treewide: Use array_size() in f2fs_kzalloc()
[linux-2.6-block.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
ffe51f01
LC
47/*
48 * Flags for dio_complete()
49 */
50#define DIO_COMPLETE_ASYNC 0x01 /* This is async IO */
51#define DIO_COMPLETE_INVALIDATE 0x02 /* Can invalidate pages */
52
1da177e4
LT
53/*
54 * This code generally works in units of "dio_blocks". A dio_block is
55 * somewhere between the hard sector size and the filesystem block size. it
56 * is determined on a per-invocation basis. When talking to the filesystem
57 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
58 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
59 * to bio_block quantities by shifting left by blkfactor.
60 *
61 * If blkfactor is zero then the user's request was aligned to the filesystem's
62 * blocksize.
1da177e4
LT
63 */
64
eb28be2b
AK
65/* dio_state only used in the submission path */
66
67struct dio_submit {
1da177e4 68 struct bio *bio; /* bio under assembly */
1da177e4
LT
69 unsigned blkbits; /* doesn't change */
70 unsigned blkfactor; /* When we're using an alignment which
71 is finer than the filesystem's soft
72 blocksize, this specifies how much
73 finer. blkfactor=2 means 1/4-block
74 alignment. Does not change */
75 unsigned start_zero_done; /* flag: sub-blocksize zeroing has
76 been performed at the start of a
77 write */
78 int pages_in_io; /* approximate total IO pages */
1da177e4
LT
79 sector_t block_in_file; /* Current offset into the underlying
80 file in dio_block units. */
81 unsigned blocks_available; /* At block_in_file. changes */
0dc2bc49 82 int reap_counter; /* rate limit reaping */
1da177e4 83 sector_t final_block_in_request;/* doesn't change */
1da177e4 84 int boundary; /* prev block is at a boundary */
1d8fa7a2 85 get_block_t *get_block; /* block mapping function */
facd07b0 86 dio_submit_t *submit_io; /* IO submition function */
eb28be2b 87
facd07b0 88 loff_t logical_offset_in_bio; /* current first logical block in bio */
1da177e4
LT
89 sector_t final_block_in_bio; /* current final block in bio + 1 */
90 sector_t next_block_for_io; /* next block to be put under IO,
91 in dio_blocks units */
1da177e4
LT
92
93 /*
94 * Deferred addition of a page to the dio. These variables are
95 * private to dio_send_cur_page(), submit_page_section() and
96 * dio_bio_add_page().
97 */
98 struct page *cur_page; /* The page */
99 unsigned cur_page_offset; /* Offset into it, in bytes */
100 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
101 sector_t cur_page_block; /* Where it starts */
facd07b0 102 loff_t cur_page_fs_offset; /* Offset in file */
1da177e4 103
7b2c99d1 104 struct iov_iter *iter;
1da177e4
LT
105 /*
106 * Page queue. These variables belong to dio_refill_pages() and
107 * dio_get_page().
108 */
1da177e4
LT
109 unsigned head; /* next page to process */
110 unsigned tail; /* last valid page + 1 */
7b2c99d1 111 size_t from, to;
eb28be2b
AK
112};
113
114/* dio_state communicated between submission path and end_io */
115struct dio {
116 int flags; /* doesn't change */
8a4c1e42
MC
117 int op;
118 int op_flags;
15c4f638 119 blk_qc_t bio_cookie;
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{
7b2c99d1 168 ssize_t ret;
1da177e4 169
2c80929c 170 ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES,
7b2c99d1 171 &sdio->from);
1da177e4 172
8a4c1e42 173 if (ret < 0 && sdio->blocks_available && (dio->op == REQ_OP_WRITE)) {
557ed1fa 174 struct page *page = ZERO_PAGE(0);
1da177e4
LT
175 /*
176 * A memory fault, but the filesystem has some outstanding
177 * mapped blocks. We need to use those blocks up to avoid
178 * leaking stale data in the file.
179 */
180 if (dio->page_errors == 0)
181 dio->page_errors = ret;
09cbfeaf 182 get_page(page);
b5810039 183 dio->pages[0] = page;
eb28be2b
AK
184 sdio->head = 0;
185 sdio->tail = 1;
7b2c99d1
AV
186 sdio->from = 0;
187 sdio->to = PAGE_SIZE;
188 return 0;
1da177e4
LT
189 }
190
191 if (ret >= 0) {
7b2c99d1
AV
192 iov_iter_advance(sdio->iter, ret);
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
5a9d929d
DW
222/*
223 * Warn about a page cache invalidation failure during a direct io write.
224 */
225void dio_warn_stale_pagecache(struct file *filp)
226{
227 static DEFINE_RATELIMIT_STATE(_rs, 86400 * HZ, DEFAULT_RATELIMIT_BURST);
228 char pathname[128];
229 struct inode *inode = file_inode(filp);
230 char *path;
231
232 errseq_set(&inode->i_mapping->wb_err, -EIO);
233 if (__ratelimit(&_rs)) {
234 path = file_path(filp, pathname, sizeof(pathname));
235 if (IS_ERR(path))
236 path = "(unknown)";
237 pr_crit("Page cache invalidation failure on direct I/O. Possible data corruption due to collision with buffered I/O!\n");
238 pr_crit("File: %s PID: %d Comm: %.20s\n", path, current->pid,
239 current->comm);
240 }
241}
242
6d544bb4
ZB
243/**
244 * dio_complete() - called when all DIO BIO I/O has been completed
245 * @offset: the byte offset in the file of the completed operation
246 *
7b7a8665
CH
247 * This drops i_dio_count, lets interested parties know that a DIO operation
248 * has completed, and calculates the resulting return code for the operation.
6d544bb4
ZB
249 *
250 * It lets the filesystem know if it registered an interest earlier via
251 * get_block. Pass the private field of the map buffer_head so that
252 * filesystems can use it to hold additional state between get_block calls and
253 * dio_complete.
1da177e4 254 */
ffe51f01 255static ssize_t dio_complete(struct dio *dio, ssize_t ret, unsigned int flags)
1da177e4 256{
716b9bc0 257 loff_t offset = dio->iocb->ki_pos;
6d544bb4 258 ssize_t transferred = 0;
332391a9 259 int err;
6d544bb4 260
8459d86a
ZB
261 /*
262 * AIO submission can race with bio completion to get here while
263 * expecting to have the last io completed by bio completion.
264 * In that case -EIOCBQUEUED is in fact not an error we want
265 * to preserve through this call.
266 */
267 if (ret == -EIOCBQUEUED)
268 ret = 0;
269
6d544bb4
ZB
270 if (dio->result) {
271 transferred = dio->result;
272
273 /* Check for short read case */
8a4c1e42
MC
274 if ((dio->op == REQ_OP_READ) &&
275 ((offset + transferred) > dio->i_size))
6d544bb4 276 transferred = dio->i_size - offset;
4038acdb
AV
277 /* ignore EFAULT if some IO has been done */
278 if (unlikely(ret == -EFAULT) && transferred)
279 ret = 0;
6d544bb4
ZB
280 }
281
6d544bb4
ZB
282 if (ret == 0)
283 ret = dio->page_errors;
284 if (ret == 0)
285 ret = dio->io_error;
286 if (ret == 0)
287 ret = transferred;
288
5e25c269
EG
289 if (dio->end_io) {
290 // XXX: ki_pos??
291 err = dio->end_io(dio->iocb, offset, ret, dio->private);
292 if (err)
293 ret = err;
294 }
295
332391a9
LC
296 /*
297 * Try again to invalidate clean pages which might have been cached by
298 * non-direct readahead, or faulted in by get_user_pages() if the source
299 * of the write was an mmap'ed region of the file we're writing. Either
300 * one is a pretty crazy thing to do, so we don't support it 100%. If
301 * this invalidation fails, tough, the write still worked...
5e25c269
EG
302 *
303 * And this page cache invalidation has to be after dio->end_io(), as
304 * some filesystems convert unwritten extents to real allocations in
305 * end_io() when necessary, otherwise a racing buffer read would cache
306 * zeros from unwritten extents.
332391a9 307 */
ffe51f01
LC
308 if (flags & DIO_COMPLETE_INVALIDATE &&
309 ret > 0 && dio->op == REQ_OP_WRITE &&
332391a9
LC
310 dio->inode->i_mapping->nrpages) {
311 err = invalidate_inode_pages2_range(dio->inode->i_mapping,
312 offset >> PAGE_SHIFT,
313 (offset + ret - 1) >> PAGE_SHIFT);
5a9d929d
DW
314 if (err)
315 dio_warn_stale_pagecache(dio->iocb->ki_filp);
332391a9
LC
316 }
317
ce3077ee 318 inode_dio_end(dio->inode);
fe0f07d0 319
ffe51f01 320 if (flags & DIO_COMPLETE_ASYNC) {
e2592217
CH
321 /*
322 * generic_write_sync expects ki_pos to have been updated
323 * already, but the submission path only does this for
324 * synchronous I/O.
325 */
326 dio->iocb->ki_pos += transferred;
02afc27f 327
8a4c1e42 328 if (dio->op == REQ_OP_WRITE)
e2592217 329 ret = generic_write_sync(dio->iocb, transferred);
04b2fa9f 330 dio->iocb->ki_complete(dio->iocb, ret, 0);
02afc27f 331 }
40e2e973 332
7b7a8665 333 kmem_cache_free(dio_cache, dio);
6d544bb4 334 return ret;
1da177e4
LT
335}
336
7b7a8665
CH
337static void dio_aio_complete_work(struct work_struct *work)
338{
339 struct dio *dio = container_of(work, struct dio, complete_work);
340
ffe51f01 341 dio_complete(dio, 0, DIO_COMPLETE_ASYNC | DIO_COMPLETE_INVALIDATE);
7b7a8665
CH
342}
343
4e4cbee9 344static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio);
7b7a8665 345
1da177e4
LT
346/*
347 * Asynchronous IO callback.
348 */
4246a0b6 349static void dio_bio_end_aio(struct bio *bio)
1da177e4
LT
350{
351 struct dio *dio = bio->bi_private;
5eb6c7a2
ZB
352 unsigned long remaining;
353 unsigned long flags;
332391a9 354 bool defer_completion = false;
1da177e4 355
1da177e4
LT
356 /* cleanup the bio */
357 dio_bio_complete(dio, bio);
0273201e 358
5eb6c7a2
ZB
359 spin_lock_irqsave(&dio->bio_lock, flags);
360 remaining = --dio->refcount;
361 if (remaining == 1 && dio->waiter)
20258b2b 362 wake_up_process(dio->waiter);
5eb6c7a2 363 spin_unlock_irqrestore(&dio->bio_lock, flags);
20258b2b 364
8459d86a 365 if (remaining == 0) {
332391a9
LC
366 /*
367 * Defer completion when defer_completion is set or
368 * when the inode has pages mapped and this is AIO write.
369 * We need to invalidate those pages because there is a
370 * chance they contain stale data in the case buffered IO
371 * went in between AIO submission and completion into the
372 * same region.
373 */
374 if (dio->result)
375 defer_completion = dio->defer_completion ||
376 (dio->op == REQ_OP_WRITE &&
377 dio->inode->i_mapping->nrpages);
378 if (defer_completion) {
7b7a8665
CH
379 INIT_WORK(&dio->complete_work, dio_aio_complete_work);
380 queue_work(dio->inode->i_sb->s_dio_done_wq,
381 &dio->complete_work);
382 } else {
ffe51f01 383 dio_complete(dio, 0, DIO_COMPLETE_ASYNC);
7b7a8665 384 }
8459d86a 385 }
1da177e4
LT
386}
387
388/*
389 * The BIO completion handler simply queues the BIO up for the process-context
390 * handler.
391 *
392 * During I/O bi_private points at the dio. After I/O, bi_private is used to
393 * implement a singly-linked list of completed BIOs, at dio->bio_list.
394 */
4246a0b6 395static void dio_bio_end_io(struct bio *bio)
1da177e4
LT
396{
397 struct dio *dio = bio->bi_private;
398 unsigned long flags;
399
1da177e4
LT
400 spin_lock_irqsave(&dio->bio_lock, flags);
401 bio->bi_private = dio->bio_list;
402 dio->bio_list = bio;
5eb6c7a2 403 if (--dio->refcount == 1 && dio->waiter)
1da177e4
LT
404 wake_up_process(dio->waiter);
405 spin_unlock_irqrestore(&dio->bio_lock, flags);
1da177e4
LT
406}
407
facd07b0
JB
408/**
409 * dio_end_io - handle the end io action for the given bio
410 * @bio: The direct io bio thats being completed
facd07b0
JB
411 *
412 * This is meant to be called by any filesystem that uses their own dio_submit_t
413 * so that the DIO specific endio actions are dealt with after the filesystem
414 * has done it's completion work.
415 */
4055351c 416void dio_end_io(struct bio *bio)
facd07b0
JB
417{
418 struct dio *dio = bio->bi_private;
419
420 if (dio->is_async)
4246a0b6 421 dio_bio_end_aio(bio);
facd07b0 422 else
4246a0b6 423 dio_bio_end_io(bio);
facd07b0
JB
424}
425EXPORT_SYMBOL_GPL(dio_end_io);
426
ba253fbf 427static inline void
eb28be2b
AK
428dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
429 struct block_device *bdev,
430 sector_t first_sector, int nr_vecs)
1da177e4
LT
431{
432 struct bio *bio;
433
20d9600c 434 /*
0eb0b63c
CH
435 * bio_alloc() is guaranteed to return a bio when allowed to sleep and
436 * we request a valid number of vectors.
20d9600c 437 */
1da177e4 438 bio = bio_alloc(GFP_KERNEL, nr_vecs);
1da177e4 439
74d46992 440 bio_set_dev(bio, bdev);
4f024f37 441 bio->bi_iter.bi_sector = first_sector;
8a4c1e42 442 bio_set_op_attrs(bio, dio->op, dio->op_flags);
1da177e4
LT
443 if (dio->is_async)
444 bio->bi_end_io = dio_bio_end_aio;
445 else
446 bio->bi_end_io = dio_bio_end_io;
447
45d06cf7
JA
448 bio->bi_write_hint = dio->iocb->ki_hint;
449
eb28be2b
AK
450 sdio->bio = bio;
451 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
1da177e4
LT
452}
453
454/*
455 * In the AIO read case we speculatively dirty the pages before starting IO.
456 * During IO completion, any of these pages which happen to have been written
457 * back will be redirtied by bio_check_pages_dirty().
0273201e
ZB
458 *
459 * bios hold a dio reference between submit_bio and ->end_io.
1da177e4 460 */
ba253fbf 461static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
1da177e4 462{
eb28be2b 463 struct bio *bio = sdio->bio;
5eb6c7a2 464 unsigned long flags;
1da177e4
LT
465
466 bio->bi_private = dio;
5eb6c7a2
ZB
467
468 spin_lock_irqsave(&dio->bio_lock, flags);
469 dio->refcount++;
470 spin_unlock_irqrestore(&dio->bio_lock, flags);
471
8a4c1e42 472 if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty)
1da177e4 473 bio_set_pages_dirty(bio);
5eb6c7a2 474
74d46992 475 dio->bio_disk = bio->bi_disk;
c1c53460 476
15c4f638 477 if (sdio->submit_io) {
8a4c1e42 478 sdio->submit_io(bio, dio->inode, sdio->logical_offset_in_bio);
15c4f638 479 dio->bio_cookie = BLK_QC_T_NONE;
c1c53460 480 } else
4e49ea4a 481 dio->bio_cookie = submit_bio(bio);
1da177e4 482
eb28be2b
AK
483 sdio->bio = NULL;
484 sdio->boundary = 0;
485 sdio->logical_offset_in_bio = 0;
1da177e4
LT
486}
487
488/*
489 * Release any resources in case of a failure
490 */
ba253fbf 491static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
1da177e4 492{
7b2c99d1 493 while (sdio->head < sdio->tail)
09cbfeaf 494 put_page(dio->pages[sdio->head++]);
1da177e4
LT
495}
496
497/*
0273201e
ZB
498 * Wait for the next BIO to complete. Remove it and return it. NULL is
499 * returned once all BIOs have been completed. This must only be called once
500 * all bios have been issued so that dio->refcount can only decrease. This
501 * requires that that the caller hold a reference on the dio.
1da177e4
LT
502 */
503static struct bio *dio_await_one(struct dio *dio)
504{
505 unsigned long flags;
0273201e 506 struct bio *bio = NULL;
1da177e4
LT
507
508 spin_lock_irqsave(&dio->bio_lock, flags);
5eb6c7a2
ZB
509
510 /*
511 * Wait as long as the list is empty and there are bios in flight. bio
512 * completion drops the count, maybe adds to the list, and wakes while
513 * holding the bio_lock so we don't need set_current_state()'s barrier
514 * and can call it after testing our condition.
515 */
516 while (dio->refcount > 1 && dio->bio_list == NULL) {
517 __set_current_state(TASK_UNINTERRUPTIBLE);
518 dio->waiter = current;
519 spin_unlock_irqrestore(&dio->bio_lock, flags);
c43c83a2 520 if (!(dio->iocb->ki_flags & IOCB_HIPRI) ||
ea435e1b 521 !blk_poll(dio->bio_disk->queue, dio->bio_cookie))
15c4f638 522 io_schedule();
5eb6c7a2
ZB
523 /* wake up sets us TASK_RUNNING */
524 spin_lock_irqsave(&dio->bio_lock, flags);
525 dio->waiter = NULL;
1da177e4 526 }
0273201e
ZB
527 if (dio->bio_list) {
528 bio = dio->bio_list;
529 dio->bio_list = bio->bi_private;
530 }
1da177e4
LT
531 spin_unlock_irqrestore(&dio->bio_lock, flags);
532 return bio;
533}
534
535/*
536 * Process one completed BIO. No locks are held.
537 */
4e4cbee9 538static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio)
1da177e4 539{
cb34e057
KO
540 struct bio_vec *bvec;
541 unsigned i;
4e4cbee9 542 blk_status_t err = bio->bi_status;
1da177e4 543
03a07c92
GR
544 if (err) {
545 if (err == BLK_STS_AGAIN && (bio->bi_opf & REQ_NOWAIT))
546 dio->io_error = -EAGAIN;
547 else
548 dio->io_error = -EIO;
549 }
1da177e4 550
8a4c1e42 551 if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty) {
7ddc971f 552 bio_check_pages_dirty(bio); /* transfers ownership */
1da177e4 553 } else {
cb34e057
KO
554 bio_for_each_segment_all(bvec, bio, i) {
555 struct page *page = bvec->bv_page;
1da177e4 556
8a4c1e42 557 if (dio->op == REQ_OP_READ && !PageCompound(page) &&
53cbf3b1 558 dio->should_dirty)
1da177e4 559 set_page_dirty_lock(page);
09cbfeaf 560 put_page(page);
1da177e4
LT
561 }
562 bio_put(bio);
563 }
9b81c842 564 return err;
1da177e4
LT
565}
566
567/*
0273201e
ZB
568 * Wait on and process all in-flight BIOs. This must only be called once
569 * all bios have been issued so that the refcount can only decrease.
570 * This just waits for all bios to make it through dio_bio_complete. IO
beb7dd86 571 * errors are propagated through dio->io_error and should be propagated via
0273201e 572 * dio_complete().
1da177e4 573 */
6d544bb4 574static void dio_await_completion(struct dio *dio)
1da177e4 575{
0273201e
ZB
576 struct bio *bio;
577 do {
578 bio = dio_await_one(dio);
579 if (bio)
580 dio_bio_complete(dio, bio);
581 } while (bio);
1da177e4
LT
582}
583
584/*
585 * A really large O_DIRECT read or write can generate a lot of BIOs. So
586 * to keep the memory consumption sane we periodically reap any completed BIOs
587 * during the BIO generation phase.
588 *
589 * This also helps to limit the peak amount of pinned userspace memory.
590 */
ba253fbf 591static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
1da177e4
LT
592{
593 int ret = 0;
594
eb28be2b 595 if (sdio->reap_counter++ >= 64) {
1da177e4
LT
596 while (dio->bio_list) {
597 unsigned long flags;
598 struct bio *bio;
599 int ret2;
600
601 spin_lock_irqsave(&dio->bio_lock, flags);
602 bio = dio->bio_list;
603 dio->bio_list = bio->bi_private;
604 spin_unlock_irqrestore(&dio->bio_lock, flags);
4e4cbee9 605 ret2 = blk_status_to_errno(dio_bio_complete(dio, bio));
1da177e4
LT
606 if (ret == 0)
607 ret = ret2;
608 }
eb28be2b 609 sdio->reap_counter = 0;
1da177e4
LT
610 }
611 return ret;
612}
613
7b7a8665
CH
614/*
615 * Create workqueue for deferred direct IO completions. We allocate the
616 * workqueue when it's first needed. This avoids creating workqueue for
617 * filesystems that don't need it and also allows us to create the workqueue
618 * late enough so the we can include s_id in the name of the workqueue.
619 */
ec1b8260 620int sb_init_dio_done_wq(struct super_block *sb)
7b7a8665 621{
45150c43 622 struct workqueue_struct *old;
7b7a8665
CH
623 struct workqueue_struct *wq = alloc_workqueue("dio/%s",
624 WQ_MEM_RECLAIM, 0,
625 sb->s_id);
626 if (!wq)
627 return -ENOMEM;
628 /*
629 * This has to be atomic as more DIOs can race to create the workqueue
630 */
45150c43 631 old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
7b7a8665 632 /* Someone created workqueue before us? Free ours... */
45150c43 633 if (old)
7b7a8665
CH
634 destroy_workqueue(wq);
635 return 0;
636}
637
638static int dio_set_defer_completion(struct dio *dio)
639{
640 struct super_block *sb = dio->inode->i_sb;
641
642 if (dio->defer_completion)
643 return 0;
644 dio->defer_completion = true;
645 if (!sb->s_dio_done_wq)
646 return sb_init_dio_done_wq(sb);
647 return 0;
648}
649
1da177e4
LT
650/*
651 * Call into the fs to map some more disk blocks. We record the current number
eb28be2b 652 * of available blocks at sdio->blocks_available. These are in units of the
93407472 653 * fs blocksize, i_blocksize(inode).
1da177e4
LT
654 *
655 * The fs is allowed to map lots of blocks at once. If it wants to do that,
656 * it uses the passed inode-relative block number as the file offset, as usual.
657 *
1d8fa7a2 658 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
1da177e4
LT
659 * has remaining to do. The fs should not map more than this number of blocks.
660 *
661 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
662 * indicate how much contiguous disk space has been made available at
663 * bh->b_blocknr.
664 *
665 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
666 * This isn't very efficient...
667 *
668 * In the case of filesystem holes: the fs may return an arbitrarily-large
669 * hole by returning an appropriate value in b_size and by clearing
670 * buffer_mapped(). However the direct-io code will only process holes one
1d8fa7a2 671 * block at a time - it will repeatedly call get_block() as it walks the hole.
1da177e4 672 */
18772641
AK
673static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
674 struct buffer_head *map_bh)
1da177e4
LT
675{
676 int ret;
1da177e4 677 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
ae55e1aa 678 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
1da177e4 679 unsigned long fs_count; /* Number of filesystem-sized blocks */
1da177e4 680 int create;
ab73857e 681 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
1da177e4
LT
682
683 /*
684 * If there was a memory error and we've overwritten all the
685 * mapped blocks then we can now return that memory error
686 */
687 ret = dio->page_errors;
688 if (ret == 0) {
eb28be2b
AK
689 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
690 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
ae55e1aa
TM
691 fs_endblk = (sdio->final_block_in_request - 1) >>
692 sdio->blkfactor;
693 fs_count = fs_endblk - fs_startblk + 1;
1da177e4 694
3c674e74 695 map_bh->b_state = 0;
ab73857e 696 map_bh->b_size = fs_count << i_blkbits;
3c674e74 697
5fe878ae 698 /*
9ecd10b7
EG
699 * For writes that could fill holes inside i_size on a
700 * DIO_SKIP_HOLES filesystem we forbid block creations: only
701 * overwrites are permitted. We will return early to the caller
702 * once we see an unmapped buffer head returned, and the caller
703 * will fall back to buffered I/O.
5fe878ae
CH
704 *
705 * Otherwise the decision is left to the get_blocks method,
706 * which may decide to handle it or also return an unmapped
707 * buffer head.
708 */
8a4c1e42 709 create = dio->op == REQ_OP_WRITE;
5fe878ae 710 if (dio->flags & DIO_SKIP_HOLES) {
9ecd10b7
EG
711 if (fs_startblk <= ((i_size_read(dio->inode) - 1) >>
712 i_blkbits))
1da177e4 713 create = 0;
1da177e4 714 }
3c674e74 715
eb28be2b 716 ret = (*sdio->get_block)(dio->inode, fs_startblk,
1da177e4 717 map_bh, create);
18772641
AK
718
719 /* Store for completion */
720 dio->private = map_bh->b_private;
7b7a8665
CH
721
722 if (ret == 0 && buffer_defer_completion(map_bh))
723 ret = dio_set_defer_completion(dio);
1da177e4
LT
724 }
725 return ret;
726}
727
728/*
729 * There is no bio. Make one now.
730 */
ba253fbf
AK
731static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
732 sector_t start_sector, struct buffer_head *map_bh)
1da177e4
LT
733{
734 sector_t sector;
735 int ret, nr_pages;
736
eb28be2b 737 ret = dio_bio_reap(dio, sdio);
1da177e4
LT
738 if (ret)
739 goto out;
eb28be2b 740 sector = start_sector << (sdio->blkbits - 9);
b54ffb73 741 nr_pages = min(sdio->pages_in_io, BIO_MAX_PAGES);
1da177e4 742 BUG_ON(nr_pages <= 0);
18772641 743 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
eb28be2b 744 sdio->boundary = 0;
1da177e4
LT
745out:
746 return ret;
747}
748
749/*
750 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
751 * that was successful then update final_block_in_bio and take a ref against
752 * the just-added page.
753 *
754 * Return zero on success. Non-zero means the caller needs to start a new BIO.
755 */
ba253fbf 756static inline int dio_bio_add_page(struct dio_submit *sdio)
1da177e4
LT
757{
758 int ret;
759
eb28be2b
AK
760 ret = bio_add_page(sdio->bio, sdio->cur_page,
761 sdio->cur_page_len, sdio->cur_page_offset);
762 if (ret == sdio->cur_page_len) {
1da177e4
LT
763 /*
764 * Decrement count only, if we are done with this page
765 */
eb28be2b
AK
766 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
767 sdio->pages_in_io--;
09cbfeaf 768 get_page(sdio->cur_page);
eb28be2b
AK
769 sdio->final_block_in_bio = sdio->cur_page_block +
770 (sdio->cur_page_len >> sdio->blkbits);
1da177e4
LT
771 ret = 0;
772 } else {
773 ret = 1;
774 }
775 return ret;
776}
777
778/*
779 * Put cur_page under IO. The section of cur_page which is described by
780 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
781 * starts on-disk at cur_page_block.
782 *
783 * We take a ref against the page here (on behalf of its presence in the bio).
784 *
785 * The caller of this function is responsible for removing cur_page from the
786 * dio, and for dropping the refcount which came from that presence.
787 */
ba253fbf
AK
788static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
789 struct buffer_head *map_bh)
1da177e4
LT
790{
791 int ret = 0;
792
eb28be2b
AK
793 if (sdio->bio) {
794 loff_t cur_offset = sdio->cur_page_fs_offset;
795 loff_t bio_next_offset = sdio->logical_offset_in_bio +
4f024f37 796 sdio->bio->bi_iter.bi_size;
c2c6ca41 797
1da177e4 798 /*
c2c6ca41
JB
799 * See whether this new request is contiguous with the old.
800 *
f0940cee
NK
801 * Btrfs cannot handle having logically non-contiguous requests
802 * submitted. For example if you have
c2c6ca41
JB
803 *
804 * Logical: [0-4095][HOLE][8192-12287]
f0940cee 805 * Physical: [0-4095] [4096-8191]
c2c6ca41
JB
806 *
807 * We cannot submit those pages together as one BIO. So if our
808 * current logical offset in the file does not equal what would
809 * be the next logical offset in the bio, submit the bio we
810 * have.
1da177e4 811 */
eb28be2b 812 if (sdio->final_block_in_bio != sdio->cur_page_block ||
c2c6ca41 813 cur_offset != bio_next_offset)
eb28be2b 814 dio_bio_submit(dio, sdio);
1da177e4
LT
815 }
816
eb28be2b 817 if (sdio->bio == NULL) {
18772641 818 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
1da177e4
LT
819 if (ret)
820 goto out;
821 }
822
eb28be2b
AK
823 if (dio_bio_add_page(sdio) != 0) {
824 dio_bio_submit(dio, sdio);
18772641 825 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
1da177e4 826 if (ret == 0) {
eb28be2b 827 ret = dio_bio_add_page(sdio);
1da177e4
LT
828 BUG_ON(ret != 0);
829 }
830 }
831out:
832 return ret;
833}
834
835/*
836 * An autonomous function to put a chunk of a page under deferred IO.
837 *
838 * The caller doesn't actually know (or care) whether this piece of page is in
839 * a BIO, or is under IO or whatever. We just take care of all possible
840 * situations here. The separation between the logic of do_direct_IO() and
841 * that of submit_page_section() is important for clarity. Please don't break.
842 *
843 * The chunk of page starts on-disk at blocknr.
844 *
845 * We perform deferred IO, by recording the last-submitted page inside our
846 * private part of the dio structure. If possible, we just expand the IO
847 * across that page here.
848 *
849 * If that doesn't work out then we put the old page into the bio and add this
850 * page to the dio instead.
851 */
ba253fbf 852static inline int
eb28be2b 853submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
18772641
AK
854 unsigned offset, unsigned len, sector_t blocknr,
855 struct buffer_head *map_bh)
1da177e4
LT
856{
857 int ret = 0;
858
8a4c1e42 859 if (dio->op == REQ_OP_WRITE) {
98c4d57d
AM
860 /*
861 * Read accounting is performed in submit_bio()
862 */
863 task_io_account_write(len);
864 }
865
1da177e4
LT
866 /*
867 * Can we just grow the current page's presence in the dio?
868 */
eb28be2b
AK
869 if (sdio->cur_page == page &&
870 sdio->cur_page_offset + sdio->cur_page_len == offset &&
871 sdio->cur_page_block +
872 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
873 sdio->cur_page_len += len;
1da177e4
LT
874 goto out;
875 }
876
877 /*
878 * If there's a deferred page already there then send it.
879 */
eb28be2b 880 if (sdio->cur_page) {
18772641 881 ret = dio_send_cur_page(dio, sdio, map_bh);
09cbfeaf 882 put_page(sdio->cur_page);
eb28be2b 883 sdio->cur_page = NULL;
1da177e4 884 if (ret)
b1058b98 885 return ret;
1da177e4
LT
886 }
887
09cbfeaf 888 get_page(page); /* It is in dio */
eb28be2b
AK
889 sdio->cur_page = page;
890 sdio->cur_page_offset = offset;
891 sdio->cur_page_len = len;
892 sdio->cur_page_block = blocknr;
893 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
1da177e4 894out:
b1058b98
JK
895 /*
896 * If sdio->boundary then we want to schedule the IO now to
897 * avoid metadata seeks.
898 */
899 if (sdio->boundary) {
900 ret = dio_send_cur_page(dio, sdio, map_bh);
899f0429
AG
901 if (sdio->bio)
902 dio_bio_submit(dio, sdio);
09cbfeaf 903 put_page(sdio->cur_page);
b1058b98
JK
904 sdio->cur_page = NULL;
905 }
1da177e4
LT
906 return ret;
907}
908
1da177e4
LT
909/*
910 * If we are not writing the entire block and get_block() allocated
911 * the block for us, we need to fill-in the unused portion of the
912 * block with zeros. This happens only if user-buffer, fileoffset or
913 * io length is not filesystem block-size multiple.
914 *
915 * `end' is zero if we're doing the start of the IO, 1 at the end of the
916 * IO.
917 */
ba253fbf
AK
918static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
919 int end, struct buffer_head *map_bh)
1da177e4
LT
920{
921 unsigned dio_blocks_per_fs_block;
922 unsigned this_chunk_blocks; /* In dio_blocks */
923 unsigned this_chunk_bytes;
924 struct page *page;
925
eb28be2b 926 sdio->start_zero_done = 1;
18772641 927 if (!sdio->blkfactor || !buffer_new(map_bh))
1da177e4
LT
928 return;
929
eb28be2b
AK
930 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
931 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
1da177e4
LT
932
933 if (!this_chunk_blocks)
934 return;
935
936 /*
937 * We need to zero out part of an fs block. It is either at the
938 * beginning or the end of the fs block.
939 */
940 if (end)
941 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
942
eb28be2b 943 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
1da177e4 944
557ed1fa 945 page = ZERO_PAGE(0);
eb28be2b 946 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
18772641 947 sdio->next_block_for_io, map_bh))
1da177e4
LT
948 return;
949
eb28be2b 950 sdio->next_block_for_io += this_chunk_blocks;
1da177e4
LT
951}
952
953/*
954 * Walk the user pages, and the file, mapping blocks to disk and generating
955 * a sequence of (page,offset,len,block) mappings. These mappings are injected
956 * into submit_page_section(), which takes care of the next stage of submission
957 *
958 * Direct IO against a blockdev is different from a file. Because we can
959 * happily perform page-sized but 512-byte aligned IOs. It is important that
960 * blockdev IO be able to have fine alignment and large sizes.
961 *
1d8fa7a2 962 * So what we do is to permit the ->get_block function to populate bh.b_size
1da177e4
LT
963 * with the size of IO which is permitted at this offset and this i_blkbits.
964 *
965 * For best results, the blockdev should be set up with 512-byte i_blkbits and
1d8fa7a2 966 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
1da177e4
LT
967 * fine alignment but still allows this function to work in PAGE_SIZE units.
968 */
18772641
AK
969static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
970 struct buffer_head *map_bh)
1da177e4 971{
eb28be2b 972 const unsigned blkbits = sdio->blkbits;
dd545b52 973 const unsigned i_blkbits = blkbits + sdio->blkfactor;
1da177e4
LT
974 int ret = 0;
975
eb28be2b 976 while (sdio->block_in_file < sdio->final_block_in_request) {
7b2c99d1
AV
977 struct page *page;
978 size_t from, to;
6fcc5420
BH
979
980 page = dio_get_page(dio, sdio);
1da177e4
LT
981 if (IS_ERR(page)) {
982 ret = PTR_ERR(page);
983 goto out;
984 }
6fcc5420
BH
985 from = sdio->head ? 0 : sdio->from;
986 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
987 sdio->head++;
1da177e4 988
7b2c99d1 989 while (from < to) {
1da177e4
LT
990 unsigned this_chunk_bytes; /* # of bytes mapped */
991 unsigned this_chunk_blocks; /* # of blocks */
992 unsigned u;
993
eb28be2b 994 if (sdio->blocks_available == 0) {
1da177e4
LT
995 /*
996 * Need to go and map some more disk
997 */
998 unsigned long blkmask;
999 unsigned long dio_remainder;
1000
18772641 1001 ret = get_more_blocks(dio, sdio, map_bh);
1da177e4 1002 if (ret) {
09cbfeaf 1003 put_page(page);
1da177e4
LT
1004 goto out;
1005 }
1006 if (!buffer_mapped(map_bh))
1007 goto do_holes;
1008
eb28be2b 1009 sdio->blocks_available =
f734c89c 1010 map_bh->b_size >> blkbits;
eb28be2b
AK
1011 sdio->next_block_for_io =
1012 map_bh->b_blocknr << sdio->blkfactor;
f734c89c
JK
1013 if (buffer_new(map_bh)) {
1014 clean_bdev_aliases(
1015 map_bh->b_bdev,
1016 map_bh->b_blocknr,
dd545b52 1017 map_bh->b_size >> i_blkbits);
f734c89c 1018 }
1da177e4 1019
eb28be2b 1020 if (!sdio->blkfactor)
1da177e4
LT
1021 goto do_holes;
1022
eb28be2b
AK
1023 blkmask = (1 << sdio->blkfactor) - 1;
1024 dio_remainder = (sdio->block_in_file & blkmask);
1da177e4
LT
1025
1026 /*
1027 * If we are at the start of IO and that IO
1028 * starts partway into a fs-block,
1029 * dio_remainder will be non-zero. If the IO
1030 * is a read then we can simply advance the IO
1031 * cursor to the first block which is to be
1032 * read. But if the IO is a write and the
1033 * block was newly allocated we cannot do that;
1034 * the start of the fs block must be zeroed out
1035 * on-disk
1036 */
1037 if (!buffer_new(map_bh))
eb28be2b
AK
1038 sdio->next_block_for_io += dio_remainder;
1039 sdio->blocks_available -= dio_remainder;
1da177e4
LT
1040 }
1041do_holes:
1042 /* Handle holes */
1043 if (!buffer_mapped(map_bh)) {
35dc8161 1044 loff_t i_size_aligned;
1da177e4
LT
1045
1046 /* AKPM: eargh, -ENOTBLK is a hack */
8a4c1e42 1047 if (dio->op == REQ_OP_WRITE) {
09cbfeaf 1048 put_page(page);
1da177e4
LT
1049 return -ENOTBLK;
1050 }
1051
35dc8161
JM
1052 /*
1053 * Be sure to account for a partial block as the
1054 * last block in the file
1055 */
1056 i_size_aligned = ALIGN(i_size_read(dio->inode),
1057 1 << blkbits);
eb28be2b 1058 if (sdio->block_in_file >=
35dc8161 1059 i_size_aligned >> blkbits) {
1da177e4 1060 /* We hit eof */
09cbfeaf 1061 put_page(page);
1da177e4
LT
1062 goto out;
1063 }
7b2c99d1 1064 zero_user(page, from, 1 << blkbits);
eb28be2b 1065 sdio->block_in_file++;
7b2c99d1 1066 from += 1 << blkbits;
3320c60b 1067 dio->result += 1 << blkbits;
1da177e4
LT
1068 goto next_block;
1069 }
1070
1071 /*
1072 * If we're performing IO which has an alignment which
1073 * is finer than the underlying fs, go check to see if
1074 * we must zero out the start of this block.
1075 */
eb28be2b 1076 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
18772641 1077 dio_zero_block(dio, sdio, 0, map_bh);
1da177e4
LT
1078
1079 /*
1080 * Work out, in this_chunk_blocks, how much disk we
1081 * can add to this page
1082 */
eb28be2b 1083 this_chunk_blocks = sdio->blocks_available;
7b2c99d1 1084 u = (to - from) >> blkbits;
1da177e4
LT
1085 if (this_chunk_blocks > u)
1086 this_chunk_blocks = u;
eb28be2b 1087 u = sdio->final_block_in_request - sdio->block_in_file;
1da177e4
LT
1088 if (this_chunk_blocks > u)
1089 this_chunk_blocks = u;
1090 this_chunk_bytes = this_chunk_blocks << blkbits;
1091 BUG_ON(this_chunk_bytes == 0);
1092
092c8d46
JK
1093 if (this_chunk_blocks == sdio->blocks_available)
1094 sdio->boundary = buffer_boundary(map_bh);
eb28be2b 1095 ret = submit_page_section(dio, sdio, page,
7b2c99d1 1096 from,
eb28be2b 1097 this_chunk_bytes,
18772641
AK
1098 sdio->next_block_for_io,
1099 map_bh);
1da177e4 1100 if (ret) {
09cbfeaf 1101 put_page(page);
1da177e4
LT
1102 goto out;
1103 }
eb28be2b 1104 sdio->next_block_for_io += this_chunk_blocks;
1da177e4 1105
eb28be2b 1106 sdio->block_in_file += this_chunk_blocks;
7b2c99d1
AV
1107 from += this_chunk_bytes;
1108 dio->result += this_chunk_bytes;
eb28be2b 1109 sdio->blocks_available -= this_chunk_blocks;
1da177e4 1110next_block:
eb28be2b
AK
1111 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1112 if (sdio->block_in_file == sdio->final_block_in_request)
1da177e4
LT
1113 break;
1114 }
1115
1116 /* Drop the ref which was taken in get_user_pages() */
09cbfeaf 1117 put_page(page);
1da177e4
LT
1118 }
1119out:
1120 return ret;
1121}
1122
847cc637 1123static inline int drop_refcount(struct dio *dio)
1da177e4 1124{
847cc637 1125 int ret2;
5eb6c7a2 1126 unsigned long flags;
1da177e4 1127
8459d86a
ZB
1128 /*
1129 * Sync will always be dropping the final ref and completing the
5eb6c7a2
ZB
1130 * operation. AIO can if it was a broken operation described above or
1131 * in fact if all the bios race to complete before we get here. In
1132 * that case dio_complete() translates the EIOCBQUEUED into the proper
04b2fa9f 1133 * return code that the caller will hand to ->complete().
5eb6c7a2
ZB
1134 *
1135 * This is managed by the bio_lock instead of being an atomic_t so that
1136 * completion paths can drop their ref and use the remaining count to
1137 * decide to wake the submission path atomically.
8459d86a 1138 */
5eb6c7a2
ZB
1139 spin_lock_irqsave(&dio->bio_lock, flags);
1140 ret2 = --dio->refcount;
1141 spin_unlock_irqrestore(&dio->bio_lock, flags);
847cc637 1142 return ret2;
1da177e4
LT
1143}
1144
eafdc7d1
CH
1145/*
1146 * This is a library function for use by filesystem drivers.
1147 *
1148 * The locking rules are governed by the flags parameter:
1149 * - if the flags value contains DIO_LOCKING we use a fancy locking
1150 * scheme for dumb filesystems.
1151 * For writes this function is called under i_mutex and returns with
1152 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1153 * taken and dropped again before returning.
eafdc7d1
CH
1154 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1155 * internal locking but rather rely on the filesystem to synchronize
1156 * direct I/O reads/writes versus each other and truncate.
df2d6f26
CH
1157 *
1158 * To help with locking against truncate we incremented the i_dio_count
1159 * counter before starting direct I/O, and decrement it once we are done.
1160 * Truncate can wait for it to reach zero to provide exclusion. It is
1161 * expected that filesystem provide exclusion between new direct I/O
1162 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1163 * but other filesystems need to take care of this on their own.
ba253fbf
AK
1164 *
1165 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1166 * is always inlined. Otherwise gcc is unable to split the structure into
1167 * individual fields and will generate much worse code. This is important
1168 * for the whole file.
eafdc7d1 1169 */
65dd2aa9 1170static inline ssize_t
17f8c842
OS
1171do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1172 struct block_device *bdev, struct iov_iter *iter,
c8b8e32d 1173 get_block_t get_block, dio_iodone_t end_io,
17f8c842 1174 dio_submit_t submit_io, int flags)
1da177e4 1175{
6aa7de05 1176 unsigned i_blkbits = READ_ONCE(inode->i_blkbits);
ab73857e 1177 unsigned blkbits = i_blkbits;
1da177e4
LT
1178 unsigned blocksize_mask = (1 << blkbits) - 1;
1179 ssize_t retval = -EINVAL;
1c0ff0f1 1180 const size_t count = iov_iter_count(iter);
c8b8e32d 1181 loff_t offset = iocb->ki_pos;
1c0ff0f1 1182 const loff_t end = offset + count;
1da177e4 1183 struct dio *dio;
eb28be2b 1184 struct dio_submit sdio = { 0, };
847cc637 1185 struct buffer_head map_bh = { 0, };
647d1e4c 1186 struct blk_plug plug;
886a3911 1187 unsigned long align = offset | iov_iter_alignment(iter);
1da177e4 1188
65dd2aa9
AK
1189 /*
1190 * Avoid references to bdev if not absolutely needed to give
1191 * the early prefetch in the caller enough time.
1192 */
1da177e4 1193
886a3911 1194 if (align & blocksize_mask) {
1da177e4 1195 if (bdev)
65dd2aa9 1196 blkbits = blksize_bits(bdev_logical_block_size(bdev));
1da177e4 1197 blocksize_mask = (1 << blkbits) - 1;
886a3911 1198 if (align & blocksize_mask)
1da177e4
LT
1199 goto out;
1200 }
1201
f9b5570d 1202 /* watch out for a 0 len io from a tricksy fs */
1c0ff0f1 1203 if (iov_iter_rw(iter) == READ && !count)
f9b5570d
CH
1204 return 0;
1205
6e8267f5 1206 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
1da177e4
LT
1207 retval = -ENOMEM;
1208 if (!dio)
1209 goto out;
23aee091
JM
1210 /*
1211 * Believe it or not, zeroing out the page array caused a .5%
1212 * performance regression in a database benchmark. So, we take
1213 * care to only zero out what's needed.
1214 */
1215 memset(dio, 0, offsetof(struct dio, pages));
1da177e4 1216
5fe878ae
CH
1217 dio->flags = flags;
1218 if (dio->flags & DIO_LOCKING) {
17f8c842 1219 if (iov_iter_rw(iter) == READ) {
5fe878ae
CH
1220 struct address_space *mapping =
1221 iocb->ki_filp->f_mapping;
1da177e4 1222
5fe878ae 1223 /* will be released by direct_io_worker */
5955102c 1224 inode_lock(inode);
1da177e4
LT
1225
1226 retval = filemap_write_and_wait_range(mapping, offset,
1227 end - 1);
1228 if (retval) {
5955102c 1229 inode_unlock(inode);
6e8267f5 1230 kmem_cache_free(dio_cache, dio);
1da177e4
LT
1231 goto out;
1232 }
1da177e4 1233 }
1da177e4
LT
1234 }
1235
74cedf9b
JK
1236 /* Once we sampled i_size check for reads beyond EOF */
1237 dio->i_size = i_size_read(inode);
1238 if (iov_iter_rw(iter) == READ && offset >= dio->i_size) {
1239 if (dio->flags & DIO_LOCKING)
5955102c 1240 inode_unlock(inode);
74cedf9b 1241 kmem_cache_free(dio_cache, dio);
2d4594ac 1242 retval = 0;
74cedf9b
JK
1243 goto out;
1244 }
1245
1da177e4 1246 /*
60392573
CH
1247 * For file extending writes updating i_size before data writeouts
1248 * complete can expose uninitialized blocks in dumb filesystems.
1249 * In that case we need to wait for I/O completion even if asked
1250 * for an asynchronous write.
1da177e4 1251 */
60392573
CH
1252 if (is_sync_kiocb(iocb))
1253 dio->is_async = false;
c8f4c36f 1254 else if (iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
60392573
CH
1255 dio->is_async = false;
1256 else
1257 dio->is_async = true;
1258
847cc637 1259 dio->inode = inode;
8a4c1e42
MC
1260 if (iov_iter_rw(iter) == WRITE) {
1261 dio->op = REQ_OP_WRITE;
70fd7614 1262 dio->op_flags = REQ_SYNC | REQ_IDLE;
03a07c92
GR
1263 if (iocb->ki_flags & IOCB_NOWAIT)
1264 dio->op_flags |= REQ_NOWAIT;
8a4c1e42
MC
1265 } else {
1266 dio->op = REQ_OP_READ;
1267 }
02afc27f
CH
1268
1269 /*
1270 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1271 * so that we can call ->fsync.
1272 */
332391a9
LC
1273 if (dio->is_async && iov_iter_rw(iter) == WRITE) {
1274 retval = 0;
d9c10e5b 1275 if (iocb->ki_flags & IOCB_DSYNC)
332391a9
LC
1276 retval = dio_set_defer_completion(dio);
1277 else if (!dio->inode->i_sb->s_dio_done_wq) {
1278 /*
1279 * In case of AIO write racing with buffered read we
1280 * need to defer completion. We can't decide this now,
1281 * however the workqueue needs to be initialized here.
1282 */
1283 retval = sb_init_dio_done_wq(dio->inode->i_sb);
1284 }
02afc27f
CH
1285 if (retval) {
1286 /*
1287 * We grab i_mutex only for reads so we don't have
1288 * to release it here
1289 */
1290 kmem_cache_free(dio_cache, dio);
1291 goto out;
1292 }
1293 }
1294
1295 /*
1296 * Will be decremented at I/O completion time.
1297 */
ce3077ee 1298 inode_dio_begin(inode);
02afc27f
CH
1299
1300 retval = 0;
847cc637 1301 sdio.blkbits = blkbits;
ab73857e 1302 sdio.blkfactor = i_blkbits - blkbits;
847cc637
AK
1303 sdio.block_in_file = offset >> blkbits;
1304
1305 sdio.get_block = get_block;
1306 dio->end_io = end_io;
1307 sdio.submit_io = submit_io;
1308 sdio.final_block_in_bio = -1;
1309 sdio.next_block_for_io = -1;
1310
1311 dio->iocb = iocb;
847cc637
AK
1312
1313 spin_lock_init(&dio->bio_lock);
1314 dio->refcount = 1;
1315
53cbf3b1 1316 dio->should_dirty = (iter->type == ITER_IOVEC);
7b2c99d1 1317 sdio.iter = iter;
1c0ff0f1 1318 sdio.final_block_in_request = end >> blkbits;
7b2c99d1 1319
847cc637
AK
1320 /*
1321 * In case of non-aligned buffers, we may need 2 more
1322 * pages since we need to zero out first and last block.
1323 */
1324 if (unlikely(sdio.blkfactor))
1325 sdio.pages_in_io = 2;
1326
f67da30c 1327 sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
847cc637 1328
647d1e4c
FW
1329 blk_start_plug(&plug);
1330
7b2c99d1
AV
1331 retval = do_direct_IO(dio, &sdio, &map_bh);
1332 if (retval)
1333 dio_cleanup(dio, &sdio);
847cc637
AK
1334
1335 if (retval == -ENOTBLK) {
1336 /*
1337 * The remaining part of the request will be
1338 * be handled by buffered I/O when we return
1339 */
1340 retval = 0;
1341 }
1342 /*
1343 * There may be some unwritten disk at the end of a part-written
1344 * fs-block-sized block. Go zero that now.
1345 */
1346 dio_zero_block(dio, &sdio, 1, &map_bh);
1347
1348 if (sdio.cur_page) {
1349 ssize_t ret2;
1350
1351 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1352 if (retval == 0)
1353 retval = ret2;
09cbfeaf 1354 put_page(sdio.cur_page);
847cc637
AK
1355 sdio.cur_page = NULL;
1356 }
1357 if (sdio.bio)
1358 dio_bio_submit(dio, &sdio);
1359
647d1e4c
FW
1360 blk_finish_plug(&plug);
1361
847cc637
AK
1362 /*
1363 * It is possible that, we return short IO due to end of file.
1364 * In that case, we need to release all the pages we got hold on.
1365 */
1366 dio_cleanup(dio, &sdio);
1367
1368 /*
1369 * All block lookups have been performed. For READ requests
1370 * we can let i_mutex go now that its achieved its purpose
1371 * of protecting us from looking up uninitialized blocks.
1372 */
17f8c842 1373 if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING))
5955102c 1374 inode_unlock(dio->inode);
847cc637
AK
1375
1376 /*
1377 * The only time we want to leave bios in flight is when a successful
1378 * partial aio read or full aio write have been setup. In that case
1379 * bio completion will call aio_complete. The only time it's safe to
1380 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1381 * This had *better* be the only place that raises -EIOCBQUEUED.
1382 */
1383 BUG_ON(retval == -EIOCBQUEUED);
1384 if (dio->is_async && retval == 0 && dio->result &&
17f8c842 1385 (iov_iter_rw(iter) == READ || dio->result == count))
847cc637 1386 retval = -EIOCBQUEUED;
af436472 1387 else
847cc637
AK
1388 dio_await_completion(dio);
1389
1390 if (drop_refcount(dio) == 0) {
ffe51f01 1391 retval = dio_complete(dio, retval, DIO_COMPLETE_INVALIDATE);
847cc637
AK
1392 } else
1393 BUG_ON(retval != -EIOCBQUEUED);
1da177e4 1394
7bb46a67 1395out:
1396 return retval;
1397}
65dd2aa9 1398
17f8c842
OS
1399ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1400 struct block_device *bdev, struct iov_iter *iter,
c8b8e32d 1401 get_block_t get_block,
17f8c842
OS
1402 dio_iodone_t end_io, dio_submit_t submit_io,
1403 int flags)
65dd2aa9
AK
1404{
1405 /*
1406 * The block device state is needed in the end to finally
1407 * submit everything. Since it's likely to be cache cold
1408 * prefetch it here as first thing to hide some of the
1409 * latency.
1410 *
1411 * Attempt to prefetch the pieces we likely need later.
1412 */
1413 prefetch(&bdev->bd_disk->part_tbl);
1414 prefetch(bdev->bd_queue);
1415 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1416
c8b8e32d 1417 return do_blockdev_direct_IO(iocb, inode, bdev, iter, get_block,
17f8c842 1418 end_io, submit_io, flags);
65dd2aa9
AK
1419}
1420
1da177e4 1421EXPORT_SYMBOL(__blockdev_direct_IO);
6e8267f5
AK
1422
1423static __init int dio_init(void)
1424{
1425 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1426 return 0;
1427}
1428module_init(dio_init)