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> |
a27bb332 | 40 | #include <linux/aio.h> |
1da177e4 LT |
41 | |
42 | /* | |
43 | * How many user pages to map in one call to get_user_pages(). This determines | |
cde1ecb3 | 44 | * the size of a structure in the slab cache |
1da177e4 LT |
45 | */ |
46 | #define DIO_PAGES 64 | |
47 | ||
48 | /* | |
49 | * This code generally works in units of "dio_blocks". A dio_block is | |
50 | * somewhere between the hard sector size and the filesystem block size. it | |
51 | * is determined on a per-invocation basis. When talking to the filesystem | |
52 | * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity | |
53 | * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted | |
54 | * to bio_block quantities by shifting left by blkfactor. | |
55 | * | |
56 | * If blkfactor is zero then the user's request was aligned to the filesystem's | |
57 | * blocksize. | |
1da177e4 LT |
58 | */ |
59 | ||
eb28be2b AK |
60 | /* dio_state only used in the submission path */ |
61 | ||
62 | struct dio_submit { | |
1da177e4 | 63 | struct bio *bio; /* bio under assembly */ |
1da177e4 LT |
64 | unsigned blkbits; /* doesn't change */ |
65 | unsigned blkfactor; /* When we're using an alignment which | |
66 | is finer than the filesystem's soft | |
67 | blocksize, this specifies how much | |
68 | finer. blkfactor=2 means 1/4-block | |
69 | alignment. Does not change */ | |
70 | unsigned start_zero_done; /* flag: sub-blocksize zeroing has | |
71 | been performed at the start of a | |
72 | write */ | |
73 | int pages_in_io; /* approximate total IO pages */ | |
74 | size_t size; /* total request size (doesn't change)*/ | |
75 | sector_t block_in_file; /* Current offset into the underlying | |
76 | file in dio_block units. */ | |
77 | unsigned blocks_available; /* At block_in_file. changes */ | |
0dc2bc49 | 78 | int reap_counter; /* rate limit reaping */ |
1da177e4 LT |
79 | sector_t final_block_in_request;/* doesn't change */ |
80 | unsigned first_block_in_page; /* doesn't change, Used only once */ | |
81 | int boundary; /* prev block is at a boundary */ | |
1d8fa7a2 | 82 | get_block_t *get_block; /* block mapping function */ |
facd07b0 | 83 | dio_submit_t *submit_io; /* IO submition function */ |
eb28be2b | 84 | |
facd07b0 | 85 | loff_t logical_offset_in_bio; /* current first logical block in bio */ |
1da177e4 LT |
86 | sector_t final_block_in_bio; /* current final block in bio + 1 */ |
87 | sector_t next_block_for_io; /* next block to be put under IO, | |
88 | in dio_blocks units */ | |
1da177e4 LT |
89 | |
90 | /* | |
91 | * Deferred addition of a page to the dio. These variables are | |
92 | * private to dio_send_cur_page(), submit_page_section() and | |
93 | * dio_bio_add_page(). | |
94 | */ | |
95 | struct page *cur_page; /* The page */ | |
96 | unsigned cur_page_offset; /* Offset into it, in bytes */ | |
97 | unsigned cur_page_len; /* Nr of bytes at cur_page_offset */ | |
98 | sector_t cur_page_block; /* Where it starts */ | |
facd07b0 | 99 | loff_t cur_page_fs_offset; /* Offset in file */ |
1da177e4 LT |
100 | |
101 | /* | |
102 | * Page fetching state. These variables belong to dio_refill_pages(). | |
103 | */ | |
104 | int curr_page; /* changes */ | |
105 | int total_pages; /* doesn't change */ | |
106 | unsigned long curr_user_address;/* changes */ | |
107 | ||
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 */ | |
eb28be2b AK |
114 | }; |
115 | ||
116 | /* dio_state communicated between submission path and end_io */ | |
117 | struct dio { | |
118 | int flags; /* doesn't change */ | |
eb28be2b | 119 | int rw; |
0dc2bc49 | 120 | struct inode *inode; |
eb28be2b AK |
121 | loff_t i_size; /* i_size when submitted */ |
122 | dio_iodone_t *end_io; /* IO completion function */ | |
eb28be2b | 123 | |
18772641 | 124 | void *private; /* copy from map_bh.b_private */ |
eb28be2b AK |
125 | |
126 | /* BIO completion state */ | |
127 | spinlock_t bio_lock; /* protects BIO fields below */ | |
0dc2bc49 AK |
128 | int page_errors; /* errno from get_user_pages() */ |
129 | int is_async; /* is IO async ? */ | |
130 | int io_error; /* IO error in completion path */ | |
eb28be2b AK |
131 | unsigned long refcount; /* direct_io_worker() and bios */ |
132 | struct bio *bio_list; /* singly linked via bi_private */ | |
133 | struct task_struct *waiter; /* waiting task (NULL if none) */ | |
134 | ||
135 | /* AIO related stuff */ | |
136 | struct kiocb *iocb; /* kiocb */ | |
eb28be2b AK |
137 | ssize_t result; /* IO result */ |
138 | ||
23aee091 JM |
139 | /* |
140 | * pages[] (and any fields placed after it) are not zeroed out at | |
141 | * allocation time. Don't add new fields after pages[] unless you | |
142 | * wish that they not be zeroed. | |
143 | */ | |
144 | struct page *pages[DIO_PAGES]; /* page buffer */ | |
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 LT |
161 | { |
162 | int ret; | |
163 | int nr_pages; | |
164 | ||
eb28be2b | 165 | nr_pages = min(sdio->total_pages - sdio->curr_page, DIO_PAGES); |
f5dd33c4 | 166 | ret = get_user_pages_fast( |
eb28be2b | 167 | sdio->curr_user_address, /* Where from? */ |
1da177e4 LT |
168 | nr_pages, /* How many pages? */ |
169 | dio->rw == READ, /* Write to memory? */ | |
f5dd33c4 | 170 | &dio->pages[0]); /* Put results here */ |
1da177e4 | 171 | |
eb28be2b | 172 | if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) { |
557ed1fa | 173 | struct page *page = ZERO_PAGE(0); |
1da177e4 LT |
174 | /* |
175 | * A memory fault, but the filesystem has some outstanding | |
176 | * mapped blocks. We need to use those blocks up to avoid | |
177 | * leaking stale data in the file. | |
178 | */ | |
179 | if (dio->page_errors == 0) | |
180 | dio->page_errors = ret; | |
b5810039 NP |
181 | page_cache_get(page); |
182 | dio->pages[0] = page; | |
eb28be2b AK |
183 | sdio->head = 0; |
184 | sdio->tail = 1; | |
1da177e4 LT |
185 | ret = 0; |
186 | goto out; | |
187 | } | |
188 | ||
189 | if (ret >= 0) { | |
eb28be2b AK |
190 | sdio->curr_user_address += ret * PAGE_SIZE; |
191 | sdio->curr_page += ret; | |
192 | sdio->head = 0; | |
193 | sdio->tail = ret; | |
1da177e4 LT |
194 | ret = 0; |
195 | } | |
196 | out: | |
197 | return ret; | |
198 | } | |
199 | ||
200 | /* | |
201 | * Get another userspace page. Returns an ERR_PTR on error. Pages are | |
202 | * buffered inside the dio so that we can call get_user_pages() against a | |
203 | * decent number of pages, less frequently. To provide nicer use of the | |
204 | * L1 cache. | |
205 | */ | |
ba253fbf AK |
206 | static inline struct page *dio_get_page(struct dio *dio, |
207 | struct dio_submit *sdio) | |
1da177e4 | 208 | { |
eb28be2b | 209 | if (dio_pages_present(sdio) == 0) { |
1da177e4 LT |
210 | int ret; |
211 | ||
eb28be2b | 212 | ret = dio_refill_pages(dio, sdio); |
1da177e4 LT |
213 | if (ret) |
214 | return ERR_PTR(ret); | |
eb28be2b | 215 | BUG_ON(dio_pages_present(sdio) == 0); |
1da177e4 | 216 | } |
eb28be2b | 217 | return dio->pages[sdio->head++]; |
1da177e4 LT |
218 | } |
219 | ||
6d544bb4 ZB |
220 | /** |
221 | * dio_complete() - called when all DIO BIO I/O has been completed | |
222 | * @offset: the byte offset in the file of the completed operation | |
223 | * | |
224 | * This releases locks as dictated by the locking type, lets interested parties | |
225 | * know that a DIO operation has completed, and calculates the resulting return | |
226 | * code for the operation. | |
227 | * | |
228 | * It lets the filesystem know if it registered an interest earlier via | |
229 | * get_block. Pass the private field of the map buffer_head so that | |
230 | * filesystems can use it to hold additional state between get_block calls and | |
231 | * dio_complete. | |
1da177e4 | 232 | */ |
cd1c584f | 233 | static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret, bool is_async) |
1da177e4 | 234 | { |
6d544bb4 ZB |
235 | ssize_t transferred = 0; |
236 | ||
8459d86a ZB |
237 | /* |
238 | * AIO submission can race with bio completion to get here while | |
239 | * expecting to have the last io completed by bio completion. | |
240 | * In that case -EIOCBQUEUED is in fact not an error we want | |
241 | * to preserve through this call. | |
242 | */ | |
243 | if (ret == -EIOCBQUEUED) | |
244 | ret = 0; | |
245 | ||
6d544bb4 ZB |
246 | if (dio->result) { |
247 | transferred = dio->result; | |
248 | ||
249 | /* Check for short read case */ | |
250 | if ((dio->rw == READ) && ((offset + transferred) > dio->i_size)) | |
251 | transferred = dio->i_size - offset; | |
252 | } | |
253 | ||
6d544bb4 ZB |
254 | if (ret == 0) |
255 | ret = dio->page_errors; | |
256 | if (ret == 0) | |
257 | ret = dio->io_error; | |
258 | if (ret == 0) | |
259 | ret = transferred; | |
260 | ||
40e2e973 CH |
261 | if (dio->end_io && dio->result) { |
262 | dio->end_io(dio->iocb, offset, transferred, | |
18772641 | 263 | dio->private, ret, is_async); |
72c5052d | 264 | } else { |
54c807e7 | 265 | inode_dio_done(dio->inode); |
72c5052d CH |
266 | if (is_async) |
267 | aio_complete(dio->iocb, ret, 0); | |
40e2e973 CH |
268 | } |
269 | ||
6d544bb4 | 270 | return ret; |
1da177e4 LT |
271 | } |
272 | ||
1da177e4 LT |
273 | static int dio_bio_complete(struct dio *dio, struct bio *bio); |
274 | /* | |
275 | * Asynchronous IO callback. | |
276 | */ | |
6712ecf8 | 277 | static void dio_bio_end_aio(struct bio *bio, int error) |
1da177e4 LT |
278 | { |
279 | struct dio *dio = bio->bi_private; | |
5eb6c7a2 ZB |
280 | unsigned long remaining; |
281 | unsigned long flags; | |
1da177e4 | 282 | |
1da177e4 LT |
283 | /* cleanup the bio */ |
284 | dio_bio_complete(dio, bio); | |
0273201e | 285 | |
5eb6c7a2 ZB |
286 | spin_lock_irqsave(&dio->bio_lock, flags); |
287 | remaining = --dio->refcount; | |
288 | if (remaining == 1 && dio->waiter) | |
20258b2b | 289 | wake_up_process(dio->waiter); |
5eb6c7a2 | 290 | spin_unlock_irqrestore(&dio->bio_lock, flags); |
20258b2b | 291 | |
8459d86a | 292 | if (remaining == 0) { |
40e2e973 | 293 | dio_complete(dio, dio->iocb->ki_pos, 0, true); |
6e8267f5 | 294 | kmem_cache_free(dio_cache, dio); |
8459d86a | 295 | } |
1da177e4 LT |
296 | } |
297 | ||
298 | /* | |
299 | * The BIO completion handler simply queues the BIO up for the process-context | |
300 | * handler. | |
301 | * | |
302 | * During I/O bi_private points at the dio. After I/O, bi_private is used to | |
303 | * implement a singly-linked list of completed BIOs, at dio->bio_list. | |
304 | */ | |
6712ecf8 | 305 | static void dio_bio_end_io(struct bio *bio, int error) |
1da177e4 LT |
306 | { |
307 | struct dio *dio = bio->bi_private; | |
308 | unsigned long flags; | |
309 | ||
1da177e4 LT |
310 | spin_lock_irqsave(&dio->bio_lock, flags); |
311 | bio->bi_private = dio->bio_list; | |
312 | dio->bio_list = bio; | |
5eb6c7a2 | 313 | if (--dio->refcount == 1 && dio->waiter) |
1da177e4 LT |
314 | wake_up_process(dio->waiter); |
315 | spin_unlock_irqrestore(&dio->bio_lock, flags); | |
1da177e4 LT |
316 | } |
317 | ||
facd07b0 JB |
318 | /** |
319 | * dio_end_io - handle the end io action for the given bio | |
320 | * @bio: The direct io bio thats being completed | |
321 | * @error: Error if there was one | |
322 | * | |
323 | * This is meant to be called by any filesystem that uses their own dio_submit_t | |
324 | * so that the DIO specific endio actions are dealt with after the filesystem | |
325 | * has done it's completion work. | |
326 | */ | |
327 | void dio_end_io(struct bio *bio, int error) | |
328 | { | |
329 | struct dio *dio = bio->bi_private; | |
330 | ||
331 | if (dio->is_async) | |
332 | dio_bio_end_aio(bio, error); | |
333 | else | |
334 | dio_bio_end_io(bio, error); | |
335 | } | |
336 | EXPORT_SYMBOL_GPL(dio_end_io); | |
337 | ||
ba253fbf | 338 | static inline void |
eb28be2b AK |
339 | dio_bio_alloc(struct dio *dio, struct dio_submit *sdio, |
340 | struct block_device *bdev, | |
341 | sector_t first_sector, int nr_vecs) | |
1da177e4 LT |
342 | { |
343 | struct bio *bio; | |
344 | ||
20d9600c DD |
345 | /* |
346 | * bio_alloc() is guaranteed to return a bio when called with | |
347 | * __GFP_WAIT and we request a valid number of vectors. | |
348 | */ | |
1da177e4 | 349 | bio = bio_alloc(GFP_KERNEL, nr_vecs); |
1da177e4 LT |
350 | |
351 | bio->bi_bdev = bdev; | |
352 | bio->bi_sector = first_sector; | |
353 | if (dio->is_async) | |
354 | bio->bi_end_io = dio_bio_end_aio; | |
355 | else | |
356 | bio->bi_end_io = dio_bio_end_io; | |
357 | ||
eb28be2b AK |
358 | sdio->bio = bio; |
359 | sdio->logical_offset_in_bio = sdio->cur_page_fs_offset; | |
1da177e4 LT |
360 | } |
361 | ||
362 | /* | |
363 | * In the AIO read case we speculatively dirty the pages before starting IO. | |
364 | * During IO completion, any of these pages which happen to have been written | |
365 | * back will be redirtied by bio_check_pages_dirty(). | |
0273201e ZB |
366 | * |
367 | * bios hold a dio reference between submit_bio and ->end_io. | |
1da177e4 | 368 | */ |
ba253fbf | 369 | static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio) |
1da177e4 | 370 | { |
eb28be2b | 371 | struct bio *bio = sdio->bio; |
5eb6c7a2 | 372 | unsigned long flags; |
1da177e4 LT |
373 | |
374 | bio->bi_private = dio; | |
5eb6c7a2 ZB |
375 | |
376 | spin_lock_irqsave(&dio->bio_lock, flags); | |
377 | dio->refcount++; | |
378 | spin_unlock_irqrestore(&dio->bio_lock, flags); | |
379 | ||
1da177e4 LT |
380 | if (dio->is_async && dio->rw == READ) |
381 | bio_set_pages_dirty(bio); | |
5eb6c7a2 | 382 | |
eb28be2b AK |
383 | if (sdio->submit_io) |
384 | sdio->submit_io(dio->rw, bio, dio->inode, | |
385 | sdio->logical_offset_in_bio); | |
facd07b0 JB |
386 | else |
387 | submit_bio(dio->rw, bio); | |
1da177e4 | 388 | |
eb28be2b AK |
389 | sdio->bio = NULL; |
390 | sdio->boundary = 0; | |
391 | sdio->logical_offset_in_bio = 0; | |
1da177e4 LT |
392 | } |
393 | ||
394 | /* | |
395 | * Release any resources in case of a failure | |
396 | */ | |
ba253fbf | 397 | static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio) |
1da177e4 | 398 | { |
eb28be2b AK |
399 | while (dio_pages_present(sdio)) |
400 | page_cache_release(dio_get_page(dio, sdio)); | |
1da177e4 LT |
401 | } |
402 | ||
403 | /* | |
0273201e ZB |
404 | * Wait for the next BIO to complete. Remove it and return it. NULL is |
405 | * returned once all BIOs have been completed. This must only be called once | |
406 | * all bios have been issued so that dio->refcount can only decrease. This | |
407 | * requires that that the caller hold a reference on the dio. | |
1da177e4 LT |
408 | */ |
409 | static struct bio *dio_await_one(struct dio *dio) | |
410 | { | |
411 | unsigned long flags; | |
0273201e | 412 | struct bio *bio = NULL; |
1da177e4 LT |
413 | |
414 | spin_lock_irqsave(&dio->bio_lock, flags); | |
5eb6c7a2 ZB |
415 | |
416 | /* | |
417 | * Wait as long as the list is empty and there are bios in flight. bio | |
418 | * completion drops the count, maybe adds to the list, and wakes while | |
419 | * holding the bio_lock so we don't need set_current_state()'s barrier | |
420 | * and can call it after testing our condition. | |
421 | */ | |
422 | while (dio->refcount > 1 && dio->bio_list == NULL) { | |
423 | __set_current_state(TASK_UNINTERRUPTIBLE); | |
424 | dio->waiter = current; | |
425 | spin_unlock_irqrestore(&dio->bio_lock, flags); | |
426 | io_schedule(); | |
427 | /* wake up sets us TASK_RUNNING */ | |
428 | spin_lock_irqsave(&dio->bio_lock, flags); | |
429 | dio->waiter = NULL; | |
1da177e4 | 430 | } |
0273201e ZB |
431 | if (dio->bio_list) { |
432 | bio = dio->bio_list; | |
433 | dio->bio_list = bio->bi_private; | |
434 | } | |
1da177e4 LT |
435 | spin_unlock_irqrestore(&dio->bio_lock, flags); |
436 | return bio; | |
437 | } | |
438 | ||
439 | /* | |
440 | * Process one completed BIO. No locks are held. | |
441 | */ | |
442 | static int dio_bio_complete(struct dio *dio, struct bio *bio) | |
443 | { | |
444 | const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); | |
cb34e057 KO |
445 | struct bio_vec *bvec; |
446 | unsigned i; | |
1da177e4 LT |
447 | |
448 | if (!uptodate) | |
174e27c6 | 449 | dio->io_error = -EIO; |
1da177e4 LT |
450 | |
451 | if (dio->is_async && dio->rw == READ) { | |
452 | bio_check_pages_dirty(bio); /* transfers ownership */ | |
453 | } else { | |
cb34e057 KO |
454 | bio_for_each_segment_all(bvec, bio, i) { |
455 | struct page *page = bvec->bv_page; | |
1da177e4 LT |
456 | |
457 | if (dio->rw == READ && !PageCompound(page)) | |
458 | set_page_dirty_lock(page); | |
459 | page_cache_release(page); | |
460 | } | |
461 | bio_put(bio); | |
462 | } | |
1da177e4 LT |
463 | return uptodate ? 0 : -EIO; |
464 | } | |
465 | ||
466 | /* | |
0273201e ZB |
467 | * Wait on and process all in-flight BIOs. This must only be called once |
468 | * all bios have been issued so that the refcount can only decrease. | |
469 | * This just waits for all bios to make it through dio_bio_complete. IO | |
beb7dd86 | 470 | * errors are propagated through dio->io_error and should be propagated via |
0273201e | 471 | * dio_complete(). |
1da177e4 | 472 | */ |
6d544bb4 | 473 | static void dio_await_completion(struct dio *dio) |
1da177e4 | 474 | { |
0273201e ZB |
475 | struct bio *bio; |
476 | do { | |
477 | bio = dio_await_one(dio); | |
478 | if (bio) | |
479 | dio_bio_complete(dio, bio); | |
480 | } while (bio); | |
1da177e4 LT |
481 | } |
482 | ||
483 | /* | |
484 | * A really large O_DIRECT read or write can generate a lot of BIOs. So | |
485 | * to keep the memory consumption sane we periodically reap any completed BIOs | |
486 | * during the BIO generation phase. | |
487 | * | |
488 | * This also helps to limit the peak amount of pinned userspace memory. | |
489 | */ | |
ba253fbf | 490 | static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio) |
1da177e4 LT |
491 | { |
492 | int ret = 0; | |
493 | ||
eb28be2b | 494 | if (sdio->reap_counter++ >= 64) { |
1da177e4 LT |
495 | while (dio->bio_list) { |
496 | unsigned long flags; | |
497 | struct bio *bio; | |
498 | int ret2; | |
499 | ||
500 | spin_lock_irqsave(&dio->bio_lock, flags); | |
501 | bio = dio->bio_list; | |
502 | dio->bio_list = bio->bi_private; | |
503 | spin_unlock_irqrestore(&dio->bio_lock, flags); | |
504 | ret2 = dio_bio_complete(dio, bio); | |
505 | if (ret == 0) | |
506 | ret = ret2; | |
507 | } | |
eb28be2b | 508 | sdio->reap_counter = 0; |
1da177e4 LT |
509 | } |
510 | return ret; | |
511 | } | |
512 | ||
513 | /* | |
514 | * Call into the fs to map some more disk blocks. We record the current number | |
eb28be2b | 515 | * of available blocks at sdio->blocks_available. These are in units of the |
1da177e4 LT |
516 | * fs blocksize, (1 << inode->i_blkbits). |
517 | * | |
518 | * The fs is allowed to map lots of blocks at once. If it wants to do that, | |
519 | * it uses the passed inode-relative block number as the file offset, as usual. | |
520 | * | |
1d8fa7a2 | 521 | * get_block() is passed the number of i_blkbits-sized blocks which direct_io |
1da177e4 LT |
522 | * has remaining to do. The fs should not map more than this number of blocks. |
523 | * | |
524 | * If the fs has mapped a lot of blocks, it should populate bh->b_size to | |
525 | * indicate how much contiguous disk space has been made available at | |
526 | * bh->b_blocknr. | |
527 | * | |
528 | * If *any* of the mapped blocks are new, then the fs must set buffer_new(). | |
529 | * This isn't very efficient... | |
530 | * | |
531 | * In the case of filesystem holes: the fs may return an arbitrarily-large | |
532 | * hole by returning an appropriate value in b_size and by clearing | |
533 | * buffer_mapped(). However the direct-io code will only process holes one | |
1d8fa7a2 | 534 | * block at a time - it will repeatedly call get_block() as it walks the hole. |
1da177e4 | 535 | */ |
18772641 AK |
536 | static int get_more_blocks(struct dio *dio, struct dio_submit *sdio, |
537 | struct buffer_head *map_bh) | |
1da177e4 LT |
538 | { |
539 | int ret; | |
1da177e4 | 540 | sector_t fs_startblk; /* Into file, in filesystem-sized blocks */ |
ae55e1aa | 541 | sector_t fs_endblk; /* Into file, in filesystem-sized blocks */ |
1da177e4 | 542 | unsigned long fs_count; /* Number of filesystem-sized blocks */ |
1da177e4 | 543 | int create; |
ab73857e | 544 | unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor; |
1da177e4 LT |
545 | |
546 | /* | |
547 | * If there was a memory error and we've overwritten all the | |
548 | * mapped blocks then we can now return that memory error | |
549 | */ | |
550 | ret = dio->page_errors; | |
551 | if (ret == 0) { | |
eb28be2b AK |
552 | BUG_ON(sdio->block_in_file >= sdio->final_block_in_request); |
553 | fs_startblk = sdio->block_in_file >> sdio->blkfactor; | |
ae55e1aa TM |
554 | fs_endblk = (sdio->final_block_in_request - 1) >> |
555 | sdio->blkfactor; | |
556 | fs_count = fs_endblk - fs_startblk + 1; | |
1da177e4 | 557 | |
3c674e74 | 558 | map_bh->b_state = 0; |
ab73857e | 559 | map_bh->b_size = fs_count << i_blkbits; |
3c674e74 | 560 | |
5fe878ae CH |
561 | /* |
562 | * For writes inside i_size on a DIO_SKIP_HOLES filesystem we | |
563 | * forbid block creations: only overwrites are permitted. | |
564 | * We will return early to the caller once we see an | |
565 | * unmapped buffer head returned, and the caller will fall | |
566 | * back to buffered I/O. | |
567 | * | |
568 | * Otherwise the decision is left to the get_blocks method, | |
569 | * which may decide to handle it or also return an unmapped | |
570 | * buffer head. | |
571 | */ | |
b31dc66a | 572 | create = dio->rw & WRITE; |
5fe878ae | 573 | if (dio->flags & DIO_SKIP_HOLES) { |
eb28be2b AK |
574 | if (sdio->block_in_file < (i_size_read(dio->inode) >> |
575 | sdio->blkbits)) | |
1da177e4 | 576 | create = 0; |
1da177e4 | 577 | } |
3c674e74 | 578 | |
eb28be2b | 579 | ret = (*sdio->get_block)(dio->inode, fs_startblk, |
1da177e4 | 580 | map_bh, create); |
18772641 AK |
581 | |
582 | /* Store for completion */ | |
583 | dio->private = map_bh->b_private; | |
1da177e4 LT |
584 | } |
585 | return ret; | |
586 | } | |
587 | ||
588 | /* | |
589 | * There is no bio. Make one now. | |
590 | */ | |
ba253fbf AK |
591 | static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio, |
592 | sector_t start_sector, struct buffer_head *map_bh) | |
1da177e4 LT |
593 | { |
594 | sector_t sector; | |
595 | int ret, nr_pages; | |
596 | ||
eb28be2b | 597 | ret = dio_bio_reap(dio, sdio); |
1da177e4 LT |
598 | if (ret) |
599 | goto out; | |
eb28be2b | 600 | sector = start_sector << (sdio->blkbits - 9); |
18772641 | 601 | nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(map_bh->b_bdev)); |
20d9600c | 602 | nr_pages = min(nr_pages, BIO_MAX_PAGES); |
1da177e4 | 603 | BUG_ON(nr_pages <= 0); |
18772641 | 604 | dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages); |
eb28be2b | 605 | sdio->boundary = 0; |
1da177e4 LT |
606 | out: |
607 | return ret; | |
608 | } | |
609 | ||
610 | /* | |
611 | * Attempt to put the current chunk of 'cur_page' into the current BIO. If | |
612 | * that was successful then update final_block_in_bio and take a ref against | |
613 | * the just-added page. | |
614 | * | |
615 | * Return zero on success. Non-zero means the caller needs to start a new BIO. | |
616 | */ | |
ba253fbf | 617 | static inline int dio_bio_add_page(struct dio_submit *sdio) |
1da177e4 LT |
618 | { |
619 | int ret; | |
620 | ||
eb28be2b AK |
621 | ret = bio_add_page(sdio->bio, sdio->cur_page, |
622 | sdio->cur_page_len, sdio->cur_page_offset); | |
623 | if (ret == sdio->cur_page_len) { | |
1da177e4 LT |
624 | /* |
625 | * Decrement count only, if we are done with this page | |
626 | */ | |
eb28be2b AK |
627 | if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE) |
628 | sdio->pages_in_io--; | |
629 | page_cache_get(sdio->cur_page); | |
630 | sdio->final_block_in_bio = sdio->cur_page_block + | |
631 | (sdio->cur_page_len >> sdio->blkbits); | |
1da177e4 LT |
632 | ret = 0; |
633 | } else { | |
634 | ret = 1; | |
635 | } | |
636 | return ret; | |
637 | } | |
638 | ||
639 | /* | |
640 | * Put cur_page under IO. The section of cur_page which is described by | |
641 | * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page | |
642 | * starts on-disk at cur_page_block. | |
643 | * | |
644 | * We take a ref against the page here (on behalf of its presence in the bio). | |
645 | * | |
646 | * The caller of this function is responsible for removing cur_page from the | |
647 | * dio, and for dropping the refcount which came from that presence. | |
648 | */ | |
ba253fbf AK |
649 | static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio, |
650 | struct buffer_head *map_bh) | |
1da177e4 LT |
651 | { |
652 | int ret = 0; | |
653 | ||
eb28be2b AK |
654 | if (sdio->bio) { |
655 | loff_t cur_offset = sdio->cur_page_fs_offset; | |
656 | loff_t bio_next_offset = sdio->logical_offset_in_bio + | |
657 | sdio->bio->bi_size; | |
c2c6ca41 | 658 | |
1da177e4 | 659 | /* |
c2c6ca41 JB |
660 | * See whether this new request is contiguous with the old. |
661 | * | |
f0940cee NK |
662 | * Btrfs cannot handle having logically non-contiguous requests |
663 | * submitted. For example if you have | |
c2c6ca41 JB |
664 | * |
665 | * Logical: [0-4095][HOLE][8192-12287] | |
f0940cee | 666 | * Physical: [0-4095] [4096-8191] |
c2c6ca41 JB |
667 | * |
668 | * We cannot submit those pages together as one BIO. So if our | |
669 | * current logical offset in the file does not equal what would | |
670 | * be the next logical offset in the bio, submit the bio we | |
671 | * have. | |
1da177e4 | 672 | */ |
eb28be2b | 673 | if (sdio->final_block_in_bio != sdio->cur_page_block || |
c2c6ca41 | 674 | cur_offset != bio_next_offset) |
eb28be2b | 675 | dio_bio_submit(dio, sdio); |
1da177e4 LT |
676 | } |
677 | ||
eb28be2b | 678 | if (sdio->bio == NULL) { |
18772641 | 679 | ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh); |
1da177e4 LT |
680 | if (ret) |
681 | goto out; | |
682 | } | |
683 | ||
eb28be2b AK |
684 | if (dio_bio_add_page(sdio) != 0) { |
685 | dio_bio_submit(dio, sdio); | |
18772641 | 686 | ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh); |
1da177e4 | 687 | if (ret == 0) { |
eb28be2b | 688 | ret = dio_bio_add_page(sdio); |
1da177e4 LT |
689 | BUG_ON(ret != 0); |
690 | } | |
691 | } | |
692 | out: | |
693 | return ret; | |
694 | } | |
695 | ||
696 | /* | |
697 | * An autonomous function to put a chunk of a page under deferred IO. | |
698 | * | |
699 | * The caller doesn't actually know (or care) whether this piece of page is in | |
700 | * a BIO, or is under IO or whatever. We just take care of all possible | |
701 | * situations here. The separation between the logic of do_direct_IO() and | |
702 | * that of submit_page_section() is important for clarity. Please don't break. | |
703 | * | |
704 | * The chunk of page starts on-disk at blocknr. | |
705 | * | |
706 | * We perform deferred IO, by recording the last-submitted page inside our | |
707 | * private part of the dio structure. If possible, we just expand the IO | |
708 | * across that page here. | |
709 | * | |
710 | * If that doesn't work out then we put the old page into the bio and add this | |
711 | * page to the dio instead. | |
712 | */ | |
6fa3eb70 S |
713 | #ifdef CONFIG_MTK_LCA_RAM_OPTIMIZE |
714 | static int | |
715 | #else | |
ba253fbf | 716 | static inline int |
6fa3eb70 | 717 | #endif |
eb28be2b | 718 | submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page, |
18772641 AK |
719 | unsigned offset, unsigned len, sector_t blocknr, |
720 | struct buffer_head *map_bh) | |
1da177e4 | 721 | { |
6fa3eb70 S |
722 | #ifdef FEATURE_STORAGE_PID_LOGGER |
723 | struct page_pid_logger *tmp_logger; | |
724 | extern unsigned char *page_logger; | |
725 | extern spinlock_t g_locker; | |
726 | unsigned long page_index; | |
727 | unsigned long flags; | |
728 | #endif | |
1da177e4 LT |
729 | int ret = 0; |
730 | ||
98c4d57d AM |
731 | if (dio->rw & WRITE) { |
732 | /* | |
733 | * Read accounting is performed in submit_bio() | |
734 | */ | |
735 | task_io_account_write(len); | |
6fa3eb70 S |
736 | #ifdef FEATURE_STORAGE_PID_LOGGER |
737 | ||
738 | if( page_logger && page) { | |
739 | //#if defined(CONFIG_FLATMEM) | |
740 | //page_index = (unsigned long)((page) - mem_map) ; | |
741 | //#else | |
742 | page_index = (unsigned long)(__page_to_pfn(page))- PHYS_PFN_OFFSET; | |
743 | //#endif | |
744 | tmp_logger =((struct page_pid_logger *)page_logger) + page_index; | |
745 | spin_lock_irqsave(&g_locker, flags); | |
746 | if( page_index < num_physpages) { | |
747 | if( tmp_logger->pid1 == 0XFFFF) | |
748 | tmp_logger->pid1 = current->pid; | |
749 | else if( tmp_logger->pid1 != current->pid ) | |
750 | tmp_logger->pid2 = current->pid; | |
751 | } | |
752 | spin_unlock_irqrestore(&g_locker, flags); | |
753 | //printk(KERN_INFO"submit_page_sction pid1:%u pid2:%u pfn:%d \n", tmp_logger->pid1, tmp_logger->pid2, page_index ); | |
754 | } | |
755 | #endif | |
98c4d57d AM |
756 | } |
757 | ||
1da177e4 LT |
758 | /* |
759 | * Can we just grow the current page's presence in the dio? | |
760 | */ | |
eb28be2b AK |
761 | if (sdio->cur_page == page && |
762 | sdio->cur_page_offset + sdio->cur_page_len == offset && | |
763 | sdio->cur_page_block + | |
764 | (sdio->cur_page_len >> sdio->blkbits) == blocknr) { | |
765 | sdio->cur_page_len += len; | |
1da177e4 LT |
766 | goto out; |
767 | } | |
768 | ||
769 | /* | |
770 | * If there's a deferred page already there then send it. | |
771 | */ | |
eb28be2b | 772 | if (sdio->cur_page) { |
18772641 | 773 | ret = dio_send_cur_page(dio, sdio, map_bh); |
eb28be2b AK |
774 | page_cache_release(sdio->cur_page); |
775 | sdio->cur_page = NULL; | |
1da177e4 | 776 | if (ret) |
b1058b98 | 777 | return ret; |
1da177e4 LT |
778 | } |
779 | ||
780 | page_cache_get(page); /* It is in dio */ | |
eb28be2b AK |
781 | sdio->cur_page = page; |
782 | sdio->cur_page_offset = offset; | |
783 | sdio->cur_page_len = len; | |
784 | sdio->cur_page_block = blocknr; | |
785 | sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits; | |
1da177e4 | 786 | out: |
b1058b98 JK |
787 | /* |
788 | * If sdio->boundary then we want to schedule the IO now to | |
789 | * avoid metadata seeks. | |
790 | */ | |
791 | if (sdio->boundary) { | |
792 | ret = dio_send_cur_page(dio, sdio, map_bh); | |
d75e8540 AG |
793 | if (sdio->bio) |
794 | dio_bio_submit(dio, sdio); | |
b1058b98 JK |
795 | page_cache_release(sdio->cur_page); |
796 | sdio->cur_page = NULL; | |
797 | } | |
1da177e4 LT |
798 | return ret; |
799 | } | |
800 | ||
801 | /* | |
802 | * Clean any dirty buffers in the blockdev mapping which alias newly-created | |
803 | * file blocks. Only called for S_ISREG files - blockdevs do not set | |
804 | * buffer_new | |
805 | */ | |
18772641 | 806 | static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh) |
1da177e4 LT |
807 | { |
808 | unsigned i; | |
809 | unsigned nblocks; | |
810 | ||
18772641 | 811 | nblocks = map_bh->b_size >> dio->inode->i_blkbits; |
1da177e4 LT |
812 | |
813 | for (i = 0; i < nblocks; i++) { | |
18772641 AK |
814 | unmap_underlying_metadata(map_bh->b_bdev, |
815 | map_bh->b_blocknr + i); | |
1da177e4 LT |
816 | } |
817 | } | |
818 | ||
819 | /* | |
820 | * If we are not writing the entire block and get_block() allocated | |
821 | * the block for us, we need to fill-in the unused portion of the | |
822 | * block with zeros. This happens only if user-buffer, fileoffset or | |
823 | * io length is not filesystem block-size multiple. | |
824 | * | |
825 | * `end' is zero if we're doing the start of the IO, 1 at the end of the | |
826 | * IO. | |
827 | */ | |
ba253fbf AK |
828 | static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio, |
829 | int end, struct buffer_head *map_bh) | |
1da177e4 LT |
830 | { |
831 | unsigned dio_blocks_per_fs_block; | |
832 | unsigned this_chunk_blocks; /* In dio_blocks */ | |
833 | unsigned this_chunk_bytes; | |
834 | struct page *page; | |
835 | ||
eb28be2b | 836 | sdio->start_zero_done = 1; |
18772641 | 837 | if (!sdio->blkfactor || !buffer_new(map_bh)) |
1da177e4 LT |
838 | return; |
839 | ||
eb28be2b AK |
840 | dio_blocks_per_fs_block = 1 << sdio->blkfactor; |
841 | this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1); | |
1da177e4 LT |
842 | |
843 | if (!this_chunk_blocks) | |
844 | return; | |
845 | ||
846 | /* | |
847 | * We need to zero out part of an fs block. It is either at the | |
848 | * beginning or the end of the fs block. | |
849 | */ | |
850 | if (end) | |
851 | this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks; | |
852 | ||
eb28be2b | 853 | this_chunk_bytes = this_chunk_blocks << sdio->blkbits; |
1da177e4 | 854 | |
557ed1fa | 855 | page = ZERO_PAGE(0); |
eb28be2b | 856 | if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes, |
18772641 | 857 | sdio->next_block_for_io, map_bh)) |
1da177e4 LT |
858 | return; |
859 | ||
eb28be2b | 860 | sdio->next_block_for_io += this_chunk_blocks; |
1da177e4 LT |
861 | } |
862 | ||
863 | /* | |
864 | * Walk the user pages, and the file, mapping blocks to disk and generating | |
865 | * a sequence of (page,offset,len,block) mappings. These mappings are injected | |
866 | * into submit_page_section(), which takes care of the next stage of submission | |
867 | * | |
868 | * Direct IO against a blockdev is different from a file. Because we can | |
869 | * happily perform page-sized but 512-byte aligned IOs. It is important that | |
870 | * blockdev IO be able to have fine alignment and large sizes. | |
871 | * | |
1d8fa7a2 | 872 | * So what we do is to permit the ->get_block function to populate bh.b_size |
1da177e4 LT |
873 | * with the size of IO which is permitted at this offset and this i_blkbits. |
874 | * | |
875 | * For best results, the blockdev should be set up with 512-byte i_blkbits and | |
1d8fa7a2 | 876 | * it should set b_size to PAGE_SIZE or more inside get_block(). This gives |
1da177e4 LT |
877 | * fine alignment but still allows this function to work in PAGE_SIZE units. |
878 | */ | |
18772641 AK |
879 | static int do_direct_IO(struct dio *dio, struct dio_submit *sdio, |
880 | struct buffer_head *map_bh) | |
1da177e4 | 881 | { |
eb28be2b | 882 | const unsigned blkbits = sdio->blkbits; |
1da177e4 LT |
883 | const unsigned blocks_per_page = PAGE_SIZE >> blkbits; |
884 | struct page *page; | |
885 | unsigned block_in_page; | |
1da177e4 LT |
886 | int ret = 0; |
887 | ||
888 | /* The I/O can start at any block offset within the first page */ | |
eb28be2b | 889 | block_in_page = sdio->first_block_in_page; |
1da177e4 | 890 | |
eb28be2b AK |
891 | while (sdio->block_in_file < sdio->final_block_in_request) { |
892 | page = dio_get_page(dio, sdio); | |
1da177e4 LT |
893 | if (IS_ERR(page)) { |
894 | ret = PTR_ERR(page); | |
895 | goto out; | |
896 | } | |
897 | ||
898 | while (block_in_page < blocks_per_page) { | |
899 | unsigned offset_in_page = block_in_page << blkbits; | |
900 | unsigned this_chunk_bytes; /* # of bytes mapped */ | |
901 | unsigned this_chunk_blocks; /* # of blocks */ | |
902 | unsigned u; | |
903 | ||
eb28be2b | 904 | if (sdio->blocks_available == 0) { |
1da177e4 LT |
905 | /* |
906 | * Need to go and map some more disk | |
907 | */ | |
908 | unsigned long blkmask; | |
909 | unsigned long dio_remainder; | |
910 | ||
18772641 | 911 | ret = get_more_blocks(dio, sdio, map_bh); |
1da177e4 LT |
912 | if (ret) { |
913 | page_cache_release(page); | |
914 | goto out; | |
915 | } | |
916 | if (!buffer_mapped(map_bh)) | |
917 | goto do_holes; | |
918 | ||
eb28be2b AK |
919 | sdio->blocks_available = |
920 | map_bh->b_size >> sdio->blkbits; | |
921 | sdio->next_block_for_io = | |
922 | map_bh->b_blocknr << sdio->blkfactor; | |
1da177e4 | 923 | if (buffer_new(map_bh)) |
18772641 | 924 | clean_blockdev_aliases(dio, map_bh); |
1da177e4 | 925 | |
eb28be2b | 926 | if (!sdio->blkfactor) |
1da177e4 LT |
927 | goto do_holes; |
928 | ||
eb28be2b AK |
929 | blkmask = (1 << sdio->blkfactor) - 1; |
930 | dio_remainder = (sdio->block_in_file & blkmask); | |
1da177e4 LT |
931 | |
932 | /* | |
933 | * If we are at the start of IO and that IO | |
934 | * starts partway into a fs-block, | |
935 | * dio_remainder will be non-zero. If the IO | |
936 | * is a read then we can simply advance the IO | |
937 | * cursor to the first block which is to be | |
938 | * read. But if the IO is a write and the | |
939 | * block was newly allocated we cannot do that; | |
940 | * the start of the fs block must be zeroed out | |
941 | * on-disk | |
942 | */ | |
943 | if (!buffer_new(map_bh)) | |
eb28be2b AK |
944 | sdio->next_block_for_io += dio_remainder; |
945 | sdio->blocks_available -= dio_remainder; | |
1da177e4 LT |
946 | } |
947 | do_holes: | |
948 | /* Handle holes */ | |
949 | if (!buffer_mapped(map_bh)) { | |
35dc8161 | 950 | loff_t i_size_aligned; |
1da177e4 LT |
951 | |
952 | /* AKPM: eargh, -ENOTBLK is a hack */ | |
b31dc66a | 953 | if (dio->rw & WRITE) { |
1da177e4 LT |
954 | page_cache_release(page); |
955 | return -ENOTBLK; | |
956 | } | |
957 | ||
35dc8161 JM |
958 | /* |
959 | * Be sure to account for a partial block as the | |
960 | * last block in the file | |
961 | */ | |
962 | i_size_aligned = ALIGN(i_size_read(dio->inode), | |
963 | 1 << blkbits); | |
eb28be2b | 964 | if (sdio->block_in_file >= |
35dc8161 | 965 | i_size_aligned >> blkbits) { |
1da177e4 LT |
966 | /* We hit eof */ |
967 | page_cache_release(page); | |
75894adb | 968 | dio_cleanup(dio, sdio); |
1da177e4 LT |
969 | goto out; |
970 | } | |
eebd2aa3 CL |
971 | zero_user(page, block_in_page << blkbits, |
972 | 1 << blkbits); | |
eb28be2b | 973 | sdio->block_in_file++; |
1da177e4 LT |
974 | block_in_page++; |
975 | goto next_block; | |
976 | } | |
977 | ||
978 | /* | |
979 | * If we're performing IO which has an alignment which | |
980 | * is finer than the underlying fs, go check to see if | |
981 | * we must zero out the start of this block. | |
982 | */ | |
eb28be2b | 983 | if (unlikely(sdio->blkfactor && !sdio->start_zero_done)) |
18772641 | 984 | dio_zero_block(dio, sdio, 0, map_bh); |
1da177e4 LT |
985 | |
986 | /* | |
987 | * Work out, in this_chunk_blocks, how much disk we | |
988 | * can add to this page | |
989 | */ | |
eb28be2b | 990 | this_chunk_blocks = sdio->blocks_available; |
1da177e4 LT |
991 | u = (PAGE_SIZE - offset_in_page) >> blkbits; |
992 | if (this_chunk_blocks > u) | |
993 | this_chunk_blocks = u; | |
eb28be2b | 994 | u = sdio->final_block_in_request - sdio->block_in_file; |
1da177e4 LT |
995 | if (this_chunk_blocks > u) |
996 | this_chunk_blocks = u; | |
997 | this_chunk_bytes = this_chunk_blocks << blkbits; | |
998 | BUG_ON(this_chunk_bytes == 0); | |
999 | ||
092c8d46 JK |
1000 | if (this_chunk_blocks == sdio->blocks_available) |
1001 | sdio->boundary = buffer_boundary(map_bh); | |
eb28be2b AK |
1002 | ret = submit_page_section(dio, sdio, page, |
1003 | offset_in_page, | |
1004 | this_chunk_bytes, | |
18772641 AK |
1005 | sdio->next_block_for_io, |
1006 | map_bh); | |
1da177e4 LT |
1007 | if (ret) { |
1008 | page_cache_release(page); | |
1009 | goto out; | |
1010 | } | |
eb28be2b | 1011 | sdio->next_block_for_io += this_chunk_blocks; |
1da177e4 | 1012 | |
eb28be2b | 1013 | sdio->block_in_file += this_chunk_blocks; |
1da177e4 | 1014 | block_in_page += this_chunk_blocks; |
eb28be2b | 1015 | sdio->blocks_available -= this_chunk_blocks; |
1da177e4 | 1016 | next_block: |
eb28be2b AK |
1017 | BUG_ON(sdio->block_in_file > sdio->final_block_in_request); |
1018 | if (sdio->block_in_file == sdio->final_block_in_request) | |
1da177e4 LT |
1019 | break; |
1020 | } | |
1021 | ||
1022 | /* Drop the ref which was taken in get_user_pages() */ | |
1023 | page_cache_release(page); | |
1024 | block_in_page = 0; | |
1025 | } | |
1026 | out: | |
1027 | return ret; | |
1028 | } | |
1029 | ||
847cc637 | 1030 | static inline int drop_refcount(struct dio *dio) |
1da177e4 | 1031 | { |
847cc637 | 1032 | int ret2; |
5eb6c7a2 | 1033 | unsigned long flags; |
1da177e4 | 1034 | |
8459d86a ZB |
1035 | /* |
1036 | * Sync will always be dropping the final ref and completing the | |
5eb6c7a2 ZB |
1037 | * operation. AIO can if it was a broken operation described above or |
1038 | * in fact if all the bios race to complete before we get here. In | |
1039 | * that case dio_complete() translates the EIOCBQUEUED into the proper | |
1040 | * return code that the caller will hand to aio_complete(). | |
1041 | * | |
1042 | * This is managed by the bio_lock instead of being an atomic_t so that | |
1043 | * completion paths can drop their ref and use the remaining count to | |
1044 | * decide to wake the submission path atomically. | |
8459d86a | 1045 | */ |
5eb6c7a2 ZB |
1046 | spin_lock_irqsave(&dio->bio_lock, flags); |
1047 | ret2 = --dio->refcount; | |
1048 | spin_unlock_irqrestore(&dio->bio_lock, flags); | |
847cc637 | 1049 | return ret2; |
1da177e4 LT |
1050 | } |
1051 | ||
eafdc7d1 CH |
1052 | /* |
1053 | * This is a library function for use by filesystem drivers. | |
1054 | * | |
1055 | * The locking rules are governed by the flags parameter: | |
1056 | * - if the flags value contains DIO_LOCKING we use a fancy locking | |
1057 | * scheme for dumb filesystems. | |
1058 | * For writes this function is called under i_mutex and returns with | |
1059 | * i_mutex held, for reads, i_mutex is not held on entry, but it is | |
1060 | * taken and dropped again before returning. | |
eafdc7d1 CH |
1061 | * - if the flags value does NOT contain DIO_LOCKING we don't use any |
1062 | * internal locking but rather rely on the filesystem to synchronize | |
1063 | * direct I/O reads/writes versus each other and truncate. | |
df2d6f26 CH |
1064 | * |
1065 | * To help with locking against truncate we incremented the i_dio_count | |
1066 | * counter before starting direct I/O, and decrement it once we are done. | |
1067 | * Truncate can wait for it to reach zero to provide exclusion. It is | |
1068 | * expected that filesystem provide exclusion between new direct I/O | |
1069 | * and truncates. For DIO_LOCKING filesystems this is done by i_mutex, | |
1070 | * but other filesystems need to take care of this on their own. | |
ba253fbf AK |
1071 | * |
1072 | * NOTE: if you pass "sdio" to anything by pointer make sure that function | |
1073 | * is always inlined. Otherwise gcc is unable to split the structure into | |
1074 | * individual fields and will generate much worse code. This is important | |
1075 | * for the whole file. | |
eafdc7d1 | 1076 | */ |
65dd2aa9 AK |
1077 | static inline ssize_t |
1078 | do_blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode, | |
1da177e4 | 1079 | struct block_device *bdev, const struct iovec *iov, loff_t offset, |
1d8fa7a2 | 1080 | unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io, |
facd07b0 | 1081 | dio_submit_t submit_io, int flags) |
1da177e4 LT |
1082 | { |
1083 | int seg; | |
1084 | size_t size; | |
1085 | unsigned long addr; | |
ab73857e LT |
1086 | unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits); |
1087 | unsigned blkbits = i_blkbits; | |
1da177e4 LT |
1088 | unsigned blocksize_mask = (1 << blkbits) - 1; |
1089 | ssize_t retval = -EINVAL; | |
1090 | loff_t end = offset; | |
1091 | struct dio *dio; | |
eb28be2b | 1092 | struct dio_submit sdio = { 0, }; |
847cc637 AK |
1093 | unsigned long user_addr; |
1094 | size_t bytes; | |
1095 | struct buffer_head map_bh = { 0, }; | |
647d1e4c | 1096 | struct blk_plug plug; |
1da177e4 LT |
1097 | |
1098 | if (rw & WRITE) | |
721a9602 | 1099 | rw = WRITE_ODIRECT; |
1da177e4 | 1100 | |
65dd2aa9 AK |
1101 | /* |
1102 | * Avoid references to bdev if not absolutely needed to give | |
1103 | * the early prefetch in the caller enough time. | |
1104 | */ | |
1da177e4 LT |
1105 | |
1106 | if (offset & blocksize_mask) { | |
1107 | if (bdev) | |
65dd2aa9 | 1108 | blkbits = blksize_bits(bdev_logical_block_size(bdev)); |
1da177e4 LT |
1109 | blocksize_mask = (1 << blkbits) - 1; |
1110 | if (offset & blocksize_mask) | |
1111 | goto out; | |
1112 | } | |
1113 | ||
1114 | /* Check the memory alignment. Blocks cannot straddle pages */ | |
1115 | for (seg = 0; seg < nr_segs; seg++) { | |
1116 | addr = (unsigned long)iov[seg].iov_base; | |
1117 | size = iov[seg].iov_len; | |
1118 | end += size; | |
65dd2aa9 AK |
1119 | if (unlikely((addr & blocksize_mask) || |
1120 | (size & blocksize_mask))) { | |
1da177e4 | 1121 | if (bdev) |
65dd2aa9 AK |
1122 | blkbits = blksize_bits( |
1123 | bdev_logical_block_size(bdev)); | |
1da177e4 | 1124 | blocksize_mask = (1 << blkbits) - 1; |
65dd2aa9 | 1125 | if ((addr & blocksize_mask) || (size & blocksize_mask)) |
1da177e4 LT |
1126 | goto out; |
1127 | } | |
1128 | } | |
1129 | ||
f9b5570d CH |
1130 | /* watch out for a 0 len io from a tricksy fs */ |
1131 | if (rw == READ && end == offset) | |
1132 | return 0; | |
1133 | ||
6e8267f5 | 1134 | dio = kmem_cache_alloc(dio_cache, GFP_KERNEL); |
1da177e4 LT |
1135 | retval = -ENOMEM; |
1136 | if (!dio) | |
1137 | goto out; | |
23aee091 JM |
1138 | /* |
1139 | * Believe it or not, zeroing out the page array caused a .5% | |
1140 | * performance regression in a database benchmark. So, we take | |
1141 | * care to only zero out what's needed. | |
1142 | */ | |
1143 | memset(dio, 0, offsetof(struct dio, pages)); | |
1da177e4 | 1144 | |
5fe878ae CH |
1145 | dio->flags = flags; |
1146 | if (dio->flags & DIO_LOCKING) { | |
f9b5570d | 1147 | if (rw == READ) { |
5fe878ae CH |
1148 | struct address_space *mapping = |
1149 | iocb->ki_filp->f_mapping; | |
1da177e4 | 1150 | |
5fe878ae CH |
1151 | /* will be released by direct_io_worker */ |
1152 | mutex_lock(&inode->i_mutex); | |
1da177e4 LT |
1153 | |
1154 | retval = filemap_write_and_wait_range(mapping, offset, | |
1155 | end - 1); | |
1156 | if (retval) { | |
5fe878ae | 1157 | mutex_unlock(&inode->i_mutex); |
6e8267f5 | 1158 | kmem_cache_free(dio_cache, dio); |
1da177e4 LT |
1159 | goto out; |
1160 | } | |
1da177e4 | 1161 | } |
1da177e4 LT |
1162 | } |
1163 | ||
df2d6f26 CH |
1164 | /* |
1165 | * Will be decremented at I/O completion time. | |
1166 | */ | |
1167 | atomic_inc(&inode->i_dio_count); | |
1168 | ||
1da177e4 LT |
1169 | /* |
1170 | * For file extending writes updating i_size before data | |
1171 | * writeouts complete can expose uninitialized blocks. So | |
1172 | * even for AIO, we need to wait for i/o to complete before | |
1173 | * returning in this case. | |
1174 | */ | |
b31dc66a | 1175 | dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) && |
1da177e4 LT |
1176 | (end > i_size_read(inode))); |
1177 | ||
847cc637 AK |
1178 | retval = 0; |
1179 | ||
1180 | dio->inode = inode; | |
1181 | dio->rw = rw; | |
1182 | sdio.blkbits = blkbits; | |
ab73857e | 1183 | sdio.blkfactor = i_blkbits - blkbits; |
847cc637 AK |
1184 | sdio.block_in_file = offset >> blkbits; |
1185 | ||
1186 | sdio.get_block = get_block; | |
1187 | dio->end_io = end_io; | |
1188 | sdio.submit_io = submit_io; | |
1189 | sdio.final_block_in_bio = -1; | |
1190 | sdio.next_block_for_io = -1; | |
1191 | ||
1192 | dio->iocb = iocb; | |
1193 | dio->i_size = i_size_read(inode); | |
1194 | ||
1195 | spin_lock_init(&dio->bio_lock); | |
1196 | dio->refcount = 1; | |
1197 | ||
1198 | /* | |
1199 | * In case of non-aligned buffers, we may need 2 more | |
1200 | * pages since we need to zero out first and last block. | |
1201 | */ | |
1202 | if (unlikely(sdio.blkfactor)) | |
1203 | sdio.pages_in_io = 2; | |
1204 | ||
1205 | for (seg = 0; seg < nr_segs; seg++) { | |
1206 | user_addr = (unsigned long)iov[seg].iov_base; | |
1207 | sdio.pages_in_io += | |
1208 | ((user_addr + iov[seg].iov_len + PAGE_SIZE-1) / | |
1209 | PAGE_SIZE - user_addr / PAGE_SIZE); | |
1210 | } | |
1211 | ||
647d1e4c FW |
1212 | blk_start_plug(&plug); |
1213 | ||
847cc637 AK |
1214 | for (seg = 0; seg < nr_segs; seg++) { |
1215 | user_addr = (unsigned long)iov[seg].iov_base; | |
1216 | sdio.size += bytes = iov[seg].iov_len; | |
1217 | ||
1218 | /* Index into the first page of the first block */ | |
1219 | sdio.first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits; | |
1220 | sdio.final_block_in_request = sdio.block_in_file + | |
1221 | (bytes >> blkbits); | |
1222 | /* Page fetching state */ | |
1223 | sdio.head = 0; | |
1224 | sdio.tail = 0; | |
1225 | sdio.curr_page = 0; | |
1226 | ||
1227 | sdio.total_pages = 0; | |
1228 | if (user_addr & (PAGE_SIZE-1)) { | |
1229 | sdio.total_pages++; | |
1230 | bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1)); | |
1231 | } | |
1232 | sdio.total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE; | |
1233 | sdio.curr_user_address = user_addr; | |
1234 | ||
1235 | retval = do_direct_IO(dio, &sdio, &map_bh); | |
1236 | ||
1237 | dio->result += iov[seg].iov_len - | |
1238 | ((sdio.final_block_in_request - sdio.block_in_file) << | |
1239 | blkbits); | |
1240 | ||
1241 | if (retval) { | |
1242 | dio_cleanup(dio, &sdio); | |
1243 | break; | |
1244 | } | |
1245 | } /* end iovec loop */ | |
1246 | ||
1247 | if (retval == -ENOTBLK) { | |
1248 | /* | |
1249 | * The remaining part of the request will be | |
1250 | * be handled by buffered I/O when we return | |
1251 | */ | |
1252 | retval = 0; | |
1253 | } | |
1254 | /* | |
1255 | * There may be some unwritten disk at the end of a part-written | |
1256 | * fs-block-sized block. Go zero that now. | |
1257 | */ | |
1258 | dio_zero_block(dio, &sdio, 1, &map_bh); | |
1259 | ||
1260 | if (sdio.cur_page) { | |
1261 | ssize_t ret2; | |
1262 | ||
1263 | ret2 = dio_send_cur_page(dio, &sdio, &map_bh); | |
1264 | if (retval == 0) | |
1265 | retval = ret2; | |
1266 | page_cache_release(sdio.cur_page); | |
1267 | sdio.cur_page = NULL; | |
1268 | } | |
1269 | if (sdio.bio) | |
1270 | dio_bio_submit(dio, &sdio); | |
1271 | ||
647d1e4c FW |
1272 | blk_finish_plug(&plug); |
1273 | ||
847cc637 AK |
1274 | /* |
1275 | * It is possible that, we return short IO due to end of file. | |
1276 | * In that case, we need to release all the pages we got hold on. | |
1277 | */ | |
1278 | dio_cleanup(dio, &sdio); | |
1279 | ||
1280 | /* | |
1281 | * All block lookups have been performed. For READ requests | |
1282 | * we can let i_mutex go now that its achieved its purpose | |
1283 | * of protecting us from looking up uninitialized blocks. | |
1284 | */ | |
1285 | if (rw == READ && (dio->flags & DIO_LOCKING)) | |
1286 | mutex_unlock(&dio->inode->i_mutex); | |
1287 | ||
1288 | /* | |
1289 | * The only time we want to leave bios in flight is when a successful | |
1290 | * partial aio read or full aio write have been setup. In that case | |
1291 | * bio completion will call aio_complete. The only time it's safe to | |
1292 | * call aio_complete is when we return -EIOCBQUEUED, so we key on that. | |
1293 | * This had *better* be the only place that raises -EIOCBQUEUED. | |
1294 | */ | |
1295 | BUG_ON(retval == -EIOCBQUEUED); | |
1296 | if (dio->is_async && retval == 0 && dio->result && | |
d187663e | 1297 | ((rw == READ) || (dio->result == sdio.size))) |
847cc637 AK |
1298 | retval = -EIOCBQUEUED; |
1299 | ||
1300 | if (retval != -EIOCBQUEUED) | |
1301 | dio_await_completion(dio); | |
1302 | ||
1303 | if (drop_refcount(dio) == 0) { | |
1304 | retval = dio_complete(dio, offset, retval, false); | |
1305 | kmem_cache_free(dio_cache, dio); | |
1306 | } else | |
1307 | BUG_ON(retval != -EIOCBQUEUED); | |
1da177e4 | 1308 | |
7bb46a67 | 1309 | out: |
1310 | return retval; | |
1311 | } | |
65dd2aa9 AK |
1312 | |
1313 | ssize_t | |
1314 | __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode, | |
1315 | struct block_device *bdev, const struct iovec *iov, loff_t offset, | |
1316 | unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io, | |
1317 | dio_submit_t submit_io, int flags) | |
1318 | { | |
6fa3eb70 | 1319 | //printk("%s:%d:DENIS \n", __FUNCTION__, __LINE__); |
65dd2aa9 AK |
1320 | /* |
1321 | * The block device state is needed in the end to finally | |
1322 | * submit everything. Since it's likely to be cache cold | |
1323 | * prefetch it here as first thing to hide some of the | |
1324 | * latency. | |
1325 | * | |
1326 | * Attempt to prefetch the pieces we likely need later. | |
1327 | */ | |
1328 | prefetch(&bdev->bd_disk->part_tbl); | |
1329 | prefetch(bdev->bd_queue); | |
1330 | prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES); | |
1331 | ||
1332 | return do_blockdev_direct_IO(rw, iocb, inode, bdev, iov, offset, | |
1333 | nr_segs, get_block, end_io, | |
1334 | submit_io, flags); | |
1335 | } | |
1336 | ||
1da177e4 | 1337 | EXPORT_SYMBOL(__blockdev_direct_IO); |
6e8267f5 AK |
1338 | |
1339 | static __init int dio_init(void) | |
1340 | { | |
1341 | dio_cache = KMEM_CACHE(dio, SLAB_PANIC); | |
1342 | return 0; | |
1343 | } | |
1344 | module_init(dio_init) |