Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * fs/direct-io.c | |
3 | * | |
4 | * Copyright (C) 2002, Linus Torvalds. | |
5 | * | |
6 | * O_DIRECT | |
7 | * | |
e1f8e874 | 8 | * 04Jul2002 Andrew Morton |
1da177e4 LT |
9 | * Initial version |
10 | * 11Sep2002 janetinc@us.ibm.com | |
11 | * added readv/writev support. | |
e1f8e874 | 12 | * 29Oct2002 Andrew Morton |
1da177e4 LT |
13 | * rewrote bio_add_page() support. |
14 | * 30Oct2002 pbadari@us.ibm.com | |
15 | * added support for non-aligned IO. | |
16 | * 06Nov2002 pbadari@us.ibm.com | |
17 | * added asynchronous IO support. | |
18 | * 21Jul2003 nathans@sgi.com | |
19 | * added IO completion notifier. | |
20 | */ | |
21 | ||
22 | #include <linux/kernel.h> | |
23 | #include <linux/module.h> | |
24 | #include <linux/types.h> | |
25 | #include <linux/fs.h> | |
26 | #include <linux/mm.h> | |
27 | #include <linux/slab.h> | |
28 | #include <linux/highmem.h> | |
29 | #include <linux/pagemap.h> | |
98c4d57d | 30 | #include <linux/task_io_accounting_ops.h> |
1da177e4 LT |
31 | #include <linux/bio.h> |
32 | #include <linux/wait.h> | |
33 | #include <linux/err.h> | |
34 | #include <linux/blkdev.h> | |
35 | #include <linux/buffer_head.h> | |
36 | #include <linux/rwsem.h> | |
37 | #include <linux/uio.h> | |
60063497 | 38 | #include <linux/atomic.h> |
65dd2aa9 | 39 | #include <linux/prefetch.h> |
1da177e4 LT |
40 | |
41 | /* | |
42 | * How many user pages to map in one call to get_user_pages(). This determines | |
cde1ecb3 | 43 | * the size of a structure in the slab cache |
1da177e4 LT |
44 | */ |
45 | #define DIO_PAGES 64 | |
46 | ||
47 | /* | |
48 | * This code generally works in units of "dio_blocks". A dio_block is | |
49 | * somewhere between the hard sector size and the filesystem block size. it | |
50 | * is determined on a per-invocation basis. When talking to the filesystem | |
51 | * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity | |
52 | * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted | |
53 | * to bio_block quantities by shifting left by blkfactor. | |
54 | * | |
55 | * If blkfactor is zero then the user's request was aligned to the filesystem's | |
56 | * blocksize. | |
1da177e4 LT |
57 | */ |
58 | ||
eb28be2b AK |
59 | /* dio_state only used in the submission path */ |
60 | ||
61 | struct dio_submit { | |
1da177e4 | 62 | struct bio *bio; /* bio under assembly */ |
1da177e4 LT |
63 | unsigned blkbits; /* doesn't change */ |
64 | unsigned blkfactor; /* When we're using an alignment which | |
65 | is finer than the filesystem's soft | |
66 | blocksize, this specifies how much | |
67 | finer. blkfactor=2 means 1/4-block | |
68 | alignment. Does not change */ | |
69 | unsigned start_zero_done; /* flag: sub-blocksize zeroing has | |
70 | been performed at the start of a | |
71 | write */ | |
72 | int pages_in_io; /* approximate total IO pages */ | |
1da177e4 LT |
73 | sector_t block_in_file; /* Current offset into the underlying |
74 | file in dio_block units. */ | |
75 | unsigned blocks_available; /* At block_in_file. changes */ | |
0dc2bc49 | 76 | int reap_counter; /* rate limit reaping */ |
1da177e4 | 77 | sector_t final_block_in_request;/* doesn't change */ |
1da177e4 | 78 | int boundary; /* prev block is at a boundary */ |
1d8fa7a2 | 79 | get_block_t *get_block; /* block mapping function */ |
facd07b0 | 80 | dio_submit_t *submit_io; /* IO submition function */ |
eb28be2b | 81 | |
facd07b0 | 82 | loff_t logical_offset_in_bio; /* current first logical block in bio */ |
1da177e4 LT |
83 | sector_t final_block_in_bio; /* current final block in bio + 1 */ |
84 | sector_t next_block_for_io; /* next block to be put under IO, | |
85 | in dio_blocks units */ | |
1da177e4 LT |
86 | |
87 | /* | |
88 | * Deferred addition of a page to the dio. These variables are | |
89 | * private to dio_send_cur_page(), submit_page_section() and | |
90 | * dio_bio_add_page(). | |
91 | */ | |
92 | struct page *cur_page; /* The page */ | |
93 | unsigned cur_page_offset; /* Offset into it, in bytes */ | |
94 | unsigned cur_page_len; /* Nr of bytes at cur_page_offset */ | |
95 | sector_t cur_page_block; /* Where it starts */ | |
facd07b0 | 96 | loff_t cur_page_fs_offset; /* Offset in file */ |
1da177e4 | 97 | |
7b2c99d1 | 98 | struct iov_iter *iter; |
1da177e4 LT |
99 | /* |
100 | * Page queue. These variables belong to dio_refill_pages() and | |
101 | * dio_get_page(). | |
102 | */ | |
1da177e4 LT |
103 | unsigned head; /* next page to process */ |
104 | unsigned tail; /* last valid page + 1 */ | |
7b2c99d1 | 105 | size_t from, to; |
eb28be2b AK |
106 | }; |
107 | ||
108 | /* dio_state communicated between submission path and end_io */ | |
109 | struct dio { | |
110 | int flags; /* doesn't change */ | |
8a4c1e42 MC |
111 | int op; |
112 | int op_flags; | |
15c4f638 | 113 | blk_qc_t bio_cookie; |
74d46992 | 114 | struct gendisk *bio_disk; |
0dc2bc49 | 115 | struct inode *inode; |
eb28be2b AK |
116 | loff_t i_size; /* i_size when submitted */ |
117 | dio_iodone_t *end_io; /* IO completion function */ | |
eb28be2b | 118 | |
18772641 | 119 | void *private; /* copy from map_bh.b_private */ |
eb28be2b AK |
120 | |
121 | /* BIO completion state */ | |
122 | spinlock_t bio_lock; /* protects BIO fields below */ | |
0dc2bc49 AK |
123 | int page_errors; /* errno from get_user_pages() */ |
124 | int is_async; /* is IO async ? */ | |
7b7a8665 | 125 | bool defer_completion; /* defer AIO completion to workqueue? */ |
53cbf3b1 | 126 | bool should_dirty; /* if pages should be dirtied */ |
0dc2bc49 | 127 | int io_error; /* IO error in completion path */ |
eb28be2b AK |
128 | unsigned long refcount; /* direct_io_worker() and bios */ |
129 | struct bio *bio_list; /* singly linked via bi_private */ | |
130 | struct task_struct *waiter; /* waiting task (NULL if none) */ | |
131 | ||
132 | /* AIO related stuff */ | |
133 | struct kiocb *iocb; /* kiocb */ | |
eb28be2b AK |
134 | ssize_t result; /* IO result */ |
135 | ||
23aee091 JM |
136 | /* |
137 | * pages[] (and any fields placed after it) are not zeroed out at | |
138 | * allocation time. Don't add new fields after pages[] unless you | |
139 | * wish that they not be zeroed. | |
140 | */ | |
7b7a8665 CH |
141 | union { |
142 | struct page *pages[DIO_PAGES]; /* page buffer */ | |
143 | struct work_struct complete_work;/* deferred AIO completion */ | |
144 | }; | |
6e8267f5 AK |
145 | } ____cacheline_aligned_in_smp; |
146 | ||
147 | static struct kmem_cache *dio_cache __read_mostly; | |
1da177e4 LT |
148 | |
149 | /* | |
150 | * How many pages are in the queue? | |
151 | */ | |
eb28be2b | 152 | static inline unsigned dio_pages_present(struct dio_submit *sdio) |
1da177e4 | 153 | { |
eb28be2b | 154 | return sdio->tail - sdio->head; |
1da177e4 LT |
155 | } |
156 | ||
157 | /* | |
158 | * Go grab and pin some userspace pages. Typically we'll get 64 at a time. | |
159 | */ | |
ba253fbf | 160 | static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio) |
1da177e4 | 161 | { |
7b2c99d1 | 162 | ssize_t ret; |
1da177e4 | 163 | |
2c80929c | 164 | ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES, |
7b2c99d1 | 165 | &sdio->from); |
1da177e4 | 166 | |
8a4c1e42 | 167 | if (ret < 0 && sdio->blocks_available && (dio->op == REQ_OP_WRITE)) { |
557ed1fa | 168 | struct page *page = ZERO_PAGE(0); |
1da177e4 LT |
169 | /* |
170 | * A memory fault, but the filesystem has some outstanding | |
171 | * mapped blocks. We need to use those blocks up to avoid | |
172 | * leaking stale data in the file. | |
173 | */ | |
174 | if (dio->page_errors == 0) | |
175 | dio->page_errors = ret; | |
09cbfeaf | 176 | get_page(page); |
b5810039 | 177 | dio->pages[0] = page; |
eb28be2b AK |
178 | sdio->head = 0; |
179 | sdio->tail = 1; | |
7b2c99d1 AV |
180 | sdio->from = 0; |
181 | sdio->to = PAGE_SIZE; | |
182 | return 0; | |
1da177e4 LT |
183 | } |
184 | ||
185 | if (ret >= 0) { | |
7b2c99d1 AV |
186 | iov_iter_advance(sdio->iter, ret); |
187 | ret += sdio->from; | |
eb28be2b | 188 | sdio->head = 0; |
7b2c99d1 AV |
189 | sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE; |
190 | sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1; | |
191 | return 0; | |
1da177e4 | 192 | } |
1da177e4 LT |
193 | return ret; |
194 | } | |
195 | ||
196 | /* | |
197 | * Get another userspace page. Returns an ERR_PTR on error. Pages are | |
198 | * buffered inside the dio so that we can call get_user_pages() against a | |
199 | * decent number of pages, less frequently. To provide nicer use of the | |
200 | * L1 cache. | |
201 | */ | |
ba253fbf | 202 | static inline struct page *dio_get_page(struct dio *dio, |
6fcc5420 | 203 | struct dio_submit *sdio) |
1da177e4 | 204 | { |
eb28be2b | 205 | if (dio_pages_present(sdio) == 0) { |
1da177e4 LT |
206 | int ret; |
207 | ||
eb28be2b | 208 | ret = dio_refill_pages(dio, sdio); |
1da177e4 LT |
209 | if (ret) |
210 | return ERR_PTR(ret); | |
eb28be2b | 211 | BUG_ON(dio_pages_present(sdio) == 0); |
1da177e4 | 212 | } |
6fcc5420 | 213 | return dio->pages[sdio->head]; |
1da177e4 LT |
214 | } |
215 | ||
6d544bb4 ZB |
216 | /** |
217 | * dio_complete() - called when all DIO BIO I/O has been completed | |
218 | * @offset: the byte offset in the file of the completed operation | |
219 | * | |
7b7a8665 CH |
220 | * This drops i_dio_count, lets interested parties know that a DIO operation |
221 | * has completed, and calculates the resulting return code for the operation. | |
6d544bb4 ZB |
222 | * |
223 | * It lets the filesystem know if it registered an interest earlier via | |
224 | * get_block. Pass the private field of the map buffer_head so that | |
225 | * filesystems can use it to hold additional state between get_block calls and | |
226 | * dio_complete. | |
1da177e4 | 227 | */ |
716b9bc0 | 228 | static ssize_t dio_complete(struct dio *dio, ssize_t ret, bool is_async) |
1da177e4 | 229 | { |
716b9bc0 | 230 | loff_t offset = dio->iocb->ki_pos; |
6d544bb4 | 231 | ssize_t transferred = 0; |
332391a9 | 232 | int err; |
6d544bb4 | 233 | |
8459d86a ZB |
234 | /* |
235 | * AIO submission can race with bio completion to get here while | |
236 | * expecting to have the last io completed by bio completion. | |
237 | * In that case -EIOCBQUEUED is in fact not an error we want | |
238 | * to preserve through this call. | |
239 | */ | |
240 | if (ret == -EIOCBQUEUED) | |
241 | ret = 0; | |
242 | ||
6d544bb4 ZB |
243 | if (dio->result) { |
244 | transferred = dio->result; | |
245 | ||
246 | /* Check for short read case */ | |
8a4c1e42 MC |
247 | if ((dio->op == REQ_OP_READ) && |
248 | ((offset + transferred) > dio->i_size)) | |
6d544bb4 | 249 | transferred = dio->i_size - offset; |
4038acdb AV |
250 | /* ignore EFAULT if some IO has been done */ |
251 | if (unlikely(ret == -EFAULT) && transferred) | |
252 | ret = 0; | |
6d544bb4 ZB |
253 | } |
254 | ||
6d544bb4 ZB |
255 | if (ret == 0) |
256 | ret = dio->page_errors; | |
257 | if (ret == 0) | |
258 | ret = dio->io_error; | |
259 | if (ret == 0) | |
260 | ret = transferred; | |
261 | ||
332391a9 LC |
262 | /* |
263 | * Try again to invalidate clean pages which might have been cached by | |
264 | * non-direct readahead, or faulted in by get_user_pages() if the source | |
265 | * of the write was an mmap'ed region of the file we're writing. Either | |
266 | * one is a pretty crazy thing to do, so we don't support it 100%. If | |
267 | * this invalidation fails, tough, the write still worked... | |
268 | */ | |
269 | if (ret > 0 && dio->op == REQ_OP_WRITE && | |
270 | dio->inode->i_mapping->nrpages) { | |
271 | err = invalidate_inode_pages2_range(dio->inode->i_mapping, | |
272 | offset >> PAGE_SHIFT, | |
273 | (offset + ret - 1) >> PAGE_SHIFT); | |
274 | WARN_ON_ONCE(err); | |
275 | } | |
276 | ||
187372a3 | 277 | if (dio->end_io) { |
187372a3 | 278 | |
e2592217 | 279 | // XXX: ki_pos?? |
187372a3 CH |
280 | err = dio->end_io(dio->iocb, offset, ret, dio->private); |
281 | if (err) | |
282 | ret = err; | |
283 | } | |
7b7a8665 | 284 | |
fe0f07d0 JA |
285 | if (!(dio->flags & DIO_SKIP_DIO_COUNT)) |
286 | inode_dio_end(dio->inode); | |
287 | ||
02afc27f | 288 | if (is_async) { |
e2592217 CH |
289 | /* |
290 | * generic_write_sync expects ki_pos to have been updated | |
291 | * already, but the submission path only does this for | |
292 | * synchronous I/O. | |
293 | */ | |
294 | dio->iocb->ki_pos += transferred; | |
02afc27f | 295 | |
8a4c1e42 | 296 | if (dio->op == REQ_OP_WRITE) |
e2592217 | 297 | ret = generic_write_sync(dio->iocb, transferred); |
04b2fa9f | 298 | dio->iocb->ki_complete(dio->iocb, ret, 0); |
02afc27f | 299 | } |
40e2e973 | 300 | |
7b7a8665 | 301 | kmem_cache_free(dio_cache, dio); |
6d544bb4 | 302 | return ret; |
1da177e4 LT |
303 | } |
304 | ||
7b7a8665 CH |
305 | static void dio_aio_complete_work(struct work_struct *work) |
306 | { | |
307 | struct dio *dio = container_of(work, struct dio, complete_work); | |
308 | ||
716b9bc0 | 309 | dio_complete(dio, 0, true); |
7b7a8665 CH |
310 | } |
311 | ||
4e4cbee9 | 312 | static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio); |
7b7a8665 | 313 | |
1da177e4 LT |
314 | /* |
315 | * Asynchronous IO callback. | |
316 | */ | |
4246a0b6 | 317 | static void dio_bio_end_aio(struct bio *bio) |
1da177e4 LT |
318 | { |
319 | struct dio *dio = bio->bi_private; | |
5eb6c7a2 ZB |
320 | unsigned long remaining; |
321 | unsigned long flags; | |
332391a9 | 322 | bool defer_completion = false; |
1da177e4 | 323 | |
1da177e4 LT |
324 | /* cleanup the bio */ |
325 | dio_bio_complete(dio, bio); | |
0273201e | 326 | |
5eb6c7a2 ZB |
327 | spin_lock_irqsave(&dio->bio_lock, flags); |
328 | remaining = --dio->refcount; | |
329 | if (remaining == 1 && dio->waiter) | |
20258b2b | 330 | wake_up_process(dio->waiter); |
5eb6c7a2 | 331 | spin_unlock_irqrestore(&dio->bio_lock, flags); |
20258b2b | 332 | |
8459d86a | 333 | if (remaining == 0) { |
332391a9 LC |
334 | /* |
335 | * Defer completion when defer_completion is set or | |
336 | * when the inode has pages mapped and this is AIO write. | |
337 | * We need to invalidate those pages because there is a | |
338 | * chance they contain stale data in the case buffered IO | |
339 | * went in between AIO submission and completion into the | |
340 | * same region. | |
341 | */ | |
342 | if (dio->result) | |
343 | defer_completion = dio->defer_completion || | |
344 | (dio->op == REQ_OP_WRITE && | |
345 | dio->inode->i_mapping->nrpages); | |
346 | if (defer_completion) { | |
7b7a8665 CH |
347 | INIT_WORK(&dio->complete_work, dio_aio_complete_work); |
348 | queue_work(dio->inode->i_sb->s_dio_done_wq, | |
349 | &dio->complete_work); | |
350 | } else { | |
716b9bc0 | 351 | dio_complete(dio, 0, true); |
7b7a8665 | 352 | } |
8459d86a | 353 | } |
1da177e4 LT |
354 | } |
355 | ||
356 | /* | |
357 | * The BIO completion handler simply queues the BIO up for the process-context | |
358 | * handler. | |
359 | * | |
360 | * During I/O bi_private points at the dio. After I/O, bi_private is used to | |
361 | * implement a singly-linked list of completed BIOs, at dio->bio_list. | |
362 | */ | |
4246a0b6 | 363 | static void dio_bio_end_io(struct bio *bio) |
1da177e4 LT |
364 | { |
365 | struct dio *dio = bio->bi_private; | |
366 | unsigned long flags; | |
367 | ||
1da177e4 LT |
368 | spin_lock_irqsave(&dio->bio_lock, flags); |
369 | bio->bi_private = dio->bio_list; | |
370 | dio->bio_list = bio; | |
5eb6c7a2 | 371 | if (--dio->refcount == 1 && dio->waiter) |
1da177e4 LT |
372 | wake_up_process(dio->waiter); |
373 | spin_unlock_irqrestore(&dio->bio_lock, flags); | |
1da177e4 LT |
374 | } |
375 | ||
facd07b0 JB |
376 | /** |
377 | * dio_end_io - handle the end io action for the given bio | |
378 | * @bio: The direct io bio thats being completed | |
facd07b0 JB |
379 | * |
380 | * This is meant to be called by any filesystem that uses their own dio_submit_t | |
381 | * so that the DIO specific endio actions are dealt with after the filesystem | |
382 | * has done it's completion work. | |
383 | */ | |
4055351c | 384 | void dio_end_io(struct bio *bio) |
facd07b0 JB |
385 | { |
386 | struct dio *dio = bio->bi_private; | |
387 | ||
388 | if (dio->is_async) | |
4246a0b6 | 389 | dio_bio_end_aio(bio); |
facd07b0 | 390 | else |
4246a0b6 | 391 | dio_bio_end_io(bio); |
facd07b0 JB |
392 | } |
393 | EXPORT_SYMBOL_GPL(dio_end_io); | |
394 | ||
ba253fbf | 395 | static inline void |
eb28be2b AK |
396 | dio_bio_alloc(struct dio *dio, struct dio_submit *sdio, |
397 | struct block_device *bdev, | |
398 | sector_t first_sector, int nr_vecs) | |
1da177e4 LT |
399 | { |
400 | struct bio *bio; | |
401 | ||
20d9600c DD |
402 | /* |
403 | * bio_alloc() is guaranteed to return a bio when called with | |
71baba4b | 404 | * __GFP_RECLAIM and we request a valid number of vectors. |
20d9600c | 405 | */ |
1da177e4 | 406 | bio = bio_alloc(GFP_KERNEL, nr_vecs); |
1da177e4 | 407 | |
74d46992 | 408 | bio_set_dev(bio, bdev); |
4f024f37 | 409 | bio->bi_iter.bi_sector = first_sector; |
8a4c1e42 | 410 | bio_set_op_attrs(bio, dio->op, dio->op_flags); |
1da177e4 LT |
411 | if (dio->is_async) |
412 | bio->bi_end_io = dio_bio_end_aio; | |
413 | else | |
414 | bio->bi_end_io = dio_bio_end_io; | |
415 | ||
45d06cf7 JA |
416 | bio->bi_write_hint = dio->iocb->ki_hint; |
417 | ||
eb28be2b AK |
418 | sdio->bio = bio; |
419 | sdio->logical_offset_in_bio = sdio->cur_page_fs_offset; | |
1da177e4 LT |
420 | } |
421 | ||
422 | /* | |
423 | * In the AIO read case we speculatively dirty the pages before starting IO. | |
424 | * During IO completion, any of these pages which happen to have been written | |
425 | * back will be redirtied by bio_check_pages_dirty(). | |
0273201e ZB |
426 | * |
427 | * bios hold a dio reference between submit_bio and ->end_io. | |
1da177e4 | 428 | */ |
ba253fbf | 429 | static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio) |
1da177e4 | 430 | { |
eb28be2b | 431 | struct bio *bio = sdio->bio; |
5eb6c7a2 | 432 | unsigned long flags; |
1da177e4 LT |
433 | |
434 | bio->bi_private = dio; | |
5eb6c7a2 ZB |
435 | |
436 | spin_lock_irqsave(&dio->bio_lock, flags); | |
437 | dio->refcount++; | |
438 | spin_unlock_irqrestore(&dio->bio_lock, flags); | |
439 | ||
8a4c1e42 | 440 | if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty) |
1da177e4 | 441 | bio_set_pages_dirty(bio); |
5eb6c7a2 | 442 | |
74d46992 | 443 | dio->bio_disk = bio->bi_disk; |
c1c53460 | 444 | |
15c4f638 | 445 | if (sdio->submit_io) { |
8a4c1e42 | 446 | sdio->submit_io(bio, dio->inode, sdio->logical_offset_in_bio); |
15c4f638 | 447 | dio->bio_cookie = BLK_QC_T_NONE; |
c1c53460 | 448 | } else |
4e49ea4a | 449 | dio->bio_cookie = submit_bio(bio); |
1da177e4 | 450 | |
eb28be2b AK |
451 | sdio->bio = NULL; |
452 | sdio->boundary = 0; | |
453 | sdio->logical_offset_in_bio = 0; | |
1da177e4 LT |
454 | } |
455 | ||
456 | /* | |
457 | * Release any resources in case of a failure | |
458 | */ | |
ba253fbf | 459 | static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio) |
1da177e4 | 460 | { |
7b2c99d1 | 461 | while (sdio->head < sdio->tail) |
09cbfeaf | 462 | put_page(dio->pages[sdio->head++]); |
1da177e4 LT |
463 | } |
464 | ||
465 | /* | |
0273201e ZB |
466 | * Wait for the next BIO to complete. Remove it and return it. NULL is |
467 | * returned once all BIOs have been completed. This must only be called once | |
468 | * all bios have been issued so that dio->refcount can only decrease. This | |
469 | * requires that that the caller hold a reference on the dio. | |
1da177e4 LT |
470 | */ |
471 | static struct bio *dio_await_one(struct dio *dio) | |
472 | { | |
473 | unsigned long flags; | |
0273201e | 474 | struct bio *bio = NULL; |
1da177e4 LT |
475 | |
476 | spin_lock_irqsave(&dio->bio_lock, flags); | |
5eb6c7a2 ZB |
477 | |
478 | /* | |
479 | * Wait as long as the list is empty and there are bios in flight. bio | |
480 | * completion drops the count, maybe adds to the list, and wakes while | |
481 | * holding the bio_lock so we don't need set_current_state()'s barrier | |
482 | * and can call it after testing our condition. | |
483 | */ | |
484 | while (dio->refcount > 1 && dio->bio_list == NULL) { | |
485 | __set_current_state(TASK_UNINTERRUPTIBLE); | |
486 | dio->waiter = current; | |
487 | spin_unlock_irqrestore(&dio->bio_lock, flags); | |
c43c83a2 | 488 | if (!(dio->iocb->ki_flags & IOCB_HIPRI) || |
74d46992 | 489 | !blk_mq_poll(dio->bio_disk->queue, dio->bio_cookie)) |
15c4f638 | 490 | io_schedule(); |
5eb6c7a2 ZB |
491 | /* wake up sets us TASK_RUNNING */ |
492 | spin_lock_irqsave(&dio->bio_lock, flags); | |
493 | dio->waiter = NULL; | |
1da177e4 | 494 | } |
0273201e ZB |
495 | if (dio->bio_list) { |
496 | bio = dio->bio_list; | |
497 | dio->bio_list = bio->bi_private; | |
498 | } | |
1da177e4 LT |
499 | spin_unlock_irqrestore(&dio->bio_lock, flags); |
500 | return bio; | |
501 | } | |
502 | ||
503 | /* | |
504 | * Process one completed BIO. No locks are held. | |
505 | */ | |
4e4cbee9 | 506 | static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio) |
1da177e4 | 507 | { |
cb34e057 KO |
508 | struct bio_vec *bvec; |
509 | unsigned i; | |
4e4cbee9 | 510 | blk_status_t err = bio->bi_status; |
1da177e4 | 511 | |
03a07c92 GR |
512 | if (err) { |
513 | if (err == BLK_STS_AGAIN && (bio->bi_opf & REQ_NOWAIT)) | |
514 | dio->io_error = -EAGAIN; | |
515 | else | |
516 | dio->io_error = -EIO; | |
517 | } | |
1da177e4 | 518 | |
8a4c1e42 | 519 | if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty) { |
7ddc971f | 520 | bio_check_pages_dirty(bio); /* transfers ownership */ |
1da177e4 | 521 | } else { |
cb34e057 KO |
522 | bio_for_each_segment_all(bvec, bio, i) { |
523 | struct page *page = bvec->bv_page; | |
1da177e4 | 524 | |
8a4c1e42 | 525 | if (dio->op == REQ_OP_READ && !PageCompound(page) && |
53cbf3b1 | 526 | dio->should_dirty) |
1da177e4 | 527 | set_page_dirty_lock(page); |
09cbfeaf | 528 | put_page(page); |
1da177e4 LT |
529 | } |
530 | bio_put(bio); | |
531 | } | |
9b81c842 | 532 | return err; |
1da177e4 LT |
533 | } |
534 | ||
535 | /* | |
0273201e ZB |
536 | * Wait on and process all in-flight BIOs. This must only be called once |
537 | * all bios have been issued so that the refcount can only decrease. | |
538 | * This just waits for all bios to make it through dio_bio_complete. IO | |
beb7dd86 | 539 | * errors are propagated through dio->io_error and should be propagated via |
0273201e | 540 | * dio_complete(). |
1da177e4 | 541 | */ |
6d544bb4 | 542 | static void dio_await_completion(struct dio *dio) |
1da177e4 | 543 | { |
0273201e ZB |
544 | struct bio *bio; |
545 | do { | |
546 | bio = dio_await_one(dio); | |
547 | if (bio) | |
548 | dio_bio_complete(dio, bio); | |
549 | } while (bio); | |
1da177e4 LT |
550 | } |
551 | ||
552 | /* | |
553 | * A really large O_DIRECT read or write can generate a lot of BIOs. So | |
554 | * to keep the memory consumption sane we periodically reap any completed BIOs | |
555 | * during the BIO generation phase. | |
556 | * | |
557 | * This also helps to limit the peak amount of pinned userspace memory. | |
558 | */ | |
ba253fbf | 559 | static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio) |
1da177e4 LT |
560 | { |
561 | int ret = 0; | |
562 | ||
eb28be2b | 563 | if (sdio->reap_counter++ >= 64) { |
1da177e4 LT |
564 | while (dio->bio_list) { |
565 | unsigned long flags; | |
566 | struct bio *bio; | |
567 | int ret2; | |
568 | ||
569 | spin_lock_irqsave(&dio->bio_lock, flags); | |
570 | bio = dio->bio_list; | |
571 | dio->bio_list = bio->bi_private; | |
572 | spin_unlock_irqrestore(&dio->bio_lock, flags); | |
4e4cbee9 | 573 | ret2 = blk_status_to_errno(dio_bio_complete(dio, bio)); |
1da177e4 LT |
574 | if (ret == 0) |
575 | ret = ret2; | |
576 | } | |
eb28be2b | 577 | sdio->reap_counter = 0; |
1da177e4 LT |
578 | } |
579 | return ret; | |
580 | } | |
581 | ||
7b7a8665 CH |
582 | /* |
583 | * Create workqueue for deferred direct IO completions. We allocate the | |
584 | * workqueue when it's first needed. This avoids creating workqueue for | |
585 | * filesystems that don't need it and also allows us to create the workqueue | |
586 | * late enough so the we can include s_id in the name of the workqueue. | |
587 | */ | |
ec1b8260 | 588 | int sb_init_dio_done_wq(struct super_block *sb) |
7b7a8665 | 589 | { |
45150c43 | 590 | struct workqueue_struct *old; |
7b7a8665 CH |
591 | struct workqueue_struct *wq = alloc_workqueue("dio/%s", |
592 | WQ_MEM_RECLAIM, 0, | |
593 | sb->s_id); | |
594 | if (!wq) | |
595 | return -ENOMEM; | |
596 | /* | |
597 | * This has to be atomic as more DIOs can race to create the workqueue | |
598 | */ | |
45150c43 | 599 | old = cmpxchg(&sb->s_dio_done_wq, NULL, wq); |
7b7a8665 | 600 | /* Someone created workqueue before us? Free ours... */ |
45150c43 | 601 | if (old) |
7b7a8665 CH |
602 | destroy_workqueue(wq); |
603 | return 0; | |
604 | } | |
605 | ||
606 | static int dio_set_defer_completion(struct dio *dio) | |
607 | { | |
608 | struct super_block *sb = dio->inode->i_sb; | |
609 | ||
610 | if (dio->defer_completion) | |
611 | return 0; | |
612 | dio->defer_completion = true; | |
613 | if (!sb->s_dio_done_wq) | |
614 | return sb_init_dio_done_wq(sb); | |
615 | return 0; | |
616 | } | |
617 | ||
1da177e4 LT |
618 | /* |
619 | * Call into the fs to map some more disk blocks. We record the current number | |
eb28be2b | 620 | * of available blocks at sdio->blocks_available. These are in units of the |
93407472 | 621 | * fs blocksize, i_blocksize(inode). |
1da177e4 LT |
622 | * |
623 | * The fs is allowed to map lots of blocks at once. If it wants to do that, | |
624 | * it uses the passed inode-relative block number as the file offset, as usual. | |
625 | * | |
1d8fa7a2 | 626 | * get_block() is passed the number of i_blkbits-sized blocks which direct_io |
1da177e4 LT |
627 | * has remaining to do. The fs should not map more than this number of blocks. |
628 | * | |
629 | * If the fs has mapped a lot of blocks, it should populate bh->b_size to | |
630 | * indicate how much contiguous disk space has been made available at | |
631 | * bh->b_blocknr. | |
632 | * | |
633 | * If *any* of the mapped blocks are new, then the fs must set buffer_new(). | |
634 | * This isn't very efficient... | |
635 | * | |
636 | * In the case of filesystem holes: the fs may return an arbitrarily-large | |
637 | * hole by returning an appropriate value in b_size and by clearing | |
638 | * buffer_mapped(). However the direct-io code will only process holes one | |
1d8fa7a2 | 639 | * block at a time - it will repeatedly call get_block() as it walks the hole. |
1da177e4 | 640 | */ |
18772641 AK |
641 | static int get_more_blocks(struct dio *dio, struct dio_submit *sdio, |
642 | struct buffer_head *map_bh) | |
1da177e4 LT |
643 | { |
644 | int ret; | |
1da177e4 | 645 | sector_t fs_startblk; /* Into file, in filesystem-sized blocks */ |
ae55e1aa | 646 | sector_t fs_endblk; /* Into file, in filesystem-sized blocks */ |
1da177e4 | 647 | unsigned long fs_count; /* Number of filesystem-sized blocks */ |
1da177e4 | 648 | int create; |
ab73857e | 649 | unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor; |
1da177e4 LT |
650 | |
651 | /* | |
652 | * If there was a memory error and we've overwritten all the | |
653 | * mapped blocks then we can now return that memory error | |
654 | */ | |
655 | ret = dio->page_errors; | |
656 | if (ret == 0) { | |
eb28be2b AK |
657 | BUG_ON(sdio->block_in_file >= sdio->final_block_in_request); |
658 | fs_startblk = sdio->block_in_file >> sdio->blkfactor; | |
ae55e1aa TM |
659 | fs_endblk = (sdio->final_block_in_request - 1) >> |
660 | sdio->blkfactor; | |
661 | fs_count = fs_endblk - fs_startblk + 1; | |
1da177e4 | 662 | |
3c674e74 | 663 | map_bh->b_state = 0; |
ab73857e | 664 | map_bh->b_size = fs_count << i_blkbits; |
3c674e74 | 665 | |
5fe878ae | 666 | /* |
9ecd10b7 EG |
667 | * For writes that could fill holes inside i_size on a |
668 | * DIO_SKIP_HOLES filesystem we forbid block creations: only | |
669 | * overwrites are permitted. We will return early to the caller | |
670 | * once we see an unmapped buffer head returned, and the caller | |
671 | * will fall back to buffered I/O. | |
5fe878ae CH |
672 | * |
673 | * Otherwise the decision is left to the get_blocks method, | |
674 | * which may decide to handle it or also return an unmapped | |
675 | * buffer head. | |
676 | */ | |
8a4c1e42 | 677 | create = dio->op == REQ_OP_WRITE; |
5fe878ae | 678 | if (dio->flags & DIO_SKIP_HOLES) { |
9ecd10b7 EG |
679 | if (fs_startblk <= ((i_size_read(dio->inode) - 1) >> |
680 | i_blkbits)) | |
1da177e4 | 681 | create = 0; |
1da177e4 | 682 | } |
3c674e74 | 683 | |
eb28be2b | 684 | ret = (*sdio->get_block)(dio->inode, fs_startblk, |
1da177e4 | 685 | map_bh, create); |
18772641 AK |
686 | |
687 | /* Store for completion */ | |
688 | dio->private = map_bh->b_private; | |
7b7a8665 CH |
689 | |
690 | if (ret == 0 && buffer_defer_completion(map_bh)) | |
691 | ret = dio_set_defer_completion(dio); | |
1da177e4 LT |
692 | } |
693 | return ret; | |
694 | } | |
695 | ||
696 | /* | |
697 | * There is no bio. Make one now. | |
698 | */ | |
ba253fbf AK |
699 | static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio, |
700 | sector_t start_sector, struct buffer_head *map_bh) | |
1da177e4 LT |
701 | { |
702 | sector_t sector; | |
703 | int ret, nr_pages; | |
704 | ||
eb28be2b | 705 | ret = dio_bio_reap(dio, sdio); |
1da177e4 LT |
706 | if (ret) |
707 | goto out; | |
eb28be2b | 708 | sector = start_sector << (sdio->blkbits - 9); |
b54ffb73 | 709 | nr_pages = min(sdio->pages_in_io, BIO_MAX_PAGES); |
1da177e4 | 710 | BUG_ON(nr_pages <= 0); |
18772641 | 711 | dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages); |
eb28be2b | 712 | sdio->boundary = 0; |
1da177e4 LT |
713 | out: |
714 | return ret; | |
715 | } | |
716 | ||
717 | /* | |
718 | * Attempt to put the current chunk of 'cur_page' into the current BIO. If | |
719 | * that was successful then update final_block_in_bio and take a ref against | |
720 | * the just-added page. | |
721 | * | |
722 | * Return zero on success. Non-zero means the caller needs to start a new BIO. | |
723 | */ | |
ba253fbf | 724 | static inline int dio_bio_add_page(struct dio_submit *sdio) |
1da177e4 LT |
725 | { |
726 | int ret; | |
727 | ||
eb28be2b AK |
728 | ret = bio_add_page(sdio->bio, sdio->cur_page, |
729 | sdio->cur_page_len, sdio->cur_page_offset); | |
730 | if (ret == sdio->cur_page_len) { | |
1da177e4 LT |
731 | /* |
732 | * Decrement count only, if we are done with this page | |
733 | */ | |
eb28be2b AK |
734 | if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE) |
735 | sdio->pages_in_io--; | |
09cbfeaf | 736 | get_page(sdio->cur_page); |
eb28be2b AK |
737 | sdio->final_block_in_bio = sdio->cur_page_block + |
738 | (sdio->cur_page_len >> sdio->blkbits); | |
1da177e4 LT |
739 | ret = 0; |
740 | } else { | |
741 | ret = 1; | |
742 | } | |
743 | return ret; | |
744 | } | |
745 | ||
746 | /* | |
747 | * Put cur_page under IO. The section of cur_page which is described by | |
748 | * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page | |
749 | * starts on-disk at cur_page_block. | |
750 | * | |
751 | * We take a ref against the page here (on behalf of its presence in the bio). | |
752 | * | |
753 | * The caller of this function is responsible for removing cur_page from the | |
754 | * dio, and for dropping the refcount which came from that presence. | |
755 | */ | |
ba253fbf AK |
756 | static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio, |
757 | struct buffer_head *map_bh) | |
1da177e4 LT |
758 | { |
759 | int ret = 0; | |
760 | ||
eb28be2b AK |
761 | if (sdio->bio) { |
762 | loff_t cur_offset = sdio->cur_page_fs_offset; | |
763 | loff_t bio_next_offset = sdio->logical_offset_in_bio + | |
4f024f37 | 764 | sdio->bio->bi_iter.bi_size; |
c2c6ca41 | 765 | |
1da177e4 | 766 | /* |
c2c6ca41 JB |
767 | * See whether this new request is contiguous with the old. |
768 | * | |
f0940cee NK |
769 | * Btrfs cannot handle having logically non-contiguous requests |
770 | * submitted. For example if you have | |
c2c6ca41 JB |
771 | * |
772 | * Logical: [0-4095][HOLE][8192-12287] | |
f0940cee | 773 | * Physical: [0-4095] [4096-8191] |
c2c6ca41 JB |
774 | * |
775 | * We cannot submit those pages together as one BIO. So if our | |
776 | * current logical offset in the file does not equal what would | |
777 | * be the next logical offset in the bio, submit the bio we | |
778 | * have. | |
1da177e4 | 779 | */ |
eb28be2b | 780 | if (sdio->final_block_in_bio != sdio->cur_page_block || |
c2c6ca41 | 781 | cur_offset != bio_next_offset) |
eb28be2b | 782 | dio_bio_submit(dio, sdio); |
1da177e4 LT |
783 | } |
784 | ||
eb28be2b | 785 | if (sdio->bio == NULL) { |
18772641 | 786 | ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh); |
1da177e4 LT |
787 | if (ret) |
788 | goto out; | |
789 | } | |
790 | ||
eb28be2b AK |
791 | if (dio_bio_add_page(sdio) != 0) { |
792 | dio_bio_submit(dio, sdio); | |
18772641 | 793 | ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh); |
1da177e4 | 794 | if (ret == 0) { |
eb28be2b | 795 | ret = dio_bio_add_page(sdio); |
1da177e4 LT |
796 | BUG_ON(ret != 0); |
797 | } | |
798 | } | |
799 | out: | |
800 | return ret; | |
801 | } | |
802 | ||
803 | /* | |
804 | * An autonomous function to put a chunk of a page under deferred IO. | |
805 | * | |
806 | * The caller doesn't actually know (or care) whether this piece of page is in | |
807 | * a BIO, or is under IO or whatever. We just take care of all possible | |
808 | * situations here. The separation between the logic of do_direct_IO() and | |
809 | * that of submit_page_section() is important for clarity. Please don't break. | |
810 | * | |
811 | * The chunk of page starts on-disk at blocknr. | |
812 | * | |
813 | * We perform deferred IO, by recording the last-submitted page inside our | |
814 | * private part of the dio structure. If possible, we just expand the IO | |
815 | * across that page here. | |
816 | * | |
817 | * If that doesn't work out then we put the old page into the bio and add this | |
818 | * page to the dio instead. | |
819 | */ | |
ba253fbf | 820 | static inline int |
eb28be2b | 821 | submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page, |
18772641 AK |
822 | unsigned offset, unsigned len, sector_t blocknr, |
823 | struct buffer_head *map_bh) | |
1da177e4 LT |
824 | { |
825 | int ret = 0; | |
826 | ||
8a4c1e42 | 827 | if (dio->op == REQ_OP_WRITE) { |
98c4d57d AM |
828 | /* |
829 | * Read accounting is performed in submit_bio() | |
830 | */ | |
831 | task_io_account_write(len); | |
832 | } | |
833 | ||
1da177e4 LT |
834 | /* |
835 | * Can we just grow the current page's presence in the dio? | |
836 | */ | |
eb28be2b AK |
837 | if (sdio->cur_page == page && |
838 | sdio->cur_page_offset + sdio->cur_page_len == offset && | |
839 | sdio->cur_page_block + | |
840 | (sdio->cur_page_len >> sdio->blkbits) == blocknr) { | |
841 | sdio->cur_page_len += len; | |
1da177e4 LT |
842 | goto out; |
843 | } | |
844 | ||
845 | /* | |
846 | * If there's a deferred page already there then send it. | |
847 | */ | |
eb28be2b | 848 | if (sdio->cur_page) { |
18772641 | 849 | ret = dio_send_cur_page(dio, sdio, map_bh); |
09cbfeaf | 850 | put_page(sdio->cur_page); |
eb28be2b | 851 | sdio->cur_page = NULL; |
1da177e4 | 852 | if (ret) |
b1058b98 | 853 | return ret; |
1da177e4 LT |
854 | } |
855 | ||
09cbfeaf | 856 | get_page(page); /* It is in dio */ |
eb28be2b AK |
857 | sdio->cur_page = page; |
858 | sdio->cur_page_offset = offset; | |
859 | sdio->cur_page_len = len; | |
860 | sdio->cur_page_block = blocknr; | |
861 | sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits; | |
1da177e4 | 862 | out: |
b1058b98 JK |
863 | /* |
864 | * If sdio->boundary then we want to schedule the IO now to | |
865 | * avoid metadata seeks. | |
866 | */ | |
867 | if (sdio->boundary) { | |
868 | ret = dio_send_cur_page(dio, sdio, map_bh); | |
869 | dio_bio_submit(dio, sdio); | |
09cbfeaf | 870 | put_page(sdio->cur_page); |
b1058b98 JK |
871 | sdio->cur_page = NULL; |
872 | } | |
1da177e4 LT |
873 | return ret; |
874 | } | |
875 | ||
1da177e4 LT |
876 | /* |
877 | * If we are not writing the entire block and get_block() allocated | |
878 | * the block for us, we need to fill-in the unused portion of the | |
879 | * block with zeros. This happens only if user-buffer, fileoffset or | |
880 | * io length is not filesystem block-size multiple. | |
881 | * | |
882 | * `end' is zero if we're doing the start of the IO, 1 at the end of the | |
883 | * IO. | |
884 | */ | |
ba253fbf AK |
885 | static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio, |
886 | int end, struct buffer_head *map_bh) | |
1da177e4 LT |
887 | { |
888 | unsigned dio_blocks_per_fs_block; | |
889 | unsigned this_chunk_blocks; /* In dio_blocks */ | |
890 | unsigned this_chunk_bytes; | |
891 | struct page *page; | |
892 | ||
eb28be2b | 893 | sdio->start_zero_done = 1; |
18772641 | 894 | if (!sdio->blkfactor || !buffer_new(map_bh)) |
1da177e4 LT |
895 | return; |
896 | ||
eb28be2b AK |
897 | dio_blocks_per_fs_block = 1 << sdio->blkfactor; |
898 | this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1); | |
1da177e4 LT |
899 | |
900 | if (!this_chunk_blocks) | |
901 | return; | |
902 | ||
903 | /* | |
904 | * We need to zero out part of an fs block. It is either at the | |
905 | * beginning or the end of the fs block. | |
906 | */ | |
907 | if (end) | |
908 | this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks; | |
909 | ||
eb28be2b | 910 | this_chunk_bytes = this_chunk_blocks << sdio->blkbits; |
1da177e4 | 911 | |
557ed1fa | 912 | page = ZERO_PAGE(0); |
eb28be2b | 913 | if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes, |
18772641 | 914 | sdio->next_block_for_io, map_bh)) |
1da177e4 LT |
915 | return; |
916 | ||
eb28be2b | 917 | sdio->next_block_for_io += this_chunk_blocks; |
1da177e4 LT |
918 | } |
919 | ||
920 | /* | |
921 | * Walk the user pages, and the file, mapping blocks to disk and generating | |
922 | * a sequence of (page,offset,len,block) mappings. These mappings are injected | |
923 | * into submit_page_section(), which takes care of the next stage of submission | |
924 | * | |
925 | * Direct IO against a blockdev is different from a file. Because we can | |
926 | * happily perform page-sized but 512-byte aligned IOs. It is important that | |
927 | * blockdev IO be able to have fine alignment and large sizes. | |
928 | * | |
1d8fa7a2 | 929 | * So what we do is to permit the ->get_block function to populate bh.b_size |
1da177e4 LT |
930 | * with the size of IO which is permitted at this offset and this i_blkbits. |
931 | * | |
932 | * For best results, the blockdev should be set up with 512-byte i_blkbits and | |
1d8fa7a2 | 933 | * it should set b_size to PAGE_SIZE or more inside get_block(). This gives |
1da177e4 LT |
934 | * fine alignment but still allows this function to work in PAGE_SIZE units. |
935 | */ | |
18772641 AK |
936 | static int do_direct_IO(struct dio *dio, struct dio_submit *sdio, |
937 | struct buffer_head *map_bh) | |
1da177e4 | 938 | { |
eb28be2b | 939 | const unsigned blkbits = sdio->blkbits; |
dd545b52 | 940 | const unsigned i_blkbits = blkbits + sdio->blkfactor; |
1da177e4 LT |
941 | int ret = 0; |
942 | ||
eb28be2b | 943 | while (sdio->block_in_file < sdio->final_block_in_request) { |
7b2c99d1 AV |
944 | struct page *page; |
945 | size_t from, to; | |
6fcc5420 BH |
946 | |
947 | page = dio_get_page(dio, sdio); | |
1da177e4 LT |
948 | if (IS_ERR(page)) { |
949 | ret = PTR_ERR(page); | |
950 | goto out; | |
951 | } | |
6fcc5420 BH |
952 | from = sdio->head ? 0 : sdio->from; |
953 | to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE; | |
954 | sdio->head++; | |
1da177e4 | 955 | |
7b2c99d1 | 956 | while (from < to) { |
1da177e4 LT |
957 | unsigned this_chunk_bytes; /* # of bytes mapped */ |
958 | unsigned this_chunk_blocks; /* # of blocks */ | |
959 | unsigned u; | |
960 | ||
eb28be2b | 961 | if (sdio->blocks_available == 0) { |
1da177e4 LT |
962 | /* |
963 | * Need to go and map some more disk | |
964 | */ | |
965 | unsigned long blkmask; | |
966 | unsigned long dio_remainder; | |
967 | ||
18772641 | 968 | ret = get_more_blocks(dio, sdio, map_bh); |
1da177e4 | 969 | if (ret) { |
09cbfeaf | 970 | put_page(page); |
1da177e4 LT |
971 | goto out; |
972 | } | |
973 | if (!buffer_mapped(map_bh)) | |
974 | goto do_holes; | |
975 | ||
eb28be2b | 976 | sdio->blocks_available = |
f734c89c | 977 | map_bh->b_size >> blkbits; |
eb28be2b AK |
978 | sdio->next_block_for_io = |
979 | map_bh->b_blocknr << sdio->blkfactor; | |
f734c89c JK |
980 | if (buffer_new(map_bh)) { |
981 | clean_bdev_aliases( | |
982 | map_bh->b_bdev, | |
983 | map_bh->b_blocknr, | |
dd545b52 | 984 | map_bh->b_size >> i_blkbits); |
f734c89c | 985 | } |
1da177e4 | 986 | |
eb28be2b | 987 | if (!sdio->blkfactor) |
1da177e4 LT |
988 | goto do_holes; |
989 | ||
eb28be2b AK |
990 | blkmask = (1 << sdio->blkfactor) - 1; |
991 | dio_remainder = (sdio->block_in_file & blkmask); | |
1da177e4 LT |
992 | |
993 | /* | |
994 | * If we are at the start of IO and that IO | |
995 | * starts partway into a fs-block, | |
996 | * dio_remainder will be non-zero. If the IO | |
997 | * is a read then we can simply advance the IO | |
998 | * cursor to the first block which is to be | |
999 | * read. But if the IO is a write and the | |
1000 | * block was newly allocated we cannot do that; | |
1001 | * the start of the fs block must be zeroed out | |
1002 | * on-disk | |
1003 | */ | |
1004 | if (!buffer_new(map_bh)) | |
eb28be2b AK |
1005 | sdio->next_block_for_io += dio_remainder; |
1006 | sdio->blocks_available -= dio_remainder; | |
1da177e4 LT |
1007 | } |
1008 | do_holes: | |
1009 | /* Handle holes */ | |
1010 | if (!buffer_mapped(map_bh)) { | |
35dc8161 | 1011 | loff_t i_size_aligned; |
1da177e4 LT |
1012 | |
1013 | /* AKPM: eargh, -ENOTBLK is a hack */ | |
8a4c1e42 | 1014 | if (dio->op == REQ_OP_WRITE) { |
09cbfeaf | 1015 | put_page(page); |
1da177e4 LT |
1016 | return -ENOTBLK; |
1017 | } | |
1018 | ||
35dc8161 JM |
1019 | /* |
1020 | * Be sure to account for a partial block as the | |
1021 | * last block in the file | |
1022 | */ | |
1023 | i_size_aligned = ALIGN(i_size_read(dio->inode), | |
1024 | 1 << blkbits); | |
eb28be2b | 1025 | if (sdio->block_in_file >= |
35dc8161 | 1026 | i_size_aligned >> blkbits) { |
1da177e4 | 1027 | /* We hit eof */ |
09cbfeaf | 1028 | put_page(page); |
1da177e4 LT |
1029 | goto out; |
1030 | } | |
7b2c99d1 | 1031 | zero_user(page, from, 1 << blkbits); |
eb28be2b | 1032 | sdio->block_in_file++; |
7b2c99d1 | 1033 | from += 1 << blkbits; |
3320c60b | 1034 | dio->result += 1 << blkbits; |
1da177e4 LT |
1035 | goto next_block; |
1036 | } | |
1037 | ||
1038 | /* | |
1039 | * If we're performing IO which has an alignment which | |
1040 | * is finer than the underlying fs, go check to see if | |
1041 | * we must zero out the start of this block. | |
1042 | */ | |
eb28be2b | 1043 | if (unlikely(sdio->blkfactor && !sdio->start_zero_done)) |
18772641 | 1044 | dio_zero_block(dio, sdio, 0, map_bh); |
1da177e4 LT |
1045 | |
1046 | /* | |
1047 | * Work out, in this_chunk_blocks, how much disk we | |
1048 | * can add to this page | |
1049 | */ | |
eb28be2b | 1050 | this_chunk_blocks = sdio->blocks_available; |
7b2c99d1 | 1051 | u = (to - from) >> blkbits; |
1da177e4 LT |
1052 | if (this_chunk_blocks > u) |
1053 | this_chunk_blocks = u; | |
eb28be2b | 1054 | u = sdio->final_block_in_request - sdio->block_in_file; |
1da177e4 LT |
1055 | if (this_chunk_blocks > u) |
1056 | this_chunk_blocks = u; | |
1057 | this_chunk_bytes = this_chunk_blocks << blkbits; | |
1058 | BUG_ON(this_chunk_bytes == 0); | |
1059 | ||
092c8d46 JK |
1060 | if (this_chunk_blocks == sdio->blocks_available) |
1061 | sdio->boundary = buffer_boundary(map_bh); | |
eb28be2b | 1062 | ret = submit_page_section(dio, sdio, page, |
7b2c99d1 | 1063 | from, |
eb28be2b | 1064 | this_chunk_bytes, |
18772641 AK |
1065 | sdio->next_block_for_io, |
1066 | map_bh); | |
1da177e4 | 1067 | if (ret) { |
09cbfeaf | 1068 | put_page(page); |
1da177e4 LT |
1069 | goto out; |
1070 | } | |
eb28be2b | 1071 | sdio->next_block_for_io += this_chunk_blocks; |
1da177e4 | 1072 | |
eb28be2b | 1073 | sdio->block_in_file += this_chunk_blocks; |
7b2c99d1 AV |
1074 | from += this_chunk_bytes; |
1075 | dio->result += this_chunk_bytes; | |
eb28be2b | 1076 | sdio->blocks_available -= this_chunk_blocks; |
1da177e4 | 1077 | next_block: |
eb28be2b AK |
1078 | BUG_ON(sdio->block_in_file > sdio->final_block_in_request); |
1079 | if (sdio->block_in_file == sdio->final_block_in_request) | |
1da177e4 LT |
1080 | break; |
1081 | } | |
1082 | ||
1083 | /* Drop the ref which was taken in get_user_pages() */ | |
09cbfeaf | 1084 | put_page(page); |
1da177e4 LT |
1085 | } |
1086 | out: | |
1087 | return ret; | |
1088 | } | |
1089 | ||
847cc637 | 1090 | static inline int drop_refcount(struct dio *dio) |
1da177e4 | 1091 | { |
847cc637 | 1092 | int ret2; |
5eb6c7a2 | 1093 | unsigned long flags; |
1da177e4 | 1094 | |
8459d86a ZB |
1095 | /* |
1096 | * Sync will always be dropping the final ref and completing the | |
5eb6c7a2 ZB |
1097 | * operation. AIO can if it was a broken operation described above or |
1098 | * in fact if all the bios race to complete before we get here. In | |
1099 | * that case dio_complete() translates the EIOCBQUEUED into the proper | |
04b2fa9f | 1100 | * return code that the caller will hand to ->complete(). |
5eb6c7a2 ZB |
1101 | * |
1102 | * This is managed by the bio_lock instead of being an atomic_t so that | |
1103 | * completion paths can drop their ref and use the remaining count to | |
1104 | * decide to wake the submission path atomically. | |
8459d86a | 1105 | */ |
5eb6c7a2 ZB |
1106 | spin_lock_irqsave(&dio->bio_lock, flags); |
1107 | ret2 = --dio->refcount; | |
1108 | spin_unlock_irqrestore(&dio->bio_lock, flags); | |
847cc637 | 1109 | return ret2; |
1da177e4 LT |
1110 | } |
1111 | ||
eafdc7d1 CH |
1112 | /* |
1113 | * This is a library function for use by filesystem drivers. | |
1114 | * | |
1115 | * The locking rules are governed by the flags parameter: | |
1116 | * - if the flags value contains DIO_LOCKING we use a fancy locking | |
1117 | * scheme for dumb filesystems. | |
1118 | * For writes this function is called under i_mutex and returns with | |
1119 | * i_mutex held, for reads, i_mutex is not held on entry, but it is | |
1120 | * taken and dropped again before returning. | |
eafdc7d1 CH |
1121 | * - if the flags value does NOT contain DIO_LOCKING we don't use any |
1122 | * internal locking but rather rely on the filesystem to synchronize | |
1123 | * direct I/O reads/writes versus each other and truncate. | |
df2d6f26 CH |
1124 | * |
1125 | * To help with locking against truncate we incremented the i_dio_count | |
1126 | * counter before starting direct I/O, and decrement it once we are done. | |
1127 | * Truncate can wait for it to reach zero to provide exclusion. It is | |
1128 | * expected that filesystem provide exclusion between new direct I/O | |
1129 | * and truncates. For DIO_LOCKING filesystems this is done by i_mutex, | |
1130 | * but other filesystems need to take care of this on their own. | |
ba253fbf AK |
1131 | * |
1132 | * NOTE: if you pass "sdio" to anything by pointer make sure that function | |
1133 | * is always inlined. Otherwise gcc is unable to split the structure into | |
1134 | * individual fields and will generate much worse code. This is important | |
1135 | * for the whole file. | |
eafdc7d1 | 1136 | */ |
65dd2aa9 | 1137 | static inline ssize_t |
17f8c842 OS |
1138 | do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode, |
1139 | struct block_device *bdev, struct iov_iter *iter, | |
c8b8e32d | 1140 | get_block_t get_block, dio_iodone_t end_io, |
17f8c842 | 1141 | dio_submit_t submit_io, int flags) |
1da177e4 | 1142 | { |
ab73857e LT |
1143 | unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits); |
1144 | unsigned blkbits = i_blkbits; | |
1da177e4 LT |
1145 | unsigned blocksize_mask = (1 << blkbits) - 1; |
1146 | ssize_t retval = -EINVAL; | |
af436472 | 1147 | size_t count = iov_iter_count(iter); |
c8b8e32d | 1148 | loff_t offset = iocb->ki_pos; |
af436472 | 1149 | loff_t end = offset + count; |
1da177e4 | 1150 | struct dio *dio; |
eb28be2b | 1151 | struct dio_submit sdio = { 0, }; |
847cc637 | 1152 | struct buffer_head map_bh = { 0, }; |
647d1e4c | 1153 | struct blk_plug plug; |
886a3911 | 1154 | unsigned long align = offset | iov_iter_alignment(iter); |
1da177e4 | 1155 | |
65dd2aa9 AK |
1156 | /* |
1157 | * Avoid references to bdev if not absolutely needed to give | |
1158 | * the early prefetch in the caller enough time. | |
1159 | */ | |
1da177e4 | 1160 | |
886a3911 | 1161 | if (align & blocksize_mask) { |
1da177e4 | 1162 | if (bdev) |
65dd2aa9 | 1163 | blkbits = blksize_bits(bdev_logical_block_size(bdev)); |
1da177e4 | 1164 | blocksize_mask = (1 << blkbits) - 1; |
886a3911 | 1165 | if (align & blocksize_mask) |
1da177e4 LT |
1166 | goto out; |
1167 | } | |
1168 | ||
f9b5570d | 1169 | /* watch out for a 0 len io from a tricksy fs */ |
17f8c842 | 1170 | if (iov_iter_rw(iter) == READ && !iov_iter_count(iter)) |
f9b5570d CH |
1171 | return 0; |
1172 | ||
6e8267f5 | 1173 | dio = kmem_cache_alloc(dio_cache, GFP_KERNEL); |
1da177e4 LT |
1174 | retval = -ENOMEM; |
1175 | if (!dio) | |
1176 | goto out; | |
23aee091 JM |
1177 | /* |
1178 | * Believe it or not, zeroing out the page array caused a .5% | |
1179 | * performance regression in a database benchmark. So, we take | |
1180 | * care to only zero out what's needed. | |
1181 | */ | |
1182 | memset(dio, 0, offsetof(struct dio, pages)); | |
1da177e4 | 1183 | |
5fe878ae CH |
1184 | dio->flags = flags; |
1185 | if (dio->flags & DIO_LOCKING) { | |
17f8c842 | 1186 | if (iov_iter_rw(iter) == READ) { |
5fe878ae CH |
1187 | struct address_space *mapping = |
1188 | iocb->ki_filp->f_mapping; | |
1da177e4 | 1189 | |
5fe878ae | 1190 | /* will be released by direct_io_worker */ |
5955102c | 1191 | inode_lock(inode); |
1da177e4 LT |
1192 | |
1193 | retval = filemap_write_and_wait_range(mapping, offset, | |
1194 | end - 1); | |
1195 | if (retval) { | |
5955102c | 1196 | inode_unlock(inode); |
6e8267f5 | 1197 | kmem_cache_free(dio_cache, dio); |
1da177e4 LT |
1198 | goto out; |
1199 | } | |
1da177e4 | 1200 | } |
1da177e4 LT |
1201 | } |
1202 | ||
74cedf9b JK |
1203 | /* Once we sampled i_size check for reads beyond EOF */ |
1204 | dio->i_size = i_size_read(inode); | |
1205 | if (iov_iter_rw(iter) == READ && offset >= dio->i_size) { | |
1206 | if (dio->flags & DIO_LOCKING) | |
5955102c | 1207 | inode_unlock(inode); |
74cedf9b | 1208 | kmem_cache_free(dio_cache, dio); |
2d4594ac | 1209 | retval = 0; |
74cedf9b JK |
1210 | goto out; |
1211 | } | |
1212 | ||
1da177e4 | 1213 | /* |
60392573 CH |
1214 | * For file extending writes updating i_size before data writeouts |
1215 | * complete can expose uninitialized blocks in dumb filesystems. | |
1216 | * In that case we need to wait for I/O completion even if asked | |
1217 | * for an asynchronous write. | |
1da177e4 | 1218 | */ |
60392573 CH |
1219 | if (is_sync_kiocb(iocb)) |
1220 | dio->is_async = false; | |
1221 | else if (!(dio->flags & DIO_ASYNC_EXTEND) && | |
17f8c842 | 1222 | iov_iter_rw(iter) == WRITE && end > i_size_read(inode)) |
60392573 CH |
1223 | dio->is_async = false; |
1224 | else | |
1225 | dio->is_async = true; | |
1226 | ||
847cc637 | 1227 | dio->inode = inode; |
8a4c1e42 MC |
1228 | if (iov_iter_rw(iter) == WRITE) { |
1229 | dio->op = REQ_OP_WRITE; | |
70fd7614 | 1230 | dio->op_flags = REQ_SYNC | REQ_IDLE; |
03a07c92 GR |
1231 | if (iocb->ki_flags & IOCB_NOWAIT) |
1232 | dio->op_flags |= REQ_NOWAIT; | |
8a4c1e42 MC |
1233 | } else { |
1234 | dio->op = REQ_OP_READ; | |
1235 | } | |
02afc27f CH |
1236 | |
1237 | /* | |
1238 | * For AIO O_(D)SYNC writes we need to defer completions to a workqueue | |
1239 | * so that we can call ->fsync. | |
1240 | */ | |
332391a9 LC |
1241 | if (dio->is_async && iov_iter_rw(iter) == WRITE) { |
1242 | retval = 0; | |
1243 | if ((iocb->ki_filp->f_flags & O_DSYNC) || | |
1244 | IS_SYNC(iocb->ki_filp->f_mapping->host)) | |
1245 | retval = dio_set_defer_completion(dio); | |
1246 | else if (!dio->inode->i_sb->s_dio_done_wq) { | |
1247 | /* | |
1248 | * In case of AIO write racing with buffered read we | |
1249 | * need to defer completion. We can't decide this now, | |
1250 | * however the workqueue needs to be initialized here. | |
1251 | */ | |
1252 | retval = sb_init_dio_done_wq(dio->inode->i_sb); | |
1253 | } | |
02afc27f CH |
1254 | if (retval) { |
1255 | /* | |
1256 | * We grab i_mutex only for reads so we don't have | |
1257 | * to release it here | |
1258 | */ | |
1259 | kmem_cache_free(dio_cache, dio); | |
1260 | goto out; | |
1261 | } | |
1262 | } | |
1263 | ||
1264 | /* | |
1265 | * Will be decremented at I/O completion time. | |
1266 | */ | |
fe0f07d0 JA |
1267 | if (!(dio->flags & DIO_SKIP_DIO_COUNT)) |
1268 | inode_dio_begin(inode); | |
02afc27f CH |
1269 | |
1270 | retval = 0; | |
847cc637 | 1271 | sdio.blkbits = blkbits; |
ab73857e | 1272 | sdio.blkfactor = i_blkbits - blkbits; |
847cc637 AK |
1273 | sdio.block_in_file = offset >> blkbits; |
1274 | ||
1275 | sdio.get_block = get_block; | |
1276 | dio->end_io = end_io; | |
1277 | sdio.submit_io = submit_io; | |
1278 | sdio.final_block_in_bio = -1; | |
1279 | sdio.next_block_for_io = -1; | |
1280 | ||
1281 | dio->iocb = iocb; | |
847cc637 AK |
1282 | |
1283 | spin_lock_init(&dio->bio_lock); | |
1284 | dio->refcount = 1; | |
1285 | ||
53cbf3b1 | 1286 | dio->should_dirty = (iter->type == ITER_IOVEC); |
7b2c99d1 AV |
1287 | sdio.iter = iter; |
1288 | sdio.final_block_in_request = | |
1289 | (offset + iov_iter_count(iter)) >> blkbits; | |
1290 | ||
847cc637 AK |
1291 | /* |
1292 | * In case of non-aligned buffers, we may need 2 more | |
1293 | * pages since we need to zero out first and last block. | |
1294 | */ | |
1295 | if (unlikely(sdio.blkfactor)) | |
1296 | sdio.pages_in_io = 2; | |
1297 | ||
f67da30c | 1298 | sdio.pages_in_io += iov_iter_npages(iter, INT_MAX); |
847cc637 | 1299 | |
647d1e4c FW |
1300 | blk_start_plug(&plug); |
1301 | ||
7b2c99d1 AV |
1302 | retval = do_direct_IO(dio, &sdio, &map_bh); |
1303 | if (retval) | |
1304 | dio_cleanup(dio, &sdio); | |
847cc637 AK |
1305 | |
1306 | if (retval == -ENOTBLK) { | |
1307 | /* | |
1308 | * The remaining part of the request will be | |
1309 | * be handled by buffered I/O when we return | |
1310 | */ | |
1311 | retval = 0; | |
1312 | } | |
1313 | /* | |
1314 | * There may be some unwritten disk at the end of a part-written | |
1315 | * fs-block-sized block. Go zero that now. | |
1316 | */ | |
1317 | dio_zero_block(dio, &sdio, 1, &map_bh); | |
1318 | ||
1319 | if (sdio.cur_page) { | |
1320 | ssize_t ret2; | |
1321 | ||
1322 | ret2 = dio_send_cur_page(dio, &sdio, &map_bh); | |
1323 | if (retval == 0) | |
1324 | retval = ret2; | |
09cbfeaf | 1325 | put_page(sdio.cur_page); |
847cc637 AK |
1326 | sdio.cur_page = NULL; |
1327 | } | |
1328 | if (sdio.bio) | |
1329 | dio_bio_submit(dio, &sdio); | |
1330 | ||
647d1e4c FW |
1331 | blk_finish_plug(&plug); |
1332 | ||
847cc637 AK |
1333 | /* |
1334 | * It is possible that, we return short IO due to end of file. | |
1335 | * In that case, we need to release all the pages we got hold on. | |
1336 | */ | |
1337 | dio_cleanup(dio, &sdio); | |
1338 | ||
1339 | /* | |
1340 | * All block lookups have been performed. For READ requests | |
1341 | * we can let i_mutex go now that its achieved its purpose | |
1342 | * of protecting us from looking up uninitialized blocks. | |
1343 | */ | |
17f8c842 | 1344 | if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING)) |
5955102c | 1345 | inode_unlock(dio->inode); |
847cc637 AK |
1346 | |
1347 | /* | |
1348 | * The only time we want to leave bios in flight is when a successful | |
1349 | * partial aio read or full aio write have been setup. In that case | |
1350 | * bio completion will call aio_complete. The only time it's safe to | |
1351 | * call aio_complete is when we return -EIOCBQUEUED, so we key on that. | |
1352 | * This had *better* be the only place that raises -EIOCBQUEUED. | |
1353 | */ | |
1354 | BUG_ON(retval == -EIOCBQUEUED); | |
1355 | if (dio->is_async && retval == 0 && dio->result && | |
17f8c842 | 1356 | (iov_iter_rw(iter) == READ || dio->result == count)) |
847cc637 | 1357 | retval = -EIOCBQUEUED; |
af436472 | 1358 | else |
847cc637 AK |
1359 | dio_await_completion(dio); |
1360 | ||
1361 | if (drop_refcount(dio) == 0) { | |
716b9bc0 | 1362 | retval = dio_complete(dio, retval, false); |
847cc637 AK |
1363 | } else |
1364 | BUG_ON(retval != -EIOCBQUEUED); | |
1da177e4 | 1365 | |
7bb46a67 | 1366 | out: |
1367 | return retval; | |
1368 | } | |
65dd2aa9 | 1369 | |
17f8c842 OS |
1370 | ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode, |
1371 | struct block_device *bdev, struct iov_iter *iter, | |
c8b8e32d | 1372 | get_block_t get_block, |
17f8c842 OS |
1373 | dio_iodone_t end_io, dio_submit_t submit_io, |
1374 | int flags) | |
65dd2aa9 AK |
1375 | { |
1376 | /* | |
1377 | * The block device state is needed in the end to finally | |
1378 | * submit everything. Since it's likely to be cache cold | |
1379 | * prefetch it here as first thing to hide some of the | |
1380 | * latency. | |
1381 | * | |
1382 | * Attempt to prefetch the pieces we likely need later. | |
1383 | */ | |
1384 | prefetch(&bdev->bd_disk->part_tbl); | |
1385 | prefetch(bdev->bd_queue); | |
1386 | prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES); | |
1387 | ||
c8b8e32d | 1388 | return do_blockdev_direct_IO(iocb, inode, bdev, iter, get_block, |
17f8c842 | 1389 | end_io, submit_io, flags); |
65dd2aa9 AK |
1390 | } |
1391 | ||
1da177e4 | 1392 | EXPORT_SYMBOL(__blockdev_direct_IO); |
6e8267f5 AK |
1393 | |
1394 | static __init int dio_init(void) | |
1395 | { | |
1396 | dio_cache = KMEM_CACHE(dio, SLAB_PANIC); | |
1397 | return 0; | |
1398 | } | |
1399 | module_init(dio_init) |