chunk_blocks << blkbits;
1020 BUG_ON(this_chunk_bytes == 0);
1021
092c8d46
JK
1022 if (this_chunk_blocks == sdio->blocks_available)
1023 sdio->boundary = buffer_boundary(map_bh);
eb28be2b 1024 ret = submit_page_section(dio, sdio, page,
7b2c99d1 1025 from,
eb28be2b 1026 this_chunk_bytes,
18772641
AK
1027 sdio->next_block_for_io,
1028 map_bh);
1da177e4
LT
1029 if (ret) {
1030 page_cache_release(page);
1031 goto out;
1032 }
eb28be2b 1033 sdio->next_block_for_io += this_chunk_blocks;
1da177e4 1034
eb28be2b 1035 sdio->block_in_file += this_chunk_blocks;
7b2c99d1
AV
1036 from += this_chunk_bytes;
1037 dio->result += this_chunk_bytes;
eb28be2b 1038 sdio->blocks_available -= this_chunk_blocks;
1da177e4 1039next_block:
eb28be2b
AK
1040 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1041 if (sdio->block_in_file == sdio->final_block_in_request)
1da177e4
LT
1042 break;
1043 }
1044
1045 /* Drop the ref which was taken in get_user_pages() */
1046 page_cache_release(page);
1da177e4
LT
1047 }
1048out:
1049 return ret;
1050}
1051
847cc637 1052static inline int drop_refcount(struct dio *dio)
1da177e4 1053{
847cc637 1054 int ret2;
5eb6c7a2 1055 unsigned long flags;
1da177e4 1056
8459d86a
ZB
1057 /*
1058 * Sync will always be dropping the final ref and completing the
5eb6c7a2
ZB
1059 * operation. AIO can if it was a broken operation described above or
1060 * in fact if all the bios race to complete before we get here. In
1061 * that case dio_complete() translates the EIOCBQUEUED into the proper
04b2fa9f 1062 * return code that the caller will hand to ->complete().
5eb6c7a2
ZB
1063 *
1064 * This is managed by the bio_lock instead of being an atomic_t so that
1065 * completion paths can drop their ref and use the remaining count to
1066 * decide to wake the submission path atomically.
8459d86a 1067 */
5eb6c7a2
ZB
1068 spin_lock_irqsave(&dio->bio_lock, flags);
1069 ret2 = --dio->refcount;
1070 spin_unlock_irqrestore(&dio->bio_lock, flags);
847cc637 1071 return ret2;
1da177e4
LT
1072}
1073
eafdc7d1
CH
1074/*
1075 * This is a library function for use by filesystem drivers.
1076 *
1077 * The locking rules are governed by the flags parameter:
1078 * - if the flags value contains DIO_LOCKING we use a fancy locking
1079 * scheme for dumb filesystems.
1080 * For writes this function is called under i_mutex and returns with
1081 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1082 * taken and dropped again before returning.
eafdc7d1
CH
1083 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1084 * internal locking but rather rely on the filesystem to synchronize
1085 * direct I/O reads/writes versus each other and truncate.
df2d6f26
CH
1086 *
1087 * To help with locking against truncate we incremented the i_dio_count
1088 * counter before starting direct I/O, and decrement it once we are done.
1089 * Truncate can wait for it to reach zero to provide exclusion. It is
1090 * expected that filesystem provide exclusion between new direct I/O
1091 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1092 * but other filesystems need to take care of this on their own.
ba253fbf
AK
1093 *
1094 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1095 * is always inlined. Otherwise gcc is unable to split the structure into
1096 * individual fields and will generate much worse code. This is important
1097 * for the whole file.
eafdc7d1 1098 */
65dd2aa9 1099static inline ssize_t
17f8c842
OS
1100do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1101 struct block_device *bdev, struct iov_iter *iter,
1102 loff_t offset, get_block_t get_block, dio_iodone_t end_io,
1103 dio_submit_t submit_io, int flags)
1da177e4 1104{
ab73857e
LT
1105 unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits);
1106 unsigned blkbits = i_blkbits;
1da177e4
LT
1107 unsigned blocksize_mask = (1 << blkbits) - 1;
1108 ssize_t retval = -EINVAL;
af436472
CH
1109 size_t count = iov_iter_count(iter);
1110 loff_t end = offset + count;
1da177e4 1111 struct dio *dio;
eb28be2b 1112 struct dio_submit sdio = { 0, };
847cc637 1113 struct buffer_head map_bh = { 0, };
647d1e4c 1114 struct blk_plug plug;
886a3911 1115 unsigned long align = offset | iov_iter_alignment(iter);
1da177e4 1116
65dd2aa9
AK
1117 /*
1118 * Avoid references to bdev if not absolutely needed to give
1119 * the early prefetch in the caller enough time.
1120 */
1da177e4 1121
886a3911 1122 if (align & blocksize_mask) {
1da177e4 1123 if (bdev)
65dd2aa9 1124 blkbits = blksize_bits(bdev_logical_block_size(bdev));
1da177e4 1125 blocksize_mask = (1 << blkbits) - 1;
886a3911 1126 if (align & blocksize_mask)
1da177e4
LT
1127 goto out;
1128 }
1129
f9b5570d 1130 /* watch out for a 0 len io from a tricksy fs */
17f8c842 1131 if (iov_iter_rw(iter) == READ && !iov_iter_count(iter))
f9b5570d
CH
1132 return 0;
1133
6e8267f5 1134 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
1da177e4
LT
1135 retval = -ENOMEM;
1136 if (!dio)
1137 goto out;
23aee091
JM
1138 /*
1139 * Believe it or not, zeroing out the page array caused a .5%
1140 * performance regression in a database benchmark. So, we take
1141 * care to only zero out what's needed.
1142 */
1143 memset(dio, 0, offsetof(struct dio, pages));
1da177e4 1144
5fe878ae
CH
1145 dio->flags = flags;
1146 if (dio->flags & DIO_LOCKING) {
17f8c842 1147 if (iov_iter_rw(iter) == READ) {
5fe878ae
CH
1148 struct address_space *mapping =
1149 iocb->ki_filp->f_mapping;
1da177e4 1150
5fe878ae
CH
1151 /* will be released by direct_io_worker */
1152 mutex_lock(&inode->i_mutex);
1da177e4
LT
1153
1154 retval = filemap_write_and_wait_range(mapping, offset,
1155 end - 1);
1156 if (retval) {
5fe878ae 1157 mutex_unlock(&inode->i_mutex);
6e8267f5 1158 kmem_cache_free(dio_cache, dio);
1da177e4
LT
1159 goto out;
1160 }
1da177e4 1161 }
1da177e4
LT
1162 }
1163
1164 /*
60392573
CH
1165 * For file extending writes updating i_size before data writeouts
1166 * complete can expose uninitialized blocks in dumb filesystems.
1167 * In that case we need to wait for I/O completion even if asked
1168 * for an asynchronous write.
1da177e4 1169 */
60392573
CH
1170 if (is_sync_kiocb(iocb))
1171 dio->is_async = false;
1172 else if (!(dio->flags & DIO_ASYNC_EXTEND) &&
17f8c842 1173 iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
60392573
CH
1174 dio->is_async = false;
1175 else
1176 dio->is_async = true;
1177
847cc637 1178 dio->inode = inode;
17f8c842 1179 dio->rw = iov_iter_rw(iter) == WRITE ? WRITE_ODIRECT : READ;
02afc27f
CH
1180
1181 /*
1182 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1183 * so that we can call ->fsync.
1184 */
17f8c842 1185 if (dio->is_async && iov_iter_rw(iter) == WRITE &&
02afc27f
CH
1186 ((iocb->ki_filp->f_flags & O_DSYNC) ||
1187 IS_SYNC(iocb->ki_filp->f_mapping->host))) {
1188 retval = dio_set_defer_completion(dio);
1189 if (retval) {
1190 /*
1191 * We grab i_mutex only for reads so we don't have
1192 * to release it here
1193 */
1194 kmem_cache_free(dio_cache, dio);
1195 goto out;
1196 }
1197 }
1198
1199 /*
1200 * Will be decremented at I/O completion time.
1201 */
fe0f07d0
JA
1202 if (!(dio->flags & DIO_SKIP_DIO_COUNT))
1203 inode_dio_begin(inode);
02afc27f
CH
1204
1205 retval = 0;
847cc637 1206 sdio.blkbits = blkbits;
ab73857e 1207 sdio.blkfactor = i_blkbits - blkbits;
847cc637
AK
1208 sdio.block_in_file = offset >> blkbits;
1209
1210 sdio.get_block = get_block;
1211 dio->end_io = end_io;
1212 sdio.submit_io = submit_io;
1213 sdio.final_block_in_bio = -1;
1214 sdio.next_block_for_io = -1;
1215
1216 dio->iocb = iocb;
1217 dio->i_size = i_size_read(inode);
1218
1219 spin_lock_init(&dio->bio_lock);
1220 dio->refcount = 1;
1221
7b2c99d1
AV
1222 sdio.iter = iter;
1223 sdio.final_block_in_request =
1224 (offset + iov_iter_count(iter)) >> blkbits;
1225
847cc637
AK
1226 /*
1227 * In case of non-aligned buffers, we may need 2 more
1228 * pages since we need to zero out first and last block.
1229 */
1230 if (unlikely(sdio.blkfactor))
1231 sdio.pages_in_io = 2;
1232
f67da30c 1233 sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
847cc637 1234
647d1e4c
FW
1235 blk_start_plug(&plug);
1236
7b2c99d1
AV
1237 retval = do_direct_IO(dio, &sdio, &map_bh);
1238 if (retval)
1239 dio_cleanup(dio, &sdio);
847cc637
AK
1240
1241 if (retval == -ENOTBLK) {
1242 /*
1243 * The remaining part of the request will be
1244 * be handled by buffered I/O when we return
1245 */
1246 retval = 0;
1247 }
1248 /*
1249 * There may be some unwritten disk at the end of a part-written
1250 * fs-block-sized block. Go zero that now.
1251 */
1252 dio_zero_block(dio, &sdio, 1, &map_bh);
1253
1254 if (sdio.cur_page) {
1255 ssize_t ret2;
1256
1257 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1258 if (retval == 0)
1259 retval = ret2;
1260 page_cache_release(sdio.cur_page);
1261 sdio.cur_page = NULL;
1262 }
1263 if (sdio.bio)
1264 dio_bio_submit(dio, &sdio);
1265
647d1e4c
FW
1266 blk_finish_plug(&plug);
1267
847cc637
AK
1268 /*
1269 * It is possible that, we return short IO due to end of file.
1270 * In that case, we need to release all the pages we got hold on.
1271 */
1272 dio_cleanup(dio, &sdio);
1273
1274 /*
1275 * All block lookups have been performed. For READ requests
1276 * we can let i_mutex go now that its achieved its purpose
1277 * of protecting us from looking up uninitialized blocks.
1278 */
17f8c842 1279 if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING))
847cc637
AK
1280 mutex_unlock(&dio->inode->i_mutex);
1281
1282 /*
1283 * The only time we want to leave bios in flight is when a successful
1284 * partial aio read or full aio write have been setup. In that case
1285 * bio completion will call aio_complete. The only time it's safe to
1286 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1287 * This had *better* be the only place that raises -EIOCBQUEUED.
1288 */
1289 BUG_ON(retval == -EIOCBQUEUED);
1290 if (dio->is_async && retval == 0 && dio->result &&
17f8c842 1291 (iov_iter_rw(iter) == READ || dio->result == count))
847cc637 1292 retval = -EIOCBQUEUED;
af436472 1293 else
847cc637
AK
1294 dio_await_completion(dio);
1295
1296 if (drop_refcount(dio) == 0) {
1297 retval = dio_complete(dio, offset, retval, false);
847cc637
AK
1298 } else
1299 BUG_ON(retval != -EIOCBQUEUED);
1da177e4 1300
7bb46a67 1301out:
1302 return retval;
1303}
65dd2aa9 1304
17f8c842
OS
1305ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1306 struct block_device *bdev, struct iov_iter *iter,
1307 loff_t offset, get_block_t get_block,
1308 dio_iodone_t end_io, dio_submit_t submit_io,
1309 int flags)
65dd2aa9
AK
1310{
1311 /*
1312 * The block device state is needed in the end to finally
1313 * submit everything. Since it's likely to be cache cold
1314 * prefetch it here as first thing to hide some of the
1315 * latency.
1316 *
1317 * Attempt to prefetch the pieces we likely need later.
1318 */
1319 prefetch(&bdev->bd_disk->part_tbl);
1320 prefetch(bdev->bd_queue);
1321 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1322
17f8c842
OS
1323 return do_blockdev_direct_IO(iocb, inode, bdev, iter, offset, get_block,
1324 end_io, submit_io, flags);
65dd2aa9
AK
1325}
1326
1da177e4 1327EXPORT_SYMBOL(__blockdev_direct_IO);
6e8267f5
AK
1328
1329static __init int dio_init(void)
1330{
1331 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1332 return 0;
1333}
1334module_init(dio_init)