Merge tag 'locking-core-2023-05-05' of git://git.kernel.org/pub/scm/linux/kernel...
[linux-block.git] / fs / direct-io.c
CommitLineData
457c8996 1// SPDX-License-Identifier: GPL-2.0-only
1da177e4
LT
2/*
3 * fs/direct-io.c
4 *
5 * Copyright (C) 2002, Linus Torvalds.
6 *
7 * O_DIRECT
8 *
e1f8e874 9 * 04Jul2002 Andrew Morton
1da177e4
LT
10 * Initial version
11 * 11Sep2002 janetinc@us.ibm.com
12 * added readv/writev support.
e1f8e874 13 * 29Oct2002 Andrew Morton
1da177e4
LT
14 * rewrote bio_add_page() support.
15 * 30Oct2002 pbadari@us.ibm.com
16 * added support for non-aligned IO.
17 * 06Nov2002 pbadari@us.ibm.com
18 * added asynchronous IO support.
19 * 21Jul2003 nathans@sgi.com
20 * added IO completion notifier.
21 */
22
23#include <linux/kernel.h>
24#include <linux/module.h>
25#include <linux/types.h>
26#include <linux/fs.h>
27#include <linux/mm.h>
28#include <linux/slab.h>
29#include <linux/highmem.h>
30#include <linux/pagemap.h>
98c4d57d 31#include <linux/task_io_accounting_ops.h>
1da177e4
LT
32#include <linux/bio.h>
33#include <linux/wait.h>
34#include <linux/err.h>
35#include <linux/blkdev.h>
36#include <linux/buffer_head.h>
37#include <linux/rwsem.h>
38#include <linux/uio.h>
60063497 39#include <linux/atomic.h>
65dd2aa9 40#include <linux/prefetch.h>
1da177e4 41
b16155a0
EB
42#include "internal.h"
43
1da177e4
LT
44/*
45 * How many user pages to map in one call to get_user_pages(). This determines
cde1ecb3 46 * the size of a structure in the slab cache
1da177e4
LT
47 */
48#define DIO_PAGES 64
49
ffe51f01
LC
50/*
51 * Flags for dio_complete()
52 */
53#define DIO_COMPLETE_ASYNC 0x01 /* This is async IO */
54#define DIO_COMPLETE_INVALIDATE 0x02 /* Can invalidate pages */
55
1da177e4
LT
56/*
57 * This code generally works in units of "dio_blocks". A dio_block is
58 * somewhere between the hard sector size and the filesystem block size. it
59 * is determined on a per-invocation basis. When talking to the filesystem
60 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
61 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
62 * to bio_block quantities by shifting left by blkfactor.
63 *
64 * If blkfactor is zero then the user's request was aligned to the filesystem's
65 * blocksize.
1da177e4
LT
66 */
67
eb28be2b
AK
68/* dio_state only used in the submission path */
69
70struct dio_submit {
1da177e4 71 struct bio *bio; /* bio under assembly */
1da177e4
LT
72 unsigned blkbits; /* doesn't change */
73 unsigned blkfactor; /* When we're using an alignment which
74 is finer than the filesystem's soft
75 blocksize, this specifies how much
76 finer. blkfactor=2 means 1/4-block
77 alignment. Does not change */
78 unsigned start_zero_done; /* flag: sub-blocksize zeroing has
79 been performed at the start of a
80 write */
81 int pages_in_io; /* approximate total IO pages */
1da177e4
LT
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 */
0dc2bc49 85 int reap_counter; /* rate limit reaping */
1da177e4 86 sector_t final_block_in_request;/* doesn't change */
1da177e4 87 int boundary; /* prev block is at a boundary */
1d8fa7a2 88 get_block_t *get_block; /* block mapping function */
eb28be2b 89
facd07b0 90 loff_t logical_offset_in_bio; /* current first logical block in bio */
1da177e4
LT
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 */
1da177e4
LT
94
95 /*
96 * Deferred addition of a page to the dio. These variables are
97 * private to dio_send_cur_page(), submit_page_section() and
98 * dio_bio_add_page().
99 */
100 struct page *cur_page; /* The page */
101 unsigned cur_page_offset; /* Offset into it, in bytes */
102 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
103 sector_t cur_page_block; /* Where it starts */
facd07b0 104 loff_t cur_page_fs_offset; /* Offset in file */
1da177e4 105
7b2c99d1 106 struct iov_iter *iter;
1da177e4
LT
107 /*
108 * Page queue. These variables belong to dio_refill_pages() and
109 * dio_get_page().
110 */
1da177e4
LT
111 unsigned head; /* next page to process */
112 unsigned tail; /* last valid page + 1 */
7b2c99d1 113 size_t from, to;
eb28be2b
AK
114};
115
116/* dio_state communicated between submission path and end_io */
117struct dio {
118 int flags; /* doesn't change */
c6293eac 119 blk_opf_t opf; /* request operation type and flags */
74d46992 120 struct gendisk *bio_disk;
0dc2bc49 121 struct inode *inode;
eb28be2b
AK
122 loff_t i_size; /* i_size when submitted */
123 dio_iodone_t *end_io; /* IO completion function */
eb28be2b 124
18772641 125 void *private; /* copy from map_bh.b_private */
eb28be2b
AK
126
127 /* BIO completion state */
128 spinlock_t bio_lock; /* protects BIO fields below */
0dc2bc49
AK
129 int page_errors; /* errno from get_user_pages() */
130 int is_async; /* is IO async ? */
7b7a8665 131 bool defer_completion; /* defer AIO completion to workqueue? */
53cbf3b1 132 bool should_dirty; /* if pages should be dirtied */
0dc2bc49 133 int io_error; /* IO error in completion path */
eb28be2b
AK
134 unsigned long refcount; /* direct_io_worker() and bios */
135 struct bio *bio_list; /* singly linked via bi_private */
136 struct task_struct *waiter; /* waiting task (NULL if none) */
137
138 /* AIO related stuff */
139 struct kiocb *iocb; /* kiocb */
eb28be2b
AK
140 ssize_t result; /* IO result */
141
23aee091
JM
142 /*
143 * pages[] (and any fields placed after it) are not zeroed out at
144 * allocation time. Don't add new fields after pages[] unless you
145 * wish that they not be zeroed.
146 */
7b7a8665
CH
147 union {
148 struct page *pages[DIO_PAGES]; /* page buffer */
149 struct work_struct complete_work;/* deferred AIO completion */
150 };
6e8267f5
AK
151} ____cacheline_aligned_in_smp;
152
153static struct kmem_cache *dio_cache __read_mostly;
1da177e4
LT
154
155/*
156 * How many pages are in the queue?
157 */
eb28be2b 158static inline unsigned dio_pages_present(struct dio_submit *sdio)
1da177e4 159{
eb28be2b 160 return sdio->tail - sdio->head;
1da177e4
LT
161}
162
163/*
164 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
165 */
ba253fbf 166static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
1da177e4 167{
c6293eac 168 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
7b2c99d1 169 ssize_t ret;
1da177e4 170
1ef255e2 171 ret = iov_iter_get_pages2(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES,
7b2c99d1 172 &sdio->from);
1da177e4 173
c6293eac 174 if (ret < 0 && sdio->blocks_available && dio_op == REQ_OP_WRITE) {
557ed1fa 175 struct page *page = ZERO_PAGE(0);
1da177e4
LT
176 /*
177 * A memory fault, but the filesystem has some outstanding
178 * mapped blocks. We need to use those blocks up to avoid
179 * leaking stale data in the file.
180 */
181 if (dio->page_errors == 0)
182 dio->page_errors = ret;
09cbfeaf 183 get_page(page);
b5810039 184 dio->pages[0] = page;
eb28be2b
AK
185 sdio->head = 0;
186 sdio->tail = 1;
7b2c99d1
AV
187 sdio->from = 0;
188 sdio->to = PAGE_SIZE;
189 return 0;
1da177e4
LT
190 }
191
192 if (ret >= 0) {
7b2c99d1 193 ret += sdio->from;
eb28be2b 194 sdio->head = 0;
7b2c99d1
AV
195 sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
196 sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1;
197 return 0;
1da177e4 198 }
1da177e4
LT
199 return ret;
200}
201
202/*
203 * Get another userspace page. Returns an ERR_PTR on error. Pages are
204 * buffered inside the dio so that we can call get_user_pages() against a
205 * decent number of pages, less frequently. To provide nicer use of the
206 * L1 cache.
207 */
ba253fbf 208static inline struct page *dio_get_page(struct dio *dio,
6fcc5420 209 struct dio_submit *sdio)
1da177e4 210{
eb28be2b 211 if (dio_pages_present(sdio) == 0) {
1da177e4
LT
212 int ret;
213
eb28be2b 214 ret = dio_refill_pages(dio, sdio);
1da177e4
LT
215 if (ret)
216 return ERR_PTR(ret);
eb28be2b 217 BUG_ON(dio_pages_present(sdio) == 0);
1da177e4 218 }
6fcc5420 219 return dio->pages[sdio->head];
1da177e4
LT
220}
221
c70d868f 222/*
6d544bb4 223 * dio_complete() - called when all DIO BIO I/O has been completed
6d544bb4 224 *
7b7a8665
CH
225 * This drops i_dio_count, lets interested parties know that a DIO operation
226 * has completed, and calculates the resulting return code for the operation.
6d544bb4
ZB
227 *
228 * It lets the filesystem know if it registered an interest earlier via
229 * get_block. Pass the private field of the map buffer_head so that
230 * filesystems can use it to hold additional state between get_block calls and
231 * dio_complete.
1da177e4 232 */
ffe51f01 233static ssize_t dio_complete(struct dio *dio, ssize_t ret, unsigned int flags)
1da177e4 234{
c6293eac 235 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
716b9bc0 236 loff_t offset = dio->iocb->ki_pos;
6d544bb4 237 ssize_t transferred = 0;
332391a9 238 int err;
6d544bb4 239
8459d86a
ZB
240 /*
241 * AIO submission can race with bio completion to get here while
242 * expecting to have the last io completed by bio completion.
243 * In that case -EIOCBQUEUED is in fact not an error we want
244 * to preserve through this call.
245 */
246 if (ret == -EIOCBQUEUED)
247 ret = 0;
248
6d544bb4
ZB
249 if (dio->result) {
250 transferred = dio->result;
251
252 /* Check for short read case */
c6293eac 253 if (dio_op == REQ_OP_READ &&
8a4c1e42 254 ((offset + transferred) > dio->i_size))
6d544bb4 255 transferred = dio->i_size - offset;
4038acdb
AV
256 /* ignore EFAULT if some IO has been done */
257 if (unlikely(ret == -EFAULT) && transferred)
258 ret = 0;
6d544bb4
ZB
259 }
260
6d544bb4
ZB
261 if (ret == 0)
262 ret = dio->page_errors;
263 if (ret == 0)
264 ret = dio->io_error;
265 if (ret == 0)
266 ret = transferred;
267
5e25c269
EG
268 if (dio->end_io) {
269 // XXX: ki_pos??
270 err = dio->end_io(dio->iocb, offset, ret, dio->private);
271 if (err)
272 ret = err;
273 }
274
332391a9
LC
275 /*
276 * Try again to invalidate clean pages which might have been cached by
277 * non-direct readahead, or faulted in by get_user_pages() if the source
278 * of the write was an mmap'ed region of the file we're writing. Either
279 * one is a pretty crazy thing to do, so we don't support it 100%. If
280 * this invalidation fails, tough, the write still worked...
5e25c269
EG
281 *
282 * And this page cache invalidation has to be after dio->end_io(), as
283 * some filesystems convert unwritten extents to real allocations in
284 * end_io() when necessary, otherwise a racing buffer read would cache
285 * zeros from unwritten extents.
332391a9 286 */
ffe51f01 287 if (flags & DIO_COMPLETE_INVALIDATE &&
c6293eac 288 ret > 0 && dio_op == REQ_OP_WRITE &&
332391a9
LC
289 dio->inode->i_mapping->nrpages) {
290 err = invalidate_inode_pages2_range(dio->inode->i_mapping,
291 offset >> PAGE_SHIFT,
292 (offset + ret - 1) >> PAGE_SHIFT);
5a9d929d
DW
293 if (err)
294 dio_warn_stale_pagecache(dio->iocb->ki_filp);
332391a9
LC
295 }
296
ce3077ee 297 inode_dio_end(dio->inode);
fe0f07d0 298
ffe51f01 299 if (flags & DIO_COMPLETE_ASYNC) {
e2592217
CH
300 /*
301 * generic_write_sync expects ki_pos to have been updated
302 * already, but the submission path only does this for
303 * synchronous I/O.
304 */
305 dio->iocb->ki_pos += transferred;
02afc27f 306
c6293eac 307 if (ret > 0 && dio_op == REQ_OP_WRITE)
41e817bc 308 ret = generic_write_sync(dio->iocb, ret);
6b19b766 309 dio->iocb->ki_complete(dio->iocb, ret);
02afc27f 310 }
40e2e973 311
7b7a8665 312 kmem_cache_free(dio_cache, dio);
6d544bb4 313 return ret;
1da177e4
LT
314}
315
7b7a8665
CH
316static void dio_aio_complete_work(struct work_struct *work)
317{
318 struct dio *dio = container_of(work, struct dio, complete_work);
319
ffe51f01 320 dio_complete(dio, 0, DIO_COMPLETE_ASYNC | DIO_COMPLETE_INVALIDATE);
7b7a8665
CH
321}
322
4e4cbee9 323static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio);
7b7a8665 324
1da177e4
LT
325/*
326 * Asynchronous IO callback.
327 */
4246a0b6 328static void dio_bio_end_aio(struct bio *bio)
1da177e4
LT
329{
330 struct dio *dio = bio->bi_private;
c6293eac 331 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
5eb6c7a2
ZB
332 unsigned long remaining;
333 unsigned long flags;
332391a9 334 bool defer_completion = false;
1da177e4 335
1da177e4
LT
336 /* cleanup the bio */
337 dio_bio_complete(dio, bio);
0273201e 338
5eb6c7a2
ZB
339 spin_lock_irqsave(&dio->bio_lock, flags);
340 remaining = --dio->refcount;
341 if (remaining == 1 && dio->waiter)
20258b2b 342 wake_up_process(dio->waiter);
5eb6c7a2 343 spin_unlock_irqrestore(&dio->bio_lock, flags);
20258b2b 344
8459d86a 345 if (remaining == 0) {
332391a9
LC
346 /*
347 * Defer completion when defer_completion is set or
348 * when the inode has pages mapped and this is AIO write.
349 * We need to invalidate those pages because there is a
350 * chance they contain stale data in the case buffered IO
351 * went in between AIO submission and completion into the
352 * same region.
353 */
354 if (dio->result)
355 defer_completion = dio->defer_completion ||
c6293eac 356 (dio_op == REQ_OP_WRITE &&
332391a9
LC
357 dio->inode->i_mapping->nrpages);
358 if (defer_completion) {
7b7a8665
CH
359 INIT_WORK(&dio->complete_work, dio_aio_complete_work);
360 queue_work(dio->inode->i_sb->s_dio_done_wq,
361 &dio->complete_work);
362 } else {
ffe51f01 363 dio_complete(dio, 0, DIO_COMPLETE_ASYNC);
7b7a8665 364 }
8459d86a 365 }
1da177e4
LT
366}
367
368/*
369 * The BIO completion handler simply queues the BIO up for the process-context
370 * handler.
371 *
372 * During I/O bi_private points at the dio. After I/O, bi_private is used to
373 * implement a singly-linked list of completed BIOs, at dio->bio_list.
374 */
4246a0b6 375static void dio_bio_end_io(struct bio *bio)
1da177e4
LT
376{
377 struct dio *dio = bio->bi_private;
378 unsigned long flags;
379
1da177e4
LT
380 spin_lock_irqsave(&dio->bio_lock, flags);
381 bio->bi_private = dio->bio_list;
382 dio->bio_list = bio;
5eb6c7a2 383 if (--dio->refcount == 1 && dio->waiter)
1da177e4
LT
384 wake_up_process(dio->waiter);
385 spin_unlock_irqrestore(&dio->bio_lock, flags);
1da177e4
LT
386}
387
ba253fbf 388static inline void
eb28be2b
AK
389dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
390 struct block_device *bdev,
391 sector_t first_sector, int nr_vecs)
1da177e4
LT
392{
393 struct bio *bio;
394
20d9600c 395 /*
0eb0b63c
CH
396 * bio_alloc() is guaranteed to return a bio when allowed to sleep and
397 * we request a valid number of vectors.
20d9600c 398 */
c6293eac 399 bio = bio_alloc(bdev, nr_vecs, dio->opf, GFP_KERNEL);
4f024f37 400 bio->bi_iter.bi_sector = first_sector;
1da177e4
LT
401 if (dio->is_async)
402 bio->bi_end_io = dio_bio_end_aio;
403 else
404 bio->bi_end_io = dio_bio_end_io;
eb28be2b
AK
405 sdio->bio = bio;
406 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
1da177e4
LT
407}
408
409/*
410 * In the AIO read case we speculatively dirty the pages before starting IO.
411 * During IO completion, any of these pages which happen to have been written
412 * back will be redirtied by bio_check_pages_dirty().
0273201e
ZB
413 *
414 * bios hold a dio reference between submit_bio and ->end_io.
1da177e4 415 */
ba253fbf 416static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
1da177e4 417{
c6293eac 418 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
eb28be2b 419 struct bio *bio = sdio->bio;
5eb6c7a2 420 unsigned long flags;
1da177e4
LT
421
422 bio->bi_private = dio;
5eb6c7a2
ZB
423
424 spin_lock_irqsave(&dio->bio_lock, flags);
425 dio->refcount++;
426 spin_unlock_irqrestore(&dio->bio_lock, flags);
427
c6293eac 428 if (dio->is_async && dio_op == REQ_OP_READ && dio->should_dirty)
1da177e4 429 bio_set_pages_dirty(bio);
5eb6c7a2 430
309dca30 431 dio->bio_disk = bio->bi_bdev->bd_disk;
c1c53460 432
0aaf08de 433 submit_bio(bio);
1da177e4 434
eb28be2b
AK
435 sdio->bio = NULL;
436 sdio->boundary = 0;
437 sdio->logical_offset_in_bio = 0;
1da177e4
LT
438}
439
440/*
441 * Release any resources in case of a failure
442 */
ba253fbf 443static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
1da177e4 444{
7b2c99d1 445 while (sdio->head < sdio->tail)
09cbfeaf 446 put_page(dio->pages[sdio->head++]);
1da177e4
LT
447}
448
449/*
0273201e
ZB
450 * Wait for the next BIO to complete. Remove it and return it. NULL is
451 * returned once all BIOs have been completed. This must only be called once
452 * all bios have been issued so that dio->refcount can only decrease. This
3d742d4b 453 * requires that the caller hold a reference on the dio.
1da177e4
LT
454 */
455static struct bio *dio_await_one(struct dio *dio)
456{
457 unsigned long flags;
0273201e 458 struct bio *bio = NULL;
1da177e4
LT
459
460 spin_lock_irqsave(&dio->bio_lock, flags);
5eb6c7a2
ZB
461
462 /*
463 * Wait as long as the list is empty and there are bios in flight. bio
464 * completion drops the count, maybe adds to the list, and wakes while
465 * holding the bio_lock so we don't need set_current_state()'s barrier
466 * and can call it after testing our condition.
467 */
468 while (dio->refcount > 1 && dio->bio_list == NULL) {
469 __set_current_state(TASK_UNINTERRUPTIBLE);
470 dio->waiter = current;
471 spin_unlock_irqrestore(&dio->bio_lock, flags);
94c2ed58 472 blk_io_schedule();
5eb6c7a2
ZB
473 /* wake up sets us TASK_RUNNING */
474 spin_lock_irqsave(&dio->bio_lock, flags);
475 dio->waiter = NULL;
1da177e4 476 }
0273201e
ZB
477 if (dio->bio_list) {
478 bio = dio->bio_list;
479 dio->bio_list = bio->bi_private;
480 }
1da177e4
LT
481 spin_unlock_irqrestore(&dio->bio_lock, flags);
482 return bio;
483}
484
485/*
486 * Process one completed BIO. No locks are held.
487 */
4e4cbee9 488static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio)
1da177e4 489{
4e4cbee9 490 blk_status_t err = bio->bi_status;
c6293eac
BVA
491 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
492 bool should_dirty = dio_op == REQ_OP_READ && dio->should_dirty;
1da177e4 493
03a07c92
GR
494 if (err) {
495 if (err == BLK_STS_AGAIN && (bio->bi_opf & REQ_NOWAIT))
496 dio->io_error = -EAGAIN;
497 else
498 dio->io_error = -EIO;
499 }
1da177e4 500
d7c8aa85 501 if (dio->is_async && should_dirty) {
7ddc971f 502 bio_check_pages_dirty(bio); /* transfers ownership */
1da177e4 503 } else {
d7c8aa85 504 bio_release_pages(bio, should_dirty);
1da177e4
LT
505 bio_put(bio);
506 }
9b81c842 507 return err;
1da177e4
LT
508}
509
510/*
0273201e
ZB
511 * Wait on and process all in-flight BIOs. This must only be called once
512 * all bios have been issued so that the refcount can only decrease.
513 * This just waits for all bios to make it through dio_bio_complete. IO
beb7dd86 514 * errors are propagated through dio->io_error and should be propagated via
0273201e 515 * dio_complete().
1da177e4 516 */
6d544bb4 517static void dio_await_completion(struct dio *dio)
1da177e4 518{
0273201e
ZB
519 struct bio *bio;
520 do {
521 bio = dio_await_one(dio);
522 if (bio)
523 dio_bio_complete(dio, bio);
524 } while (bio);
1da177e4
LT
525}
526
527/*
528 * A really large O_DIRECT read or write can generate a lot of BIOs. So
529 * to keep the memory consumption sane we periodically reap any completed BIOs
530 * during the BIO generation phase.
531 *
532 * This also helps to limit the peak amount of pinned userspace memory.
533 */
ba253fbf 534static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
1da177e4
LT
535{
536 int ret = 0;
537
eb28be2b 538 if (sdio->reap_counter++ >= 64) {
1da177e4
LT
539 while (dio->bio_list) {
540 unsigned long flags;
541 struct bio *bio;
542 int ret2;
543
544 spin_lock_irqsave(&dio->bio_lock, flags);
545 bio = dio->bio_list;
546 dio->bio_list = bio->bi_private;
547 spin_unlock_irqrestore(&dio->bio_lock, flags);
4e4cbee9 548 ret2 = blk_status_to_errno(dio_bio_complete(dio, bio));
1da177e4
LT
549 if (ret == 0)
550 ret = ret2;
551 }
eb28be2b 552 sdio->reap_counter = 0;
1da177e4
LT
553 }
554 return ret;
555}
556
7b7a8665
CH
557static int dio_set_defer_completion(struct dio *dio)
558{
559 struct super_block *sb = dio->inode->i_sb;
560
561 if (dio->defer_completion)
562 return 0;
563 dio->defer_completion = true;
564 if (!sb->s_dio_done_wq)
565 return sb_init_dio_done_wq(sb);
566 return 0;
567}
568
1da177e4
LT
569/*
570 * Call into the fs to map some more disk blocks. We record the current number
eb28be2b 571 * of available blocks at sdio->blocks_available. These are in units of the
93407472 572 * fs blocksize, i_blocksize(inode).
1da177e4
LT
573 *
574 * The fs is allowed to map lots of blocks at once. If it wants to do that,
575 * it uses the passed inode-relative block number as the file offset, as usual.
576 *
1d8fa7a2 577 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
1da177e4
LT
578 * has remaining to do. The fs should not map more than this number of blocks.
579 *
580 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
581 * indicate how much contiguous disk space has been made available at
582 * bh->b_blocknr.
583 *
584 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
585 * This isn't very efficient...
586 *
587 * In the case of filesystem holes: the fs may return an arbitrarily-large
588 * hole by returning an appropriate value in b_size and by clearing
589 * buffer_mapped(). However the direct-io code will only process holes one
1d8fa7a2 590 * block at a time - it will repeatedly call get_block() as it walks the hole.
1da177e4 591 */
18772641
AK
592static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
593 struct buffer_head *map_bh)
1da177e4 594{
c6293eac 595 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
1da177e4 596 int ret;
1da177e4 597 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
ae55e1aa 598 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
1da177e4 599 unsigned long fs_count; /* Number of filesystem-sized blocks */
1da177e4 600 int create;
ab73857e 601 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
8b9433eb 602 loff_t i_size;
1da177e4
LT
603
604 /*
605 * If there was a memory error and we've overwritten all the
606 * mapped blocks then we can now return that memory error
607 */
608 ret = dio->page_errors;
609 if (ret == 0) {
eb28be2b
AK
610 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
611 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
ae55e1aa
TM
612 fs_endblk = (sdio->final_block_in_request - 1) >>
613 sdio->blkfactor;
614 fs_count = fs_endblk - fs_startblk + 1;
1da177e4 615
3c674e74 616 map_bh->b_state = 0;
ab73857e 617 map_bh->b_size = fs_count << i_blkbits;
3c674e74 618
5fe878ae 619 /*
9ecd10b7
EG
620 * For writes that could fill holes inside i_size on a
621 * DIO_SKIP_HOLES filesystem we forbid block creations: only
622 * overwrites are permitted. We will return early to the caller
623 * once we see an unmapped buffer head returned, and the caller
624 * will fall back to buffered I/O.
5fe878ae
CH
625 *
626 * Otherwise the decision is left to the get_blocks method,
627 * which may decide to handle it or also return an unmapped
628 * buffer head.
629 */
c6293eac 630 create = dio_op == REQ_OP_WRITE;
5fe878ae 631 if (dio->flags & DIO_SKIP_HOLES) {
8b9433eb
EF
632 i_size = i_size_read(dio->inode);
633 if (i_size && fs_startblk <= (i_size - 1) >> i_blkbits)
1da177e4 634 create = 0;
1da177e4 635 }
3c674e74 636
eb28be2b 637 ret = (*sdio->get_block)(dio->inode, fs_startblk,
1da177e4 638 map_bh, create);
18772641
AK
639
640 /* Store for completion */
641 dio->private = map_bh->b_private;
7b7a8665
CH
642
643 if (ret == 0 && buffer_defer_completion(map_bh))
644 ret = dio_set_defer_completion(dio);
1da177e4
LT
645 }
646 return ret;
647}
648
649/*
650 * There is no bio. Make one now.
651 */
ba253fbf
AK
652static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
653 sector_t start_sector, struct buffer_head *map_bh)
1da177e4
LT
654{
655 sector_t sector;
656 int ret, nr_pages;
657
eb28be2b 658 ret = dio_bio_reap(dio, sdio);
1da177e4
LT
659 if (ret)
660 goto out;
eb28be2b 661 sector = start_sector << (sdio->blkbits - 9);
5f7136db 662 nr_pages = bio_max_segs(sdio->pages_in_io);
1da177e4 663 BUG_ON(nr_pages <= 0);
18772641 664 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
eb28be2b 665 sdio->boundary = 0;
1da177e4
LT
666out:
667 return ret;
668}
669
670/*
671 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
672 * that was successful then update final_block_in_bio and take a ref against
673 * the just-added page.
674 *
675 * Return zero on success. Non-zero means the caller needs to start a new BIO.
676 */
ba253fbf 677static inline int dio_bio_add_page(struct dio_submit *sdio)
1da177e4
LT
678{
679 int ret;
680
eb28be2b
AK
681 ret = bio_add_page(sdio->bio, sdio->cur_page,
682 sdio->cur_page_len, sdio->cur_page_offset);
683 if (ret == sdio->cur_page_len) {
1da177e4
LT
684 /*
685 * Decrement count only, if we are done with this page
686 */
eb28be2b
AK
687 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
688 sdio->pages_in_io--;
09cbfeaf 689 get_page(sdio->cur_page);
eb28be2b
AK
690 sdio->final_block_in_bio = sdio->cur_page_block +
691 (sdio->cur_page_len >> sdio->blkbits);
1da177e4
LT
692 ret = 0;
693 } else {
694 ret = 1;
695 }
696 return ret;
697}
698
699/*
700 * Put cur_page under IO. The section of cur_page which is described by
701 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
702 * starts on-disk at cur_page_block.
703 *
704 * We take a ref against the page here (on behalf of its presence in the bio).
705 *
706 * The caller of this function is responsible for removing cur_page from the
707 * dio, and for dropping the refcount which came from that presence.
708 */
ba253fbf
AK
709static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
710 struct buffer_head *map_bh)
1da177e4
LT
711{
712 int ret = 0;
713
eb28be2b
AK
714 if (sdio->bio) {
715 loff_t cur_offset = sdio->cur_page_fs_offset;
716 loff_t bio_next_offset = sdio->logical_offset_in_bio +
4f024f37 717 sdio->bio->bi_iter.bi_size;
c2c6ca41 718
1da177e4 719 /*
c2c6ca41
JB
720 * See whether this new request is contiguous with the old.
721 *
f0940cee
NK
722 * Btrfs cannot handle having logically non-contiguous requests
723 * submitted. For example if you have
c2c6ca41
JB
724 *
725 * Logical: [0-4095][HOLE][8192-12287]
f0940cee 726 * Physical: [0-4095] [4096-8191]
c2c6ca41
JB
727 *
728 * We cannot submit those pages together as one BIO. So if our
729 * current logical offset in the file does not equal what would
730 * be the next logical offset in the bio, submit the bio we
731 * have.
1da177e4 732 */
eb28be2b 733 if (sdio->final_block_in_bio != sdio->cur_page_block ||
c2c6ca41 734 cur_offset != bio_next_offset)
eb28be2b 735 dio_bio_submit(dio, sdio);
1da177e4
LT
736 }
737
eb28be2b 738 if (sdio->bio == NULL) {
18772641 739 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
1da177e4
LT
740 if (ret)
741 goto out;
742 }
743
eb28be2b
AK
744 if (dio_bio_add_page(sdio) != 0) {
745 dio_bio_submit(dio, sdio);
18772641 746 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
1da177e4 747 if (ret == 0) {
eb28be2b 748 ret = dio_bio_add_page(sdio);
1da177e4
LT
749 BUG_ON(ret != 0);
750 }
751 }
752out:
753 return ret;
754}
755
756/*
757 * An autonomous function to put a chunk of a page under deferred IO.
758 *
759 * The caller doesn't actually know (or care) whether this piece of page is in
760 * a BIO, or is under IO or whatever. We just take care of all possible
761 * situations here. The separation between the logic of do_direct_IO() and
762 * that of submit_page_section() is important for clarity. Please don't break.
763 *
764 * The chunk of page starts on-disk at blocknr.
765 *
766 * We perform deferred IO, by recording the last-submitted page inside our
767 * private part of the dio structure. If possible, we just expand the IO
768 * across that page here.
769 *
770 * If that doesn't work out then we put the old page into the bio and add this
771 * page to the dio instead.
772 */
ba253fbf 773static inline int
eb28be2b 774submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
18772641
AK
775 unsigned offset, unsigned len, sector_t blocknr,
776 struct buffer_head *map_bh)
1da177e4 777{
c6293eac 778 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
1da177e4 779 int ret = 0;
df41872b 780 int boundary = sdio->boundary; /* dio_send_cur_page may clear it */
1da177e4 781
c6293eac 782 if (dio_op == REQ_OP_WRITE) {
98c4d57d
AM
783 /*
784 * Read accounting is performed in submit_bio()
785 */
786 task_io_account_write(len);
787 }
788
1da177e4
LT
789 /*
790 * Can we just grow the current page's presence in the dio?
791 */
eb28be2b
AK
792 if (sdio->cur_page == page &&
793 sdio->cur_page_offset + sdio->cur_page_len == offset &&
794 sdio->cur_page_block +
795 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
796 sdio->cur_page_len += len;
1da177e4
LT
797 goto out;
798 }
799
800 /*
801 * If there's a deferred page already there then send it.
802 */
eb28be2b 803 if (sdio->cur_page) {
18772641 804 ret = dio_send_cur_page(dio, sdio, map_bh);
09cbfeaf 805 put_page(sdio->cur_page);
eb28be2b 806 sdio->cur_page = NULL;
1da177e4 807 if (ret)
b1058b98 808 return ret;
1da177e4
LT
809 }
810
09cbfeaf 811 get_page(page); /* It is in dio */
eb28be2b
AK
812 sdio->cur_page = page;
813 sdio->cur_page_offset = offset;
814 sdio->cur_page_len = len;
815 sdio->cur_page_block = blocknr;
816 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
1da177e4 817out:
b1058b98 818 /*
df41872b 819 * If boundary then we want to schedule the IO now to
b1058b98
JK
820 * avoid metadata seeks.
821 */
df41872b 822 if (boundary) {
b1058b98 823 ret = dio_send_cur_page(dio, sdio, map_bh);
899f0429
AG
824 if (sdio->bio)
825 dio_bio_submit(dio, sdio);
09cbfeaf 826 put_page(sdio->cur_page);
b1058b98
JK
827 sdio->cur_page = NULL;
828 }
1da177e4
LT
829 return ret;
830}
831
1da177e4
LT
832/*
833 * If we are not writing the entire block and get_block() allocated
834 * the block for us, we need to fill-in the unused portion of the
835 * block with zeros. This happens only if user-buffer, fileoffset or
836 * io length is not filesystem block-size multiple.
837 *
838 * `end' is zero if we're doing the start of the IO, 1 at the end of the
839 * IO.
840 */
ba253fbf
AK
841static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
842 int end, struct buffer_head *map_bh)
1da177e4
LT
843{
844 unsigned dio_blocks_per_fs_block;
845 unsigned this_chunk_blocks; /* In dio_blocks */
846 unsigned this_chunk_bytes;
847 struct page *page;
848
eb28be2b 849 sdio->start_zero_done = 1;
18772641 850 if (!sdio->blkfactor || !buffer_new(map_bh))
1da177e4
LT
851 return;
852
eb28be2b
AK
853 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
854 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
1da177e4
LT
855
856 if (!this_chunk_blocks)
857 return;
858
859 /*
860 * We need to zero out part of an fs block. It is either at the
861 * beginning or the end of the fs block.
862 */
863 if (end)
864 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
865
eb28be2b 866 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
1da177e4 867
557ed1fa 868 page = ZERO_PAGE(0);
eb28be2b 869 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
18772641 870 sdio->next_block_for_io, map_bh))
1da177e4
LT
871 return;
872
eb28be2b 873 sdio->next_block_for_io += this_chunk_blocks;
1da177e4
LT
874}
875
876/*
877 * Walk the user pages, and the file, mapping blocks to disk and generating
878 * a sequence of (page,offset,len,block) mappings. These mappings are injected
879 * into submit_page_section(), which takes care of the next stage of submission
880 *
881 * Direct IO against a blockdev is different from a file. Because we can
882 * happily perform page-sized but 512-byte aligned IOs. It is important that
883 * blockdev IO be able to have fine alignment and large sizes.
884 *
1d8fa7a2 885 * So what we do is to permit the ->get_block function to populate bh.b_size
1da177e4
LT
886 * with the size of IO which is permitted at this offset and this i_blkbits.
887 *
888 * For best results, the blockdev should be set up with 512-byte i_blkbits and
1d8fa7a2 889 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
1da177e4
LT
890 * fine alignment but still allows this function to work in PAGE_SIZE units.
891 */
18772641
AK
892static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
893 struct buffer_head *map_bh)
1da177e4 894{
c6293eac 895 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
eb28be2b 896 const unsigned blkbits = sdio->blkbits;
dd545b52 897 const unsigned i_blkbits = blkbits + sdio->blkfactor;
1da177e4
LT
898 int ret = 0;
899
eb28be2b 900 while (sdio->block_in_file < sdio->final_block_in_request) {
7b2c99d1
AV
901 struct page *page;
902 size_t from, to;
6fcc5420
BH
903
904 page = dio_get_page(dio, sdio);
1da177e4
LT
905 if (IS_ERR(page)) {
906 ret = PTR_ERR(page);
907 goto out;
908 }
6fcc5420
BH
909 from = sdio->head ? 0 : sdio->from;
910 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
911 sdio->head++;
1da177e4 912
7b2c99d1 913 while (from < to) {
1da177e4
LT
914 unsigned this_chunk_bytes; /* # of bytes mapped */
915 unsigned this_chunk_blocks; /* # of blocks */
916 unsigned u;
917
eb28be2b 918 if (sdio->blocks_available == 0) {
1da177e4
LT
919 /*
920 * Need to go and map some more disk
921 */
922 unsigned long blkmask;
923 unsigned long dio_remainder;
924
18772641 925 ret = get_more_blocks(dio, sdio, map_bh);
1da177e4 926 if (ret) {
09cbfeaf 927 put_page(page);
1da177e4
LT
928 goto out;
929 }
930 if (!buffer_mapped(map_bh))
931 goto do_holes;
932
eb28be2b 933 sdio->blocks_available =
f734c89c 934 map_bh->b_size >> blkbits;
eb28be2b
AK
935 sdio->next_block_for_io =
936 map_bh->b_blocknr << sdio->blkfactor;
f734c89c
JK
937 if (buffer_new(map_bh)) {
938 clean_bdev_aliases(
939 map_bh->b_bdev,
940 map_bh->b_blocknr,
dd545b52 941 map_bh->b_size >> i_blkbits);
f734c89c 942 }
1da177e4 943
eb28be2b 944 if (!sdio->blkfactor)
1da177e4
LT
945 goto do_holes;
946
eb28be2b
AK
947 blkmask = (1 << sdio->blkfactor) - 1;
948 dio_remainder = (sdio->block_in_file & blkmask);
1da177e4
LT
949
950 /*
951 * If we are at the start of IO and that IO
952 * starts partway into a fs-block,
953 * dio_remainder will be non-zero. If the IO
954 * is a read then we can simply advance the IO
955 * cursor to the first block which is to be
956 * read. But if the IO is a write and the
957 * block was newly allocated we cannot do that;
958 * the start of the fs block must be zeroed out
959 * on-disk
960 */
961 if (!buffer_new(map_bh))
eb28be2b
AK
962 sdio->next_block_for_io += dio_remainder;
963 sdio->blocks_available -= dio_remainder;
1da177e4
LT
964 }
965do_holes:
966 /* Handle holes */
967 if (!buffer_mapped(map_bh)) {
35dc8161 968 loff_t i_size_aligned;
1da177e4
LT
969
970 /* AKPM: eargh, -ENOTBLK is a hack */
c6293eac 971 if (dio_op == REQ_OP_WRITE) {
09cbfeaf 972 put_page(page);
1da177e4
LT
973 return -ENOTBLK;
974 }
975
35dc8161
JM
976 /*
977 * Be sure to account for a partial block as the
978 * last block in the file
979 */
980 i_size_aligned = ALIGN(i_size_read(dio->inode),
981 1 << blkbits);
eb28be2b 982 if (sdio->block_in_file >=
35dc8161 983 i_size_aligned >> blkbits) {
1da177e4 984 /* We hit eof */
09cbfeaf 985 put_page(page);
1da177e4
LT
986 goto out;
987 }
7b2c99d1 988 zero_user(page, from, 1 << blkbits);
eb28be2b 989 sdio->block_in_file++;
7b2c99d1 990 from += 1 << blkbits;
3320c60b 991 dio->result += 1 << blkbits;
1da177e4
LT
992 goto next_block;
993 }
994
995 /*
996 * If we're performing IO which has an alignment which
997 * is finer than the underlying fs, go check to see if
998 * we must zero out the start of this block.
999 */
eb28be2b 1000 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
18772641 1001 dio_zero_block(dio, sdio, 0, map_bh);
1da177e4
LT
1002
1003 /*
1004 * Work out, in this_chunk_blocks, how much disk we
1005 * can add to this page
1006 */
eb28be2b 1007 this_chunk_blocks = sdio->blocks_available;
7b2c99d1 1008 u = (to - from) >> blkbits;
1da177e4
LT
1009 if (this_chunk_blocks > u)
1010 this_chunk_blocks = u;
eb28be2b 1011 u = sdio->final_block_in_request - sdio->block_in_file;
1da177e4
LT
1012 if (this_chunk_blocks > u)
1013 this_chunk_blocks = u;
1014 this_chunk_bytes = this_chunk_blocks << blkbits;
1015 BUG_ON(this_chunk_bytes == 0);
1016
092c8d46
JK
1017 if (this_chunk_blocks == sdio->blocks_available)
1018 sdio->boundary = buffer_boundary(map_bh);
eb28be2b 1019 ret = submit_page_section(dio, sdio, page,
7b2c99d1 1020 from,
eb28be2b 1021 this_chunk_bytes,
18772641
AK
1022 sdio->next_block_for_io,
1023 map_bh);
1da177e4 1024 if (ret) {
09cbfeaf 1025 put_page(page);
1da177e4
LT
1026 goto out;
1027 }
eb28be2b 1028 sdio->next_block_for_io += this_chunk_blocks;
1da177e4 1029
eb28be2b 1030 sdio->block_in_file += this_chunk_blocks;
7b2c99d1
AV
1031 from += this_chunk_bytes;
1032 dio->result += this_chunk_bytes;
eb28be2b 1033 sdio->blocks_available -= this_chunk_blocks;
1da177e4 1034next_block:
eb28be2b
AK
1035 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1036 if (sdio->block_in_file == sdio->final_block_in_request)
1da177e4
LT
1037 break;
1038 }
1039
1040 /* Drop the ref which was taken in get_user_pages() */
09cbfeaf 1041 put_page(page);
1da177e4
LT
1042 }
1043out:
1044 return ret;
1045}
1046
847cc637 1047static inline int drop_refcount(struct dio *dio)
1da177e4 1048{
847cc637 1049 int ret2;
5eb6c7a2 1050 unsigned long flags;
1da177e4 1051
8459d86a
ZB
1052 /*
1053 * Sync will always be dropping the final ref and completing the
5eb6c7a2
ZB
1054 * operation. AIO can if it was a broken operation described above or
1055 * in fact if all the bios race to complete before we get here. In
1056 * that case dio_complete() translates the EIOCBQUEUED into the proper
04b2fa9f 1057 * return code that the caller will hand to ->complete().
5eb6c7a2
ZB
1058 *
1059 * This is managed by the bio_lock instead of being an atomic_t so that
1060 * completion paths can drop their ref and use the remaining count to
1061 * decide to wake the submission path atomically.
8459d86a 1062 */
5eb6c7a2
ZB
1063 spin_lock_irqsave(&dio->bio_lock, flags);
1064 ret2 = --dio->refcount;
1065 spin_unlock_irqrestore(&dio->bio_lock, flags);
847cc637 1066 return ret2;
1da177e4
LT
1067}
1068
eafdc7d1
CH
1069/*
1070 * This is a library function for use by filesystem drivers.
1071 *
1072 * The locking rules are governed by the flags parameter:
1073 * - if the flags value contains DIO_LOCKING we use a fancy locking
1074 * scheme for dumb filesystems.
1075 * For writes this function is called under i_mutex and returns with
1076 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1077 * taken and dropped again before returning.
eafdc7d1
CH
1078 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1079 * internal locking but rather rely on the filesystem to synchronize
1080 * direct I/O reads/writes versus each other and truncate.
df2d6f26
CH
1081 *
1082 * To help with locking against truncate we incremented the i_dio_count
1083 * counter before starting direct I/O, and decrement it once we are done.
1084 * Truncate can wait for it to reach zero to provide exclusion. It is
1085 * expected that filesystem provide exclusion between new direct I/O
1086 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1087 * but other filesystems need to take care of this on their own.
ba253fbf
AK
1088 *
1089 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1090 * is always inlined. Otherwise gcc is unable to split the structure into
1091 * individual fields and will generate much worse code. This is important
1092 * for the whole file.
eafdc7d1 1093 */
c22198e7
CH
1094ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1095 struct block_device *bdev, struct iov_iter *iter,
1096 get_block_t get_block, dio_iodone_t end_io,
0aaf08de 1097 int flags)
1da177e4 1098{
6aa7de05 1099 unsigned i_blkbits = READ_ONCE(inode->i_blkbits);
ab73857e 1100 unsigned blkbits = i_blkbits;
1da177e4
LT
1101 unsigned blocksize_mask = (1 << blkbits) - 1;
1102 ssize_t retval = -EINVAL;
1c0ff0f1 1103 const size_t count = iov_iter_count(iter);
c8b8e32d 1104 loff_t offset = iocb->ki_pos;
1c0ff0f1 1105 const loff_t end = offset + count;
1da177e4 1106 struct dio *dio;
eb28be2b 1107 struct dio_submit sdio = { 0, };
847cc637 1108 struct buffer_head map_bh = { 0, };
647d1e4c 1109 struct blk_plug plug;
886a3911 1110 unsigned long align = offset | iov_iter_alignment(iter);
1da177e4 1111
65dd2aa9
AK
1112 /*
1113 * Avoid references to bdev if not absolutely needed to give
1114 * the early prefetch in the caller enough time.
1115 */
1da177e4 1116
f9b5570d 1117 /* watch out for a 0 len io from a tricksy fs */
1c0ff0f1 1118 if (iov_iter_rw(iter) == READ && !count)
f9b5570d
CH
1119 return 0;
1120
6e8267f5 1121 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
1da177e4 1122 if (!dio)
46d71602 1123 return -ENOMEM;
23aee091
JM
1124 /*
1125 * Believe it or not, zeroing out the page array caused a .5%
1126 * performance regression in a database benchmark. So, we take
1127 * care to only zero out what's needed.
1128 */
1129 memset(dio, 0, offsetof(struct dio, pages));
1da177e4 1130
5fe878ae 1131 dio->flags = flags;
0a9164cb
GKB
1132 if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ) {
1133 /* will be released by direct_io_worker */
1134 inode_lock(inode);
1da177e4
LT
1135 }
1136
74cedf9b
JK
1137 /* Once we sampled i_size check for reads beyond EOF */
1138 dio->i_size = i_size_read(inode);
1139 if (iov_iter_rw(iter) == READ && offset >= dio->i_size) {
2d4594ac 1140 retval = 0;
46d71602 1141 goto fail_dio;
74cedf9b
JK
1142 }
1143
41b21af3
GKB
1144 if (align & blocksize_mask) {
1145 if (bdev)
1146 blkbits = blksize_bits(bdev_logical_block_size(bdev));
1147 blocksize_mask = (1 << blkbits) - 1;
1148 if (align & blocksize_mask)
1149 goto fail_dio;
1150 }
1151
0a9164cb
GKB
1152 if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ) {
1153 struct address_space *mapping = iocb->ki_filp->f_mapping;
1154
1155 retval = filemap_write_and_wait_range(mapping, offset, end - 1);
1156 if (retval)
1157 goto fail_dio;
74cedf9b
JK
1158 }
1159
1da177e4 1160 /*
60392573
CH
1161 * For file extending writes updating i_size before data writeouts
1162 * complete can expose uninitialized blocks in dumb filesystems.
1163 * In that case we need to wait for I/O completion even if asked
1164 * for an asynchronous write.
1da177e4 1165 */
60392573
CH
1166 if (is_sync_kiocb(iocb))
1167 dio->is_async = false;
c8f4c36f 1168 else if (iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
60392573
CH
1169 dio->is_async = false;
1170 else
1171 dio->is_async = true;
1172
847cc637 1173 dio->inode = inode;
8a4c1e42 1174 if (iov_iter_rw(iter) == WRITE) {
c6293eac 1175 dio->opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
03a07c92 1176 if (iocb->ki_flags & IOCB_NOWAIT)
c6293eac 1177 dio->opf |= REQ_NOWAIT;
8a4c1e42 1178 } else {
c6293eac 1179 dio->opf = REQ_OP_READ;
8a4c1e42 1180 }
02afc27f
CH
1181
1182 /*
1183 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1184 * so that we can call ->fsync.
1185 */
332391a9
LC
1186 if (dio->is_async && iov_iter_rw(iter) == WRITE) {
1187 retval = 0;
91b94c5d 1188 if (iocb_is_dsync(iocb))
332391a9
LC
1189 retval = dio_set_defer_completion(dio);
1190 else if (!dio->inode->i_sb->s_dio_done_wq) {
1191 /*
1192 * In case of AIO write racing with buffered read we
1193 * need to defer completion. We can't decide this now,
1194 * however the workqueue needs to be initialized here.
1195 */
1196 retval = sb_init_dio_done_wq(dio->inode->i_sb);
1197 }
46d71602
GKB
1198 if (retval)
1199 goto fail_dio;
02afc27f
CH
1200 }
1201
1202 /*
1203 * Will be decremented at I/O completion time.
1204 */
ce3077ee 1205 inode_dio_begin(inode);
02afc27f
CH
1206
1207 retval = 0;
847cc637 1208 sdio.blkbits = blkbits;
ab73857e 1209 sdio.blkfactor = i_blkbits - blkbits;
847cc637
AK
1210 sdio.block_in_file = offset >> blkbits;
1211
1212 sdio.get_block = get_block;
1213 dio->end_io = end_io;
847cc637
AK
1214 sdio.final_block_in_bio = -1;
1215 sdio.next_block_for_io = -1;
1216
1217 dio->iocb = iocb;
847cc637
AK
1218
1219 spin_lock_init(&dio->bio_lock);
1220 dio->refcount = 1;
1221
fcb14cb1 1222 dio->should_dirty = user_backed_iter(iter) && iov_iter_rw(iter) == READ;
7b2c99d1 1223 sdio.iter = iter;
1c0ff0f1 1224 sdio.final_block_in_request = end >> blkbits;
7b2c99d1 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
3d742d4b 1244 * handled by buffered I/O when we return
847cc637
AK
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;
09cbfeaf 1260 put_page(sdio.cur_page);
847cc637
AK
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))
5955102c 1280 inode_unlock(dio->inode);
847cc637
AK
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) {
ffe51f01 1297 retval = dio_complete(dio, retval, DIO_COMPLETE_INVALIDATE);
847cc637
AK
1298 } else
1299 BUG_ON(retval != -EIOCBQUEUED);
1da177e4 1300
46d71602
GKB
1301 return retval;
1302
1303fail_dio:
1304 if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ)
1305 inode_unlock(inode);
1306
1307 kmem_cache_free(dio_cache, dio);
7bb46a67 1308 return retval;
1309}
1da177e4 1310EXPORT_SYMBOL(__blockdev_direct_IO);
6e8267f5
AK
1311
1312static __init int dio_init(void)
1313{
1314 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1315 return 0;
1316}
1317module_init(dio_init)