Commit | Line | Data |
---|---|---|
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 | ||
70 | struct 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 */ |
facd07b0 | 89 | dio_submit_t *submit_io; /* IO submition function */ |
eb28be2b | 90 | |
facd07b0 | 91 | loff_t logical_offset_in_bio; /* current first logical block in bio */ |
1da177e4 LT |
92 | sector_t final_block_in_bio; /* current final block in bio + 1 */ |
93 | sector_t next_block_for_io; /* next block to be put under IO, | |
94 | in dio_blocks units */ | |
1da177e4 LT |
95 | |
96 | /* | |
97 | * Deferred addition of a page to the dio. These variables are | |
98 | * private to dio_send_cur_page(), submit_page_section() and | |
99 | * dio_bio_add_page(). | |
100 | */ | |
101 | struct page *cur_page; /* The page */ | |
102 | unsigned cur_page_offset; /* Offset into it, in bytes */ | |
103 | unsigned cur_page_len; /* Nr of bytes at cur_page_offset */ | |
104 | sector_t cur_page_block; /* Where it starts */ | |
facd07b0 | 105 | loff_t cur_page_fs_offset; /* Offset in file */ |
1da177e4 | 106 | |
7b2c99d1 | 107 | struct iov_iter *iter; |
1da177e4 LT |
108 | /* |
109 | * Page queue. These variables belong to dio_refill_pages() and | |
110 | * dio_get_page(). | |
111 | */ | |
1da177e4 LT |
112 | unsigned head; /* next page to process */ |
113 | unsigned tail; /* last valid page + 1 */ | |
7b2c99d1 | 114 | size_t from, to; |
eb28be2b AK |
115 | }; |
116 | ||
117 | /* dio_state communicated between submission path and end_io */ | |
118 | struct dio { | |
119 | int flags; /* doesn't change */ | |
8a4c1e42 MC |
120 | int op; |
121 | int op_flags; | |
74d46992 | 122 | struct gendisk *bio_disk; |
0dc2bc49 | 123 | struct inode *inode; |
eb28be2b AK |
124 | loff_t i_size; /* i_size when submitted */ |
125 | dio_iodone_t *end_io; /* IO completion function */ | |
eb28be2b | 126 | |
18772641 | 127 | void *private; /* copy from map_bh.b_private */ |
eb28be2b AK |
128 | |
129 | /* BIO completion state */ | |
130 | spinlock_t bio_lock; /* protects BIO fields below */ | |
0dc2bc49 AK |
131 | int page_errors; /* errno from get_user_pages() */ |
132 | int is_async; /* is IO async ? */ | |
7b7a8665 | 133 | bool defer_completion; /* defer AIO completion to workqueue? */ |
53cbf3b1 | 134 | bool should_dirty; /* if pages should be dirtied */ |
0dc2bc49 | 135 | int io_error; /* IO error in completion path */ |
eb28be2b AK |
136 | unsigned long refcount; /* direct_io_worker() and bios */ |
137 | struct bio *bio_list; /* singly linked via bi_private */ | |
138 | struct task_struct *waiter; /* waiting task (NULL if none) */ | |
139 | ||
140 | /* AIO related stuff */ | |
141 | struct kiocb *iocb; /* kiocb */ | |
eb28be2b AK |
142 | ssize_t result; /* IO result */ |
143 | ||
23aee091 JM |
144 | /* |
145 | * pages[] (and any fields placed after it) are not zeroed out at | |
146 | * allocation time. Don't add new fields after pages[] unless you | |
147 | * wish that they not be zeroed. | |
148 | */ | |
7b7a8665 CH |
149 | union { |
150 | struct page *pages[DIO_PAGES]; /* page buffer */ | |
151 | struct work_struct complete_work;/* deferred AIO completion */ | |
152 | }; | |
6e8267f5 AK |
153 | } ____cacheline_aligned_in_smp; |
154 | ||
155 | static struct kmem_cache *dio_cache __read_mostly; | |
1da177e4 LT |
156 | |
157 | /* | |
158 | * How many pages are in the queue? | |
159 | */ | |
eb28be2b | 160 | static inline unsigned dio_pages_present(struct dio_submit *sdio) |
1da177e4 | 161 | { |
eb28be2b | 162 | return sdio->tail - sdio->head; |
1da177e4 LT |
163 | } |
164 | ||
165 | /* | |
166 | * Go grab and pin some userspace pages. Typically we'll get 64 at a time. | |
167 | */ | |
ba253fbf | 168 | static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio) |
1da177e4 | 169 | { |
7b2c99d1 | 170 | ssize_t ret; |
1da177e4 | 171 | |
2c80929c | 172 | ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES, |
7b2c99d1 | 173 | &sdio->from); |
1da177e4 | 174 | |
8a4c1e42 | 175 | if (ret < 0 && sdio->blocks_available && (dio->op == REQ_OP_WRITE)) { |
557ed1fa | 176 | struct page *page = ZERO_PAGE(0); |
1da177e4 LT |
177 | /* |
178 | * A memory fault, but the filesystem has some outstanding | |
179 | * mapped blocks. We need to use those blocks up to avoid | |
180 | * leaking stale data in the file. | |
181 | */ | |
182 | if (dio->page_errors == 0) | |
183 | dio->page_errors = ret; | |
09cbfeaf | 184 | get_page(page); |
b5810039 | 185 | dio->pages[0] = page; |
eb28be2b AK |
186 | sdio->head = 0; |
187 | sdio->tail = 1; | |
7b2c99d1 AV |
188 | sdio->from = 0; |
189 | sdio->to = PAGE_SIZE; | |
190 | return 0; | |
1da177e4 LT |
191 | } |
192 | ||
193 | if (ret >= 0) { | |
7b2c99d1 AV |
194 | iov_iter_advance(sdio->iter, ret); |
195 | ret += sdio->from; | |
eb28be2b | 196 | sdio->head = 0; |
7b2c99d1 AV |
197 | sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE; |
198 | sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1; | |
199 | return 0; | |
1da177e4 | 200 | } |
1da177e4 LT |
201 | return ret; |
202 | } | |
203 | ||
204 | /* | |
205 | * Get another userspace page. Returns an ERR_PTR on error. Pages are | |
206 | * buffered inside the dio so that we can call get_user_pages() against a | |
207 | * decent number of pages, less frequently. To provide nicer use of the | |
208 | * L1 cache. | |
209 | */ | |
ba253fbf | 210 | static inline struct page *dio_get_page(struct dio *dio, |
6fcc5420 | 211 | struct dio_submit *sdio) |
1da177e4 | 212 | { |
eb28be2b | 213 | if (dio_pages_present(sdio) == 0) { |
1da177e4 LT |
214 | int ret; |
215 | ||
eb28be2b | 216 | ret = dio_refill_pages(dio, sdio); |
1da177e4 LT |
217 | if (ret) |
218 | return ERR_PTR(ret); | |
eb28be2b | 219 | BUG_ON(dio_pages_present(sdio) == 0); |
1da177e4 | 220 | } |
6fcc5420 | 221 | return dio->pages[sdio->head]; |
1da177e4 LT |
222 | } |
223 | ||
c70d868f | 224 | /* |
6d544bb4 | 225 | * dio_complete() - called when all DIO BIO I/O has been completed |
6d544bb4 | 226 | * |
7b7a8665 CH |
227 | * This drops i_dio_count, lets interested parties know that a DIO operation |
228 | * has completed, and calculates the resulting return code for the operation. | |
6d544bb4 ZB |
229 | * |
230 | * It lets the filesystem know if it registered an interest earlier via | |
231 | * get_block. Pass the private field of the map buffer_head so that | |
232 | * filesystems can use it to hold additional state between get_block calls and | |
233 | * dio_complete. | |
1da177e4 | 234 | */ |
ffe51f01 | 235 | static ssize_t dio_complete(struct dio *dio, ssize_t ret, unsigned int flags) |
1da177e4 | 236 | { |
716b9bc0 | 237 | loff_t offset = dio->iocb->ki_pos; |
6d544bb4 | 238 | ssize_t transferred = 0; |
332391a9 | 239 | int err; |
6d544bb4 | 240 | |
8459d86a ZB |
241 | /* |
242 | * AIO submission can race with bio completion to get here while | |
243 | * expecting to have the last io completed by bio completion. | |
244 | * In that case -EIOCBQUEUED is in fact not an error we want | |
245 | * to preserve through this call. | |
246 | */ | |
247 | if (ret == -EIOCBQUEUED) | |
248 | ret = 0; | |
249 | ||
6d544bb4 ZB |
250 | if (dio->result) { |
251 | transferred = dio->result; | |
252 | ||
253 | /* Check for short read case */ | |
8a4c1e42 MC |
254 | if ((dio->op == REQ_OP_READ) && |
255 | ((offset + transferred) > dio->i_size)) | |
6d544bb4 | 256 | transferred = dio->i_size - offset; |
4038acdb AV |
257 | /* ignore EFAULT if some IO has been done */ |
258 | if (unlikely(ret == -EFAULT) && transferred) | |
259 | ret = 0; | |
6d544bb4 ZB |
260 | } |
261 | ||
6d544bb4 ZB |
262 | if (ret == 0) |
263 | ret = dio->page_errors; | |
264 | if (ret == 0) | |
265 | ret = dio->io_error; | |
266 | if (ret == 0) | |
267 | ret = transferred; | |
268 | ||
5e25c269 EG |
269 | if (dio->end_io) { |
270 | // XXX: ki_pos?? | |
271 | err = dio->end_io(dio->iocb, offset, ret, dio->private); | |
272 | if (err) | |
273 | ret = err; | |
274 | } | |
275 | ||
332391a9 LC |
276 | /* |
277 | * Try again to invalidate clean pages which might have been cached by | |
278 | * non-direct readahead, or faulted in by get_user_pages() if the source | |
279 | * of the write was an mmap'ed region of the file we're writing. Either | |
280 | * one is a pretty crazy thing to do, so we don't support it 100%. If | |
281 | * this invalidation fails, tough, the write still worked... | |
5e25c269 EG |
282 | * |
283 | * And this page cache invalidation has to be after dio->end_io(), as | |
284 | * some filesystems convert unwritten extents to real allocations in | |
285 | * end_io() when necessary, otherwise a racing buffer read would cache | |
286 | * zeros from unwritten extents. | |
332391a9 | 287 | */ |
ffe51f01 LC |
288 | if (flags & DIO_COMPLETE_INVALIDATE && |
289 | ret > 0 && dio->op == REQ_OP_WRITE && | |
332391a9 LC |
290 | dio->inode->i_mapping->nrpages) { |
291 | err = invalidate_inode_pages2_range(dio->inode->i_mapping, | |
292 | offset >> PAGE_SHIFT, | |
293 | (offset + ret - 1) >> PAGE_SHIFT); | |
5a9d929d DW |
294 | if (err) |
295 | dio_warn_stale_pagecache(dio->iocb->ki_filp); | |
332391a9 LC |
296 | } |
297 | ||
ce3077ee | 298 | inode_dio_end(dio->inode); |
fe0f07d0 | 299 | |
ffe51f01 | 300 | if (flags & DIO_COMPLETE_ASYNC) { |
e2592217 CH |
301 | /* |
302 | * generic_write_sync expects ki_pos to have been updated | |
303 | * already, but the submission path only does this for | |
304 | * synchronous I/O. | |
305 | */ | |
306 | dio->iocb->ki_pos += transferred; | |
02afc27f | 307 | |
41e817bc MH |
308 | if (ret > 0 && dio->op == REQ_OP_WRITE) |
309 | ret = generic_write_sync(dio->iocb, ret); | |
6b19b766 | 310 | dio->iocb->ki_complete(dio->iocb, ret); |
02afc27f | 311 | } |
40e2e973 | 312 | |
7b7a8665 | 313 | kmem_cache_free(dio_cache, dio); |
6d544bb4 | 314 | return ret; |
1da177e4 LT |
315 | } |
316 | ||
7b7a8665 CH |
317 | static void dio_aio_complete_work(struct work_struct *work) |
318 | { | |
319 | struct dio *dio = container_of(work, struct dio, complete_work); | |
320 | ||
ffe51f01 | 321 | dio_complete(dio, 0, DIO_COMPLETE_ASYNC | DIO_COMPLETE_INVALIDATE); |
7b7a8665 CH |
322 | } |
323 | ||
4e4cbee9 | 324 | static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio); |
7b7a8665 | 325 | |
1da177e4 LT |
326 | /* |
327 | * Asynchronous IO callback. | |
328 | */ | |
4246a0b6 | 329 | static void dio_bio_end_aio(struct bio *bio) |
1da177e4 LT |
330 | { |
331 | struct dio *dio = bio->bi_private; | |
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 || | |
356 | (dio->op == REQ_OP_WRITE && | |
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 | 375 | static 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 | 388 | static inline void |
eb28be2b AK |
389 | dio_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 | */ |
1da177e4 | 399 | bio = bio_alloc(GFP_KERNEL, nr_vecs); |
1da177e4 | 400 | |
74d46992 | 401 | bio_set_dev(bio, bdev); |
4f024f37 | 402 | bio->bi_iter.bi_sector = first_sector; |
8a4c1e42 | 403 | bio_set_op_attrs(bio, dio->op, dio->op_flags); |
1da177e4 LT |
404 | if (dio->is_async) |
405 | bio->bi_end_io = dio_bio_end_aio; | |
406 | else | |
407 | bio->bi_end_io = dio_bio_end_io; | |
408 | ||
45d06cf7 JA |
409 | bio->bi_write_hint = dio->iocb->ki_hint; |
410 | ||
eb28be2b AK |
411 | sdio->bio = bio; |
412 | sdio->logical_offset_in_bio = sdio->cur_page_fs_offset; | |
1da177e4 LT |
413 | } |
414 | ||
415 | /* | |
416 | * In the AIO read case we speculatively dirty the pages before starting IO. | |
417 | * During IO completion, any of these pages which happen to have been written | |
418 | * back will be redirtied by bio_check_pages_dirty(). | |
0273201e ZB |
419 | * |
420 | * bios hold a dio reference between submit_bio and ->end_io. | |
1da177e4 | 421 | */ |
ba253fbf | 422 | static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio) |
1da177e4 | 423 | { |
eb28be2b | 424 | struct bio *bio = sdio->bio; |
5eb6c7a2 | 425 | unsigned long flags; |
1da177e4 LT |
426 | |
427 | bio->bi_private = dio; | |
0cf41e5e PB |
428 | /* don't account direct I/O as memory stall */ |
429 | bio_clear_flag(bio, BIO_WORKINGSET); | |
5eb6c7a2 ZB |
430 | |
431 | spin_lock_irqsave(&dio->bio_lock, flags); | |
432 | dio->refcount++; | |
433 | spin_unlock_irqrestore(&dio->bio_lock, flags); | |
434 | ||
8a4c1e42 | 435 | if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty) |
1da177e4 | 436 | bio_set_pages_dirty(bio); |
5eb6c7a2 | 437 | |
309dca30 | 438 | dio->bio_disk = bio->bi_bdev->bd_disk; |
c1c53460 | 439 | |
94c2ed58 | 440 | if (sdio->submit_io) |
8a4c1e42 | 441 | sdio->submit_io(bio, dio->inode, sdio->logical_offset_in_bio); |
94c2ed58 CH |
442 | else |
443 | submit_bio(bio); | |
1da177e4 | 444 | |
eb28be2b AK |
445 | sdio->bio = NULL; |
446 | sdio->boundary = 0; | |
447 | sdio->logical_offset_in_bio = 0; | |
1da177e4 LT |
448 | } |
449 | ||
450 | /* | |
451 | * Release any resources in case of a failure | |
452 | */ | |
ba253fbf | 453 | static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio) |
1da177e4 | 454 | { |
7b2c99d1 | 455 | while (sdio->head < sdio->tail) |
09cbfeaf | 456 | put_page(dio->pages[sdio->head++]); |
1da177e4 LT |
457 | } |
458 | ||
459 | /* | |
0273201e ZB |
460 | * Wait for the next BIO to complete. Remove it and return it. NULL is |
461 | * returned once all BIOs have been completed. This must only be called once | |
462 | * all bios have been issued so that dio->refcount can only decrease. This | |
3d742d4b | 463 | * requires that the caller hold a reference on the dio. |
1da177e4 LT |
464 | */ |
465 | static struct bio *dio_await_one(struct dio *dio) | |
466 | { | |
467 | unsigned long flags; | |
0273201e | 468 | struct bio *bio = NULL; |
1da177e4 LT |
469 | |
470 | spin_lock_irqsave(&dio->bio_lock, flags); | |
5eb6c7a2 ZB |
471 | |
472 | /* | |
473 | * Wait as long as the list is empty and there are bios in flight. bio | |
474 | * completion drops the count, maybe adds to the list, and wakes while | |
475 | * holding the bio_lock so we don't need set_current_state()'s barrier | |
476 | * and can call it after testing our condition. | |
477 | */ | |
478 | while (dio->refcount > 1 && dio->bio_list == NULL) { | |
479 | __set_current_state(TASK_UNINTERRUPTIBLE); | |
480 | dio->waiter = current; | |
481 | spin_unlock_irqrestore(&dio->bio_lock, flags); | |
94c2ed58 | 482 | blk_io_schedule(); |
5eb6c7a2 ZB |
483 | /* wake up sets us TASK_RUNNING */ |
484 | spin_lock_irqsave(&dio->bio_lock, flags); | |
485 | dio->waiter = NULL; | |
1da177e4 | 486 | } |
0273201e ZB |
487 | if (dio->bio_list) { |
488 | bio = dio->bio_list; | |
489 | dio->bio_list = bio->bi_private; | |
490 | } | |
1da177e4 LT |
491 | spin_unlock_irqrestore(&dio->bio_lock, flags); |
492 | return bio; | |
493 | } | |
494 | ||
495 | /* | |
496 | * Process one completed BIO. No locks are held. | |
497 | */ | |
4e4cbee9 | 498 | static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio) |
1da177e4 | 499 | { |
4e4cbee9 | 500 | blk_status_t err = bio->bi_status; |
d7c8aa85 | 501 | bool should_dirty = dio->op == REQ_OP_READ && dio->should_dirty; |
1da177e4 | 502 | |
03a07c92 GR |
503 | if (err) { |
504 | if (err == BLK_STS_AGAIN && (bio->bi_opf & REQ_NOWAIT)) | |
505 | dio->io_error = -EAGAIN; | |
506 | else | |
507 | dio->io_error = -EIO; | |
508 | } | |
1da177e4 | 509 | |
d7c8aa85 | 510 | if (dio->is_async && should_dirty) { |
7ddc971f | 511 | bio_check_pages_dirty(bio); /* transfers ownership */ |
1da177e4 | 512 | } else { |
d7c8aa85 | 513 | bio_release_pages(bio, should_dirty); |
1da177e4 LT |
514 | bio_put(bio); |
515 | } | |
9b81c842 | 516 | return err; |
1da177e4 LT |
517 | } |
518 | ||
519 | /* | |
0273201e ZB |
520 | * Wait on and process all in-flight BIOs. This must only be called once |
521 | * all bios have been issued so that the refcount can only decrease. | |
522 | * This just waits for all bios to make it through dio_bio_complete. IO | |
beb7dd86 | 523 | * errors are propagated through dio->io_error and should be propagated via |
0273201e | 524 | * dio_complete(). |
1da177e4 | 525 | */ |
6d544bb4 | 526 | static void dio_await_completion(struct dio *dio) |
1da177e4 | 527 | { |
0273201e ZB |
528 | struct bio *bio; |
529 | do { | |
530 | bio = dio_await_one(dio); | |
531 | if (bio) | |
532 | dio_bio_complete(dio, bio); | |
533 | } while (bio); | |
1da177e4 LT |
534 | } |
535 | ||
536 | /* | |
537 | * A really large O_DIRECT read or write can generate a lot of BIOs. So | |
538 | * to keep the memory consumption sane we periodically reap any completed BIOs | |
539 | * during the BIO generation phase. | |
540 | * | |
541 | * This also helps to limit the peak amount of pinned userspace memory. | |
542 | */ | |
ba253fbf | 543 | static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio) |
1da177e4 LT |
544 | { |
545 | int ret = 0; | |
546 | ||
eb28be2b | 547 | if (sdio->reap_counter++ >= 64) { |
1da177e4 LT |
548 | while (dio->bio_list) { |
549 | unsigned long flags; | |
550 | struct bio *bio; | |
551 | int ret2; | |
552 | ||
553 | spin_lock_irqsave(&dio->bio_lock, flags); | |
554 | bio = dio->bio_list; | |
555 | dio->bio_list = bio->bi_private; | |
556 | spin_unlock_irqrestore(&dio->bio_lock, flags); | |
4e4cbee9 | 557 | ret2 = blk_status_to_errno(dio_bio_complete(dio, bio)); |
1da177e4 LT |
558 | if (ret == 0) |
559 | ret = ret2; | |
560 | } | |
eb28be2b | 561 | sdio->reap_counter = 0; |
1da177e4 LT |
562 | } |
563 | return ret; | |
564 | } | |
565 | ||
7b7a8665 CH |
566 | /* |
567 | * Create workqueue for deferred direct IO completions. We allocate the | |
568 | * workqueue when it's first needed. This avoids creating workqueue for | |
569 | * filesystems that don't need it and also allows us to create the workqueue | |
570 | * late enough so the we can include s_id in the name of the workqueue. | |
571 | */ | |
ec1b8260 | 572 | int sb_init_dio_done_wq(struct super_block *sb) |
7b7a8665 | 573 | { |
45150c43 | 574 | struct workqueue_struct *old; |
7b7a8665 CH |
575 | struct workqueue_struct *wq = alloc_workqueue("dio/%s", |
576 | WQ_MEM_RECLAIM, 0, | |
577 | sb->s_id); | |
578 | if (!wq) | |
579 | return -ENOMEM; | |
580 | /* | |
581 | * This has to be atomic as more DIOs can race to create the workqueue | |
582 | */ | |
45150c43 | 583 | old = cmpxchg(&sb->s_dio_done_wq, NULL, wq); |
7b7a8665 | 584 | /* Someone created workqueue before us? Free ours... */ |
45150c43 | 585 | if (old) |
7b7a8665 CH |
586 | destroy_workqueue(wq); |
587 | return 0; | |
588 | } | |
589 | ||
590 | static int dio_set_defer_completion(struct dio *dio) | |
591 | { | |
592 | struct super_block *sb = dio->inode->i_sb; | |
593 | ||
594 | if (dio->defer_completion) | |
595 | return 0; | |
596 | dio->defer_completion = true; | |
597 | if (!sb->s_dio_done_wq) | |
598 | return sb_init_dio_done_wq(sb); | |
599 | return 0; | |
600 | } | |
601 | ||
1da177e4 LT |
602 | /* |
603 | * Call into the fs to map some more disk blocks. We record the current number | |
eb28be2b | 604 | * of available blocks at sdio->blocks_available. These are in units of the |
93407472 | 605 | * fs blocksize, i_blocksize(inode). |
1da177e4 LT |
606 | * |
607 | * The fs is allowed to map lots of blocks at once. If it wants to do that, | |
608 | * it uses the passed inode-relative block number as the file offset, as usual. | |
609 | * | |
1d8fa7a2 | 610 | * get_block() is passed the number of i_blkbits-sized blocks which direct_io |
1da177e4 LT |
611 | * has remaining to do. The fs should not map more than this number of blocks. |
612 | * | |
613 | * If the fs has mapped a lot of blocks, it should populate bh->b_size to | |
614 | * indicate how much contiguous disk space has been made available at | |
615 | * bh->b_blocknr. | |
616 | * | |
617 | * If *any* of the mapped blocks are new, then the fs must set buffer_new(). | |
618 | * This isn't very efficient... | |
619 | * | |
620 | * In the case of filesystem holes: the fs may return an arbitrarily-large | |
621 | * hole by returning an appropriate value in b_size and by clearing | |
622 | * buffer_mapped(). However the direct-io code will only process holes one | |
1d8fa7a2 | 623 | * block at a time - it will repeatedly call get_block() as it walks the hole. |
1da177e4 | 624 | */ |
18772641 AK |
625 | static int get_more_blocks(struct dio *dio, struct dio_submit *sdio, |
626 | struct buffer_head *map_bh) | |
1da177e4 LT |
627 | { |
628 | int ret; | |
1da177e4 | 629 | sector_t fs_startblk; /* Into file, in filesystem-sized blocks */ |
ae55e1aa | 630 | sector_t fs_endblk; /* Into file, in filesystem-sized blocks */ |
1da177e4 | 631 | unsigned long fs_count; /* Number of filesystem-sized blocks */ |
1da177e4 | 632 | int create; |
ab73857e | 633 | unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor; |
8b9433eb | 634 | loff_t i_size; |
1da177e4 LT |
635 | |
636 | /* | |
637 | * If there was a memory error and we've overwritten all the | |
638 | * mapped blocks then we can now return that memory error | |
639 | */ | |
640 | ret = dio->page_errors; | |
641 | if (ret == 0) { | |
eb28be2b AK |
642 | BUG_ON(sdio->block_in_file >= sdio->final_block_in_request); |
643 | fs_startblk = sdio->block_in_file >> sdio->blkfactor; | |
ae55e1aa TM |
644 | fs_endblk = (sdio->final_block_in_request - 1) >> |
645 | sdio->blkfactor; | |
646 | fs_count = fs_endblk - fs_startblk + 1; | |
1da177e4 | 647 | |
3c674e74 | 648 | map_bh->b_state = 0; |
ab73857e | 649 | map_bh->b_size = fs_count << i_blkbits; |
3c674e74 | 650 | |
5fe878ae | 651 | /* |
9ecd10b7 EG |
652 | * For writes that could fill holes inside i_size on a |
653 | * DIO_SKIP_HOLES filesystem we forbid block creations: only | |
654 | * overwrites are permitted. We will return early to the caller | |
655 | * once we see an unmapped buffer head returned, and the caller | |
656 | * will fall back to buffered I/O. | |
5fe878ae CH |
657 | * |
658 | * Otherwise the decision is left to the get_blocks method, | |
659 | * which may decide to handle it or also return an unmapped | |
660 | * buffer head. | |
661 | */ | |
8a4c1e42 | 662 | create = dio->op == REQ_OP_WRITE; |
5fe878ae | 663 | if (dio->flags & DIO_SKIP_HOLES) { |
8b9433eb EF |
664 | i_size = i_size_read(dio->inode); |
665 | if (i_size && fs_startblk <= (i_size - 1) >> i_blkbits) | |
1da177e4 | 666 | create = 0; |
1da177e4 | 667 | } |
3c674e74 | 668 | |
eb28be2b | 669 | ret = (*sdio->get_block)(dio->inode, fs_startblk, |
1da177e4 | 670 | map_bh, create); |
18772641 AK |
671 | |
672 | /* Store for completion */ | |
673 | dio->private = map_bh->b_private; | |
7b7a8665 CH |
674 | |
675 | if (ret == 0 && buffer_defer_completion(map_bh)) | |
676 | ret = dio_set_defer_completion(dio); | |
1da177e4 LT |
677 | } |
678 | return ret; | |
679 | } | |
680 | ||
681 | /* | |
682 | * There is no bio. Make one now. | |
683 | */ | |
ba253fbf AK |
684 | static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio, |
685 | sector_t start_sector, struct buffer_head *map_bh) | |
1da177e4 LT |
686 | { |
687 | sector_t sector; | |
688 | int ret, nr_pages; | |
689 | ||
eb28be2b | 690 | ret = dio_bio_reap(dio, sdio); |
1da177e4 LT |
691 | if (ret) |
692 | goto out; | |
eb28be2b | 693 | sector = start_sector << (sdio->blkbits - 9); |
5f7136db | 694 | nr_pages = bio_max_segs(sdio->pages_in_io); |
1da177e4 | 695 | BUG_ON(nr_pages <= 0); |
18772641 | 696 | dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages); |
eb28be2b | 697 | sdio->boundary = 0; |
1da177e4 LT |
698 | out: |
699 | return ret; | |
700 | } | |
701 | ||
702 | /* | |
703 | * Attempt to put the current chunk of 'cur_page' into the current BIO. If | |
704 | * that was successful then update final_block_in_bio and take a ref against | |
705 | * the just-added page. | |
706 | * | |
707 | * Return zero on success. Non-zero means the caller needs to start a new BIO. | |
708 | */ | |
ba253fbf | 709 | static inline int dio_bio_add_page(struct dio_submit *sdio) |
1da177e4 LT |
710 | { |
711 | int ret; | |
712 | ||
eb28be2b AK |
713 | ret = bio_add_page(sdio->bio, sdio->cur_page, |
714 | sdio->cur_page_len, sdio->cur_page_offset); | |
715 | if (ret == sdio->cur_page_len) { | |
1da177e4 LT |
716 | /* |
717 | * Decrement count only, if we are done with this page | |
718 | */ | |
eb28be2b AK |
719 | if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE) |
720 | sdio->pages_in_io--; | |
09cbfeaf | 721 | get_page(sdio->cur_page); |
eb28be2b AK |
722 | sdio->final_block_in_bio = sdio->cur_page_block + |
723 | (sdio->cur_page_len >> sdio->blkbits); | |
1da177e4 LT |
724 | ret = 0; |
725 | } else { | |
726 | ret = 1; | |
727 | } | |
728 | return ret; | |
729 | } | |
730 | ||
731 | /* | |
732 | * Put cur_page under IO. The section of cur_page which is described by | |
733 | * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page | |
734 | * starts on-disk at cur_page_block. | |
735 | * | |
736 | * We take a ref against the page here (on behalf of its presence in the bio). | |
737 | * | |
738 | * The caller of this function is responsible for removing cur_page from the | |
739 | * dio, and for dropping the refcount which came from that presence. | |
740 | */ | |
ba253fbf AK |
741 | static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio, |
742 | struct buffer_head *map_bh) | |
1da177e4 LT |
743 | { |
744 | int ret = 0; | |
745 | ||
eb28be2b AK |
746 | if (sdio->bio) { |
747 | loff_t cur_offset = sdio->cur_page_fs_offset; | |
748 | loff_t bio_next_offset = sdio->logical_offset_in_bio + | |
4f024f37 | 749 | sdio->bio->bi_iter.bi_size; |
c2c6ca41 | 750 | |
1da177e4 | 751 | /* |
c2c6ca41 JB |
752 | * See whether this new request is contiguous with the old. |
753 | * | |
f0940cee NK |
754 | * Btrfs cannot handle having logically non-contiguous requests |
755 | * submitted. For example if you have | |
c2c6ca41 JB |
756 | * |
757 | * Logical: [0-4095][HOLE][8192-12287] | |
f0940cee | 758 | * Physical: [0-4095] [4096-8191] |
c2c6ca41 JB |
759 | * |
760 | * We cannot submit those pages together as one BIO. So if our | |
761 | * current logical offset in the file does not equal what would | |
762 | * be the next logical offset in the bio, submit the bio we | |
763 | * have. | |
1da177e4 | 764 | */ |
eb28be2b | 765 | if (sdio->final_block_in_bio != sdio->cur_page_block || |
c2c6ca41 | 766 | cur_offset != bio_next_offset) |
eb28be2b | 767 | dio_bio_submit(dio, sdio); |
1da177e4 LT |
768 | } |
769 | ||
eb28be2b | 770 | if (sdio->bio == NULL) { |
18772641 | 771 | ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh); |
1da177e4 LT |
772 | if (ret) |
773 | goto out; | |
774 | } | |
775 | ||
eb28be2b AK |
776 | if (dio_bio_add_page(sdio) != 0) { |
777 | dio_bio_submit(dio, sdio); | |
18772641 | 778 | ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh); |
1da177e4 | 779 | if (ret == 0) { |
eb28be2b | 780 | ret = dio_bio_add_page(sdio); |
1da177e4 LT |
781 | BUG_ON(ret != 0); |
782 | } | |
783 | } | |
784 | out: | |
785 | return ret; | |
786 | } | |
787 | ||
788 | /* | |
789 | * An autonomous function to put a chunk of a page under deferred IO. | |
790 | * | |
791 | * The caller doesn't actually know (or care) whether this piece of page is in | |
792 | * a BIO, or is under IO or whatever. We just take care of all possible | |
793 | * situations here. The separation between the logic of do_direct_IO() and | |
794 | * that of submit_page_section() is important for clarity. Please don't break. | |
795 | * | |
796 | * The chunk of page starts on-disk at blocknr. | |
797 | * | |
798 | * We perform deferred IO, by recording the last-submitted page inside our | |
799 | * private part of the dio structure. If possible, we just expand the IO | |
800 | * across that page here. | |
801 | * | |
802 | * If that doesn't work out then we put the old page into the bio and add this | |
803 | * page to the dio instead. | |
804 | */ | |
ba253fbf | 805 | static inline int |
eb28be2b | 806 | submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page, |
18772641 AK |
807 | unsigned offset, unsigned len, sector_t blocknr, |
808 | struct buffer_head *map_bh) | |
1da177e4 LT |
809 | { |
810 | int ret = 0; | |
df41872b | 811 | int boundary = sdio->boundary; /* dio_send_cur_page may clear it */ |
1da177e4 | 812 | |
8a4c1e42 | 813 | if (dio->op == REQ_OP_WRITE) { |
98c4d57d AM |
814 | /* |
815 | * Read accounting is performed in submit_bio() | |
816 | */ | |
817 | task_io_account_write(len); | |
818 | } | |
819 | ||
1da177e4 LT |
820 | /* |
821 | * Can we just grow the current page's presence in the dio? | |
822 | */ | |
eb28be2b AK |
823 | if (sdio->cur_page == page && |
824 | sdio->cur_page_offset + sdio->cur_page_len == offset && | |
825 | sdio->cur_page_block + | |
826 | (sdio->cur_page_len >> sdio->blkbits) == blocknr) { | |
827 | sdio->cur_page_len += len; | |
1da177e4 LT |
828 | goto out; |
829 | } | |
830 | ||
831 | /* | |
832 | * If there's a deferred page already there then send it. | |
833 | */ | |
eb28be2b | 834 | if (sdio->cur_page) { |
18772641 | 835 | ret = dio_send_cur_page(dio, sdio, map_bh); |
09cbfeaf | 836 | put_page(sdio->cur_page); |
eb28be2b | 837 | sdio->cur_page = NULL; |
1da177e4 | 838 | if (ret) |
b1058b98 | 839 | return ret; |
1da177e4 LT |
840 | } |
841 | ||
09cbfeaf | 842 | get_page(page); /* It is in dio */ |
eb28be2b AK |
843 | sdio->cur_page = page; |
844 | sdio->cur_page_offset = offset; | |
845 | sdio->cur_page_len = len; | |
846 | sdio->cur_page_block = blocknr; | |
847 | sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits; | |
1da177e4 | 848 | out: |
b1058b98 | 849 | /* |
df41872b | 850 | * If boundary then we want to schedule the IO now to |
b1058b98 JK |
851 | * avoid metadata seeks. |
852 | */ | |
df41872b | 853 | if (boundary) { |
b1058b98 | 854 | ret = dio_send_cur_page(dio, sdio, map_bh); |
899f0429 AG |
855 | if (sdio->bio) |
856 | dio_bio_submit(dio, sdio); | |
09cbfeaf | 857 | put_page(sdio->cur_page); |
b1058b98 JK |
858 | sdio->cur_page = NULL; |
859 | } | |
1da177e4 LT |
860 | return ret; |
861 | } | |
862 | ||
1da177e4 LT |
863 | /* |
864 | * If we are not writing the entire block and get_block() allocated | |
865 | * the block for us, we need to fill-in the unused portion of the | |
866 | * block with zeros. This happens only if user-buffer, fileoffset or | |
867 | * io length is not filesystem block-size multiple. | |
868 | * | |
869 | * `end' is zero if we're doing the start of the IO, 1 at the end of the | |
870 | * IO. | |
871 | */ | |
ba253fbf AK |
872 | static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio, |
873 | int end, struct buffer_head *map_bh) | |
1da177e4 LT |
874 | { |
875 | unsigned dio_blocks_per_fs_block; | |
876 | unsigned this_chunk_blocks; /* In dio_blocks */ | |
877 | unsigned this_chunk_bytes; | |
878 | struct page *page; | |
879 | ||
eb28be2b | 880 | sdio->start_zero_done = 1; |
18772641 | 881 | if (!sdio->blkfactor || !buffer_new(map_bh)) |
1da177e4 LT |
882 | return; |
883 | ||
eb28be2b AK |
884 | dio_blocks_per_fs_block = 1 << sdio->blkfactor; |
885 | this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1); | |
1da177e4 LT |
886 | |
887 | if (!this_chunk_blocks) | |
888 | return; | |
889 | ||
890 | /* | |
891 | * We need to zero out part of an fs block. It is either at the | |
892 | * beginning or the end of the fs block. | |
893 | */ | |
894 | if (end) | |
895 | this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks; | |
896 | ||
eb28be2b | 897 | this_chunk_bytes = this_chunk_blocks << sdio->blkbits; |
1da177e4 | 898 | |
557ed1fa | 899 | page = ZERO_PAGE(0); |
eb28be2b | 900 | if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes, |
18772641 | 901 | sdio->next_block_for_io, map_bh)) |
1da177e4 LT |
902 | return; |
903 | ||
eb28be2b | 904 | sdio->next_block_for_io += this_chunk_blocks; |
1da177e4 LT |
905 | } |
906 | ||
907 | /* | |
908 | * Walk the user pages, and the file, mapping blocks to disk and generating | |
909 | * a sequence of (page,offset,len,block) mappings. These mappings are injected | |
910 | * into submit_page_section(), which takes care of the next stage of submission | |
911 | * | |
912 | * Direct IO against a blockdev is different from a file. Because we can | |
913 | * happily perform page-sized but 512-byte aligned IOs. It is important that | |
914 | * blockdev IO be able to have fine alignment and large sizes. | |
915 | * | |
1d8fa7a2 | 916 | * So what we do is to permit the ->get_block function to populate bh.b_size |
1da177e4 LT |
917 | * with the size of IO which is permitted at this offset and this i_blkbits. |
918 | * | |
919 | * For best results, the blockdev should be set up with 512-byte i_blkbits and | |
1d8fa7a2 | 920 | * it should set b_size to PAGE_SIZE or more inside get_block(). This gives |
1da177e4 LT |
921 | * fine alignment but still allows this function to work in PAGE_SIZE units. |
922 | */ | |
18772641 AK |
923 | static int do_direct_IO(struct dio *dio, struct dio_submit *sdio, |
924 | struct buffer_head *map_bh) | |
1da177e4 | 925 | { |
eb28be2b | 926 | const unsigned blkbits = sdio->blkbits; |
dd545b52 | 927 | const unsigned i_blkbits = blkbits + sdio->blkfactor; |
1da177e4 LT |
928 | int ret = 0; |
929 | ||
eb28be2b | 930 | while (sdio->block_in_file < sdio->final_block_in_request) { |
7b2c99d1 AV |
931 | struct page *page; |
932 | size_t from, to; | |
6fcc5420 BH |
933 | |
934 | page = dio_get_page(dio, sdio); | |
1da177e4 LT |
935 | if (IS_ERR(page)) { |
936 | ret = PTR_ERR(page); | |
937 | goto out; | |
938 | } | |
6fcc5420 BH |
939 | from = sdio->head ? 0 : sdio->from; |
940 | to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE; | |
941 | sdio->head++; | |
1da177e4 | 942 | |
7b2c99d1 | 943 | while (from < to) { |
1da177e4 LT |
944 | unsigned this_chunk_bytes; /* # of bytes mapped */ |
945 | unsigned this_chunk_blocks; /* # of blocks */ | |
946 | unsigned u; | |
947 | ||
eb28be2b | 948 | if (sdio->blocks_available == 0) { |
1da177e4 LT |
949 | /* |
950 | * Need to go and map some more disk | |
951 | */ | |
952 | unsigned long blkmask; | |
953 | unsigned long dio_remainder; | |
954 | ||
18772641 | 955 | ret = get_more_blocks(dio, sdio, map_bh); |
1da177e4 | 956 | if (ret) { |
09cbfeaf | 957 | put_page(page); |
1da177e4 LT |
958 | goto out; |
959 | } | |
960 | if (!buffer_mapped(map_bh)) | |
961 | goto do_holes; | |
962 | ||
eb28be2b | 963 | sdio->blocks_available = |
f734c89c | 964 | map_bh->b_size >> blkbits; |
eb28be2b AK |
965 | sdio->next_block_for_io = |
966 | map_bh->b_blocknr << sdio->blkfactor; | |
f734c89c JK |
967 | if (buffer_new(map_bh)) { |
968 | clean_bdev_aliases( | |
969 | map_bh->b_bdev, | |
970 | map_bh->b_blocknr, | |
dd545b52 | 971 | map_bh->b_size >> i_blkbits); |
f734c89c | 972 | } |
1da177e4 | 973 | |
eb28be2b | 974 | if (!sdio->blkfactor) |
1da177e4 LT |
975 | goto do_holes; |
976 | ||
eb28be2b AK |
977 | blkmask = (1 << sdio->blkfactor) - 1; |
978 | dio_remainder = (sdio->block_in_file & blkmask); | |
1da177e4 LT |
979 | |
980 | /* | |
981 | * If we are at the start of IO and that IO | |
982 | * starts partway into a fs-block, | |
983 | * dio_remainder will be non-zero. If the IO | |
984 | * is a read then we can simply advance the IO | |
985 | * cursor to the first block which is to be | |
986 | * read. But if the IO is a write and the | |
987 | * block was newly allocated we cannot do that; | |
988 | * the start of the fs block must be zeroed out | |
989 | * on-disk | |
990 | */ | |
991 | if (!buffer_new(map_bh)) | |
eb28be2b AK |
992 | sdio->next_block_for_io += dio_remainder; |
993 | sdio->blocks_available -= dio_remainder; | |
1da177e4 LT |
994 | } |
995 | do_holes: | |
996 | /* Handle holes */ | |
997 | if (!buffer_mapped(map_bh)) { | |
35dc8161 | 998 | loff_t i_size_aligned; |
1da177e4 LT |
999 | |
1000 | /* AKPM: eargh, -ENOTBLK is a hack */ | |
8a4c1e42 | 1001 | if (dio->op == REQ_OP_WRITE) { |
09cbfeaf | 1002 | put_page(page); |
1da177e4 LT |
1003 | return -ENOTBLK; |
1004 | } | |
1005 | ||
35dc8161 JM |
1006 | /* |
1007 | * Be sure to account for a partial block as the | |
1008 | * last block in the file | |
1009 | */ | |
1010 | i_size_aligned = ALIGN(i_size_read(dio->inode), | |
1011 | 1 << blkbits); | |
eb28be2b | 1012 | if (sdio->block_in_file >= |
35dc8161 | 1013 | i_size_aligned >> blkbits) { |
1da177e4 | 1014 | /* We hit eof */ |
09cbfeaf | 1015 | put_page(page); |
1da177e4 LT |
1016 | goto out; |
1017 | } | |
7b2c99d1 | 1018 | zero_user(page, from, 1 << blkbits); |
eb28be2b | 1019 | sdio->block_in_file++; |
7b2c99d1 | 1020 | from += 1 << blkbits; |
3320c60b | 1021 | dio->result += 1 << blkbits; |
1da177e4 LT |
1022 | goto next_block; |
1023 | } | |
1024 | ||
1025 | /* | |
1026 | * If we're performing IO which has an alignment which | |
1027 | * is finer than the underlying fs, go check to see if | |
1028 | * we must zero out the start of this block. | |
1029 | */ | |
eb28be2b | 1030 | if (unlikely(sdio->blkfactor && !sdio->start_zero_done)) |
18772641 | 1031 | dio_zero_block(dio, sdio, 0, map_bh); |
1da177e4 LT |
1032 | |
1033 | /* | |
1034 | * Work out, in this_chunk_blocks, how much disk we | |
1035 | * can add to this page | |
1036 | */ | |
eb28be2b | 1037 | this_chunk_blocks = sdio->blocks_available; |
7b2c99d1 | 1038 | u = (to - from) >> blkbits; |
1da177e4 LT |
1039 | if (this_chunk_blocks > u) |
1040 | this_chunk_blocks = u; | |
eb28be2b | 1041 | u = sdio->final_block_in_request - sdio->block_in_file; |
1da177e4 LT |
1042 | if (this_chunk_blocks > u) |
1043 | this_chunk_blocks = u; | |
1044 | this_chunk_bytes = this_chunk_blocks << blkbits; | |
1045 | BUG_ON(this_chunk_bytes == 0); | |
1046 | ||
092c8d46 JK |
1047 | if (this_chunk_blocks == sdio->blocks_available) |
1048 | sdio->boundary = buffer_boundary(map_bh); | |
eb28be2b | 1049 | ret = submit_page_section(dio, sdio, page, |
7b2c99d1 | 1050 | from, |
eb28be2b | 1051 | this_chunk_bytes, |
18772641 AK |
1052 | sdio->next_block_for_io, |
1053 | map_bh); | |
1da177e4 | 1054 | if (ret) { |
09cbfeaf | 1055 | put_page(page); |
1da177e4 LT |
1056 | goto out; |
1057 | } | |
eb28be2b | 1058 | sdio->next_block_for_io += this_chunk_blocks; |
1da177e4 | 1059 | |
eb28be2b | 1060 | sdio->block_in_file += this_chunk_blocks; |
7b2c99d1 AV |
1061 | from += this_chunk_bytes; |
1062 | dio->result += this_chunk_bytes; | |
eb28be2b | 1063 | sdio->blocks_available -= this_chunk_blocks; |
1da177e4 | 1064 | next_block: |
eb28be2b AK |
1065 | BUG_ON(sdio->block_in_file > sdio->final_block_in_request); |
1066 | if (sdio->block_in_file == sdio->final_block_in_request) | |
1da177e4 LT |
1067 | break; |
1068 | } | |
1069 | ||
1070 | /* Drop the ref which was taken in get_user_pages() */ | |
09cbfeaf | 1071 | put_page(page); |
1da177e4 LT |
1072 | } |
1073 | out: | |
1074 | return ret; | |
1075 | } | |
1076 | ||
847cc637 | 1077 | static inline int drop_refcount(struct dio *dio) |
1da177e4 | 1078 | { |
847cc637 | 1079 | int ret2; |
5eb6c7a2 | 1080 | unsigned long flags; |
1da177e4 | 1081 | |
8459d86a ZB |
1082 | /* |
1083 | * Sync will always be dropping the final ref and completing the | |
5eb6c7a2 ZB |
1084 | * operation. AIO can if it was a broken operation described above or |
1085 | * in fact if all the bios race to complete before we get here. In | |
1086 | * that case dio_complete() translates the EIOCBQUEUED into the proper | |
04b2fa9f | 1087 | * return code that the caller will hand to ->complete(). |
5eb6c7a2 ZB |
1088 | * |
1089 | * This is managed by the bio_lock instead of being an atomic_t so that | |
1090 | * completion paths can drop their ref and use the remaining count to | |
1091 | * decide to wake the submission path atomically. | |
8459d86a | 1092 | */ |
5eb6c7a2 ZB |
1093 | spin_lock_irqsave(&dio->bio_lock, flags); |
1094 | ret2 = --dio->refcount; | |
1095 | spin_unlock_irqrestore(&dio->bio_lock, flags); | |
847cc637 | 1096 | return ret2; |
1da177e4 LT |
1097 | } |
1098 | ||
eafdc7d1 CH |
1099 | /* |
1100 | * This is a library function for use by filesystem drivers. | |
1101 | * | |
1102 | * The locking rules are governed by the flags parameter: | |
1103 | * - if the flags value contains DIO_LOCKING we use a fancy locking | |
1104 | * scheme for dumb filesystems. | |
1105 | * For writes this function is called under i_mutex and returns with | |
1106 | * i_mutex held, for reads, i_mutex is not held on entry, but it is | |
1107 | * taken and dropped again before returning. | |
eafdc7d1 CH |
1108 | * - if the flags value does NOT contain DIO_LOCKING we don't use any |
1109 | * internal locking but rather rely on the filesystem to synchronize | |
1110 | * direct I/O reads/writes versus each other and truncate. | |
df2d6f26 CH |
1111 | * |
1112 | * To help with locking against truncate we incremented the i_dio_count | |
1113 | * counter before starting direct I/O, and decrement it once we are done. | |
1114 | * Truncate can wait for it to reach zero to provide exclusion. It is | |
1115 | * expected that filesystem provide exclusion between new direct I/O | |
1116 | * and truncates. For DIO_LOCKING filesystems this is done by i_mutex, | |
1117 | * but other filesystems need to take care of this on their own. | |
ba253fbf AK |
1118 | * |
1119 | * NOTE: if you pass "sdio" to anything by pointer make sure that function | |
1120 | * is always inlined. Otherwise gcc is unable to split the structure into | |
1121 | * individual fields and will generate much worse code. This is important | |
1122 | * for the whole file. | |
eafdc7d1 | 1123 | */ |
65dd2aa9 | 1124 | static inline ssize_t |
17f8c842 OS |
1125 | do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode, |
1126 | struct block_device *bdev, struct iov_iter *iter, | |
c8b8e32d | 1127 | get_block_t get_block, dio_iodone_t end_io, |
17f8c842 | 1128 | dio_submit_t submit_io, int flags) |
1da177e4 | 1129 | { |
6aa7de05 | 1130 | unsigned i_blkbits = READ_ONCE(inode->i_blkbits); |
ab73857e | 1131 | unsigned blkbits = i_blkbits; |
1da177e4 LT |
1132 | unsigned blocksize_mask = (1 << blkbits) - 1; |
1133 | ssize_t retval = -EINVAL; | |
1c0ff0f1 | 1134 | const size_t count = iov_iter_count(iter); |
c8b8e32d | 1135 | loff_t offset = iocb->ki_pos; |
1c0ff0f1 | 1136 | const loff_t end = offset + count; |
1da177e4 | 1137 | struct dio *dio; |
eb28be2b | 1138 | struct dio_submit sdio = { 0, }; |
847cc637 | 1139 | struct buffer_head map_bh = { 0, }; |
647d1e4c | 1140 | struct blk_plug plug; |
886a3911 | 1141 | unsigned long align = offset | iov_iter_alignment(iter); |
1da177e4 | 1142 | |
65dd2aa9 AK |
1143 | /* |
1144 | * Avoid references to bdev if not absolutely needed to give | |
1145 | * the early prefetch in the caller enough time. | |
1146 | */ | |
1da177e4 | 1147 | |
f9b5570d | 1148 | /* watch out for a 0 len io from a tricksy fs */ |
1c0ff0f1 | 1149 | if (iov_iter_rw(iter) == READ && !count) |
f9b5570d CH |
1150 | return 0; |
1151 | ||
6e8267f5 | 1152 | dio = kmem_cache_alloc(dio_cache, GFP_KERNEL); |
1da177e4 | 1153 | if (!dio) |
46d71602 | 1154 | return -ENOMEM; |
23aee091 JM |
1155 | /* |
1156 | * Believe it or not, zeroing out the page array caused a .5% | |
1157 | * performance regression in a database benchmark. So, we take | |
1158 | * care to only zero out what's needed. | |
1159 | */ | |
1160 | memset(dio, 0, offsetof(struct dio, pages)); | |
1da177e4 | 1161 | |
5fe878ae | 1162 | dio->flags = flags; |
0a9164cb GKB |
1163 | if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ) { |
1164 | /* will be released by direct_io_worker */ | |
1165 | inode_lock(inode); | |
1da177e4 LT |
1166 | } |
1167 | ||
74cedf9b JK |
1168 | /* Once we sampled i_size check for reads beyond EOF */ |
1169 | dio->i_size = i_size_read(inode); | |
1170 | if (iov_iter_rw(iter) == READ && offset >= dio->i_size) { | |
2d4594ac | 1171 | retval = 0; |
46d71602 | 1172 | goto fail_dio; |
74cedf9b JK |
1173 | } |
1174 | ||
41b21af3 GKB |
1175 | if (align & blocksize_mask) { |
1176 | if (bdev) | |
1177 | blkbits = blksize_bits(bdev_logical_block_size(bdev)); | |
1178 | blocksize_mask = (1 << blkbits) - 1; | |
1179 | if (align & blocksize_mask) | |
1180 | goto fail_dio; | |
1181 | } | |
1182 | ||
0a9164cb GKB |
1183 | if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ) { |
1184 | struct address_space *mapping = iocb->ki_filp->f_mapping; | |
1185 | ||
1186 | retval = filemap_write_and_wait_range(mapping, offset, end - 1); | |
1187 | if (retval) | |
1188 | goto fail_dio; | |
74cedf9b JK |
1189 | } |
1190 | ||
1da177e4 | 1191 | /* |
60392573 CH |
1192 | * For file extending writes updating i_size before data writeouts |
1193 | * complete can expose uninitialized blocks in dumb filesystems. | |
1194 | * In that case we need to wait for I/O completion even if asked | |
1195 | * for an asynchronous write. | |
1da177e4 | 1196 | */ |
60392573 CH |
1197 | if (is_sync_kiocb(iocb)) |
1198 | dio->is_async = false; | |
c8f4c36f | 1199 | else if (iov_iter_rw(iter) == WRITE && end > i_size_read(inode)) |
60392573 CH |
1200 | dio->is_async = false; |
1201 | else | |
1202 | dio->is_async = true; | |
1203 | ||
847cc637 | 1204 | dio->inode = inode; |
8a4c1e42 MC |
1205 | if (iov_iter_rw(iter) == WRITE) { |
1206 | dio->op = REQ_OP_WRITE; | |
70fd7614 | 1207 | dio->op_flags = REQ_SYNC | REQ_IDLE; |
03a07c92 GR |
1208 | if (iocb->ki_flags & IOCB_NOWAIT) |
1209 | dio->op_flags |= REQ_NOWAIT; | |
8a4c1e42 MC |
1210 | } else { |
1211 | dio->op = REQ_OP_READ; | |
1212 | } | |
02afc27f CH |
1213 | |
1214 | /* | |
1215 | * For AIO O_(D)SYNC writes we need to defer completions to a workqueue | |
1216 | * so that we can call ->fsync. | |
1217 | */ | |
332391a9 LC |
1218 | if (dio->is_async && iov_iter_rw(iter) == WRITE) { |
1219 | retval = 0; | |
d9c10e5b | 1220 | if (iocb->ki_flags & IOCB_DSYNC) |
332391a9 LC |
1221 | retval = dio_set_defer_completion(dio); |
1222 | else if (!dio->inode->i_sb->s_dio_done_wq) { | |
1223 | /* | |
1224 | * In case of AIO write racing with buffered read we | |
1225 | * need to defer completion. We can't decide this now, | |
1226 | * however the workqueue needs to be initialized here. | |
1227 | */ | |
1228 | retval = sb_init_dio_done_wq(dio->inode->i_sb); | |
1229 | } | |
46d71602 GKB |
1230 | if (retval) |
1231 | goto fail_dio; | |
02afc27f CH |
1232 | } |
1233 | ||
1234 | /* | |
1235 | * Will be decremented at I/O completion time. | |
1236 | */ | |
ce3077ee | 1237 | inode_dio_begin(inode); |
02afc27f CH |
1238 | |
1239 | retval = 0; | |
847cc637 | 1240 | sdio.blkbits = blkbits; |
ab73857e | 1241 | sdio.blkfactor = i_blkbits - blkbits; |
847cc637 AK |
1242 | sdio.block_in_file = offset >> blkbits; |
1243 | ||
1244 | sdio.get_block = get_block; | |
1245 | dio->end_io = end_io; | |
1246 | sdio.submit_io = submit_io; | |
1247 | sdio.final_block_in_bio = -1; | |
1248 | sdio.next_block_for_io = -1; | |
1249 | ||
1250 | dio->iocb = iocb; | |
847cc637 AK |
1251 | |
1252 | spin_lock_init(&dio->bio_lock); | |
1253 | dio->refcount = 1; | |
1254 | ||
00e23707 | 1255 | dio->should_dirty = iter_is_iovec(iter) && iov_iter_rw(iter) == READ; |
7b2c99d1 | 1256 | sdio.iter = iter; |
1c0ff0f1 | 1257 | sdio.final_block_in_request = end >> blkbits; |
7b2c99d1 | 1258 | |
847cc637 AK |
1259 | /* |
1260 | * In case of non-aligned buffers, we may need 2 more | |
1261 | * pages since we need to zero out first and last block. | |
1262 | */ | |
1263 | if (unlikely(sdio.blkfactor)) | |
1264 | sdio.pages_in_io = 2; | |
1265 | ||
f67da30c | 1266 | sdio.pages_in_io += iov_iter_npages(iter, INT_MAX); |
847cc637 | 1267 | |
647d1e4c FW |
1268 | blk_start_plug(&plug); |
1269 | ||
7b2c99d1 AV |
1270 | retval = do_direct_IO(dio, &sdio, &map_bh); |
1271 | if (retval) | |
1272 | dio_cleanup(dio, &sdio); | |
847cc637 AK |
1273 | |
1274 | if (retval == -ENOTBLK) { | |
1275 | /* | |
1276 | * The remaining part of the request will be | |
3d742d4b | 1277 | * handled by buffered I/O when we return |
847cc637 AK |
1278 | */ |
1279 | retval = 0; | |
1280 | } | |
1281 | /* | |
1282 | * There may be some unwritten disk at the end of a part-written | |
1283 | * fs-block-sized block. Go zero that now. | |
1284 | */ | |
1285 | dio_zero_block(dio, &sdio, 1, &map_bh); | |
1286 | ||
1287 | if (sdio.cur_page) { | |
1288 | ssize_t ret2; | |
1289 | ||
1290 | ret2 = dio_send_cur_page(dio, &sdio, &map_bh); | |
1291 | if (retval == 0) | |
1292 | retval = ret2; | |
09cbfeaf | 1293 | put_page(sdio.cur_page); |
847cc637 AK |
1294 | sdio.cur_page = NULL; |
1295 | } | |
1296 | if (sdio.bio) | |
1297 | dio_bio_submit(dio, &sdio); | |
1298 | ||
647d1e4c FW |
1299 | blk_finish_plug(&plug); |
1300 | ||
847cc637 AK |
1301 | /* |
1302 | * It is possible that, we return short IO due to end of file. | |
1303 | * In that case, we need to release all the pages we got hold on. | |
1304 | */ | |
1305 | dio_cleanup(dio, &sdio); | |
1306 | ||
1307 | /* | |
1308 | * All block lookups have been performed. For READ requests | |
1309 | * we can let i_mutex go now that its achieved its purpose | |
1310 | * of protecting us from looking up uninitialized blocks. | |
1311 | */ | |
17f8c842 | 1312 | if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING)) |
5955102c | 1313 | inode_unlock(dio->inode); |
847cc637 AK |
1314 | |
1315 | /* | |
1316 | * The only time we want to leave bios in flight is when a successful | |
1317 | * partial aio read or full aio write have been setup. In that case | |
1318 | * bio completion will call aio_complete. The only time it's safe to | |
1319 | * call aio_complete is when we return -EIOCBQUEUED, so we key on that. | |
1320 | * This had *better* be the only place that raises -EIOCBQUEUED. | |
1321 | */ | |
1322 | BUG_ON(retval == -EIOCBQUEUED); | |
1323 | if (dio->is_async && retval == 0 && dio->result && | |
17f8c842 | 1324 | (iov_iter_rw(iter) == READ || dio->result == count)) |
847cc637 | 1325 | retval = -EIOCBQUEUED; |
af436472 | 1326 | else |
847cc637 AK |
1327 | dio_await_completion(dio); |
1328 | ||
1329 | if (drop_refcount(dio) == 0) { | |
ffe51f01 | 1330 | retval = dio_complete(dio, retval, DIO_COMPLETE_INVALIDATE); |
847cc637 AK |
1331 | } else |
1332 | BUG_ON(retval != -EIOCBQUEUED); | |
1da177e4 | 1333 | |
46d71602 GKB |
1334 | return retval; |
1335 | ||
1336 | fail_dio: | |
1337 | if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ) | |
1338 | inode_unlock(inode); | |
1339 | ||
1340 | kmem_cache_free(dio_cache, dio); | |
7bb46a67 | 1341 | return retval; |
1342 | } | |
65dd2aa9 | 1343 | |
17f8c842 OS |
1344 | ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode, |
1345 | struct block_device *bdev, struct iov_iter *iter, | |
c8b8e32d | 1346 | get_block_t get_block, |
17f8c842 OS |
1347 | dio_iodone_t end_io, dio_submit_t submit_io, |
1348 | int flags) | |
65dd2aa9 AK |
1349 | { |
1350 | /* | |
1351 | * The block device state is needed in the end to finally | |
1352 | * submit everything. Since it's likely to be cache cold | |
1353 | * prefetch it here as first thing to hide some of the | |
1354 | * latency. | |
1355 | * | |
1356 | * Attempt to prefetch the pieces we likely need later. | |
1357 | */ | |
1358 | prefetch(&bdev->bd_disk->part_tbl); | |
e556f6ba CH |
1359 | prefetch(bdev->bd_disk->queue); |
1360 | prefetch((char *)bdev->bd_disk->queue + SMP_CACHE_BYTES); | |
65dd2aa9 | 1361 | |
c8b8e32d | 1362 | return do_blockdev_direct_IO(iocb, inode, bdev, iter, get_block, |
17f8c842 | 1363 | end_io, submit_io, flags); |
65dd2aa9 AK |
1364 | } |
1365 | ||
1da177e4 | 1366 | EXPORT_SYMBOL(__blockdev_direct_IO); |
6e8267f5 AK |
1367 | |
1368 | static __init int dio_init(void) | |
1369 | { | |
1370 | dio_cache = KMEM_CACHE(dio, SLAB_PANIC); | |
1371 | return 0; | |
1372 | } | |
1373 | module_init(dio_init) |