Root/
1 | /* |
2 | * fs/direct-io.c |
3 | * |
4 | * Copyright (C) 2002, Linus Torvalds. |
5 | * |
6 | * O_DIRECT |
7 | * |
8 | * 04Jul2002 Andrew Morton |
9 | * Initial version |
10 | * 11Sep2002 janetinc@us.ibm.com |
11 | * added readv/writev support. |
12 | * 29Oct2002 Andrew Morton |
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> |
30 | #include <linux/task_io_accounting_ops.h> |
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> |
38 | #include <linux/atomic.h> |
39 | #include <linux/prefetch.h> |
40 | #include <linux/aio.h> |
41 | |
42 | /* |
43 | * How many user pages to map in one call to get_user_pages(). This determines |
44 | * the size of a structure in the slab cache |
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. |
58 | */ |
59 | |
60 | /* dio_state only used in the submission path */ |
61 | |
62 | struct dio_submit { |
63 | struct bio *bio; /* bio under assembly */ |
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 */ |
78 | int reap_counter; /* rate limit reaping */ |
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 */ |
82 | get_block_t *get_block; /* block mapping function */ |
83 | dio_submit_t *submit_io; /* IO submition function */ |
84 | |
85 | loff_t logical_offset_in_bio; /* current first logical block in bio */ |
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 */ |
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 */ |
99 | loff_t cur_page_fs_offset; /* Offset in file */ |
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 | */ |
112 | unsigned head; /* next page to process */ |
113 | unsigned tail; /* last valid page + 1 */ |
114 | }; |
115 | |
116 | /* dio_state communicated between submission path and end_io */ |
117 | struct dio { |
118 | int flags; /* doesn't change */ |
119 | int rw; |
120 | struct inode *inode; |
121 | loff_t i_size; /* i_size when submitted */ |
122 | dio_iodone_t *end_io; /* IO completion function */ |
123 | |
124 | void *private; /* copy from map_bh.b_private */ |
125 | |
126 | /* BIO completion state */ |
127 | spinlock_t bio_lock; /* protects BIO fields below */ |
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 */ |
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 */ |
137 | ssize_t result; /* IO result */ |
138 | |
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 */ |
145 | } ____cacheline_aligned_in_smp; |
146 | |
147 | static struct kmem_cache *dio_cache __read_mostly; |
148 | |
149 | /* |
150 | * How many pages are in the queue? |
151 | */ |
152 | static inline unsigned dio_pages_present(struct dio_submit *sdio) |
153 | { |
154 | return sdio->tail - sdio->head; |
155 | } |
156 | |
157 | /* |
158 | * Go grab and pin some userspace pages. Typically we'll get 64 at a time. |
159 | */ |
160 | static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio) |
161 | { |
162 | int ret; |
163 | int nr_pages; |
164 | |
165 | nr_pages = min(sdio->total_pages - sdio->curr_page, DIO_PAGES); |
166 | ret = get_user_pages_fast( |
167 | sdio->curr_user_address, /* Where from? */ |
168 | nr_pages, /* How many pages? */ |
169 | dio->rw == READ, /* Write to memory? */ |
170 | &dio->pages[0]); /* Put results here */ |
171 | |
172 | if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) { |
173 | struct page *page = ZERO_PAGE(0); |
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; |
181 | page_cache_get(page); |
182 | dio->pages[0] = page; |
183 | sdio->head = 0; |
184 | sdio->tail = 1; |
185 | ret = 0; |
186 | goto out; |
187 | } |
188 | |
189 | if (ret >= 0) { |
190 | sdio->curr_user_address += ret * PAGE_SIZE; |
191 | sdio->curr_page += ret; |
192 | sdio->head = 0; |
193 | sdio->tail = ret; |
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 | */ |
206 | static inline struct page *dio_get_page(struct dio *dio, |
207 | struct dio_submit *sdio) |
208 | { |
209 | if (dio_pages_present(sdio) == 0) { |
210 | int ret; |
211 | |
212 | ret = dio_refill_pages(dio, sdio); |
213 | if (ret) |
214 | return ERR_PTR(ret); |
215 | BUG_ON(dio_pages_present(sdio) == 0); |
216 | } |
217 | return dio->pages[sdio->head++]; |
218 | } |
219 | |
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. |
232 | */ |
233 | static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret, bool is_async) |
234 | { |
235 | ssize_t transferred = 0; |
236 | |
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 | |
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 | |
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 | |
261 | if (dio->end_io && dio->result) { |
262 | dio->end_io(dio->iocb, offset, transferred, |
263 | dio->private, ret, is_async); |
264 | } else { |
265 | inode_dio_done(dio->inode); |
266 | if (is_async) |
267 | aio_complete(dio->iocb, ret, 0); |
268 | } |
269 | |
270 | return ret; |
271 | } |
272 | |
273 | static int dio_bio_complete(struct dio *dio, struct bio *bio); |
274 | /* |
275 | * Asynchronous IO callback. |
276 | */ |
277 | static void dio_bio_end_aio(struct bio *bio, int error) |
278 | { |
279 | struct dio *dio = bio->bi_private; |
280 | unsigned long remaining; |
281 | unsigned long flags; |
282 | |
283 | /* cleanup the bio */ |
284 | dio_bio_complete(dio, bio); |
285 | |
286 | spin_lock_irqsave(&dio->bio_lock, flags); |
287 | remaining = --dio->refcount; |
288 | if (remaining == 1 && dio->waiter) |
289 | wake_up_process(dio->waiter); |
290 | spin_unlock_irqrestore(&dio->bio_lock, flags); |
291 | |
292 | if (remaining == 0) { |
293 | dio_complete(dio, dio->iocb->ki_pos, 0, true); |
294 | kmem_cache_free(dio_cache, dio); |
295 | } |
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 | */ |
305 | static void dio_bio_end_io(struct bio *bio, int error) |
306 | { |
307 | struct dio *dio = bio->bi_private; |
308 | unsigned long flags; |
309 | |
310 | spin_lock_irqsave(&dio->bio_lock, flags); |
311 | bio->bi_private = dio->bio_list; |
312 | dio->bio_list = bio; |
313 | if (--dio->refcount == 1 && dio->waiter) |
314 | wake_up_process(dio->waiter); |
315 | spin_unlock_irqrestore(&dio->bio_lock, flags); |
316 | } |
317 | |
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 | |
338 | static inline void |
339 | dio_bio_alloc(struct dio *dio, struct dio_submit *sdio, |
340 | struct block_device *bdev, |
341 | sector_t first_sector, int nr_vecs) |
342 | { |
343 | struct bio *bio; |
344 | |
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 | */ |
349 | bio = bio_alloc(GFP_KERNEL, nr_vecs); |
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 | |
358 | sdio->bio = bio; |
359 | sdio->logical_offset_in_bio = sdio->cur_page_fs_offset; |
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(). |
366 | * |
367 | * bios hold a dio reference between submit_bio and ->end_io. |
368 | */ |
369 | static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio) |
370 | { |
371 | struct bio *bio = sdio->bio; |
372 | unsigned long flags; |
373 | |
374 | bio->bi_private = dio; |
375 | |
376 | spin_lock_irqsave(&dio->bio_lock, flags); |
377 | dio->refcount++; |
378 | spin_unlock_irqrestore(&dio->bio_lock, flags); |
379 | |
380 | if (dio->is_async && dio->rw == READ) |
381 | bio_set_pages_dirty(bio); |
382 | |
383 | if (sdio->submit_io) |
384 | sdio->submit_io(dio->rw, bio, dio->inode, |
385 | sdio->logical_offset_in_bio); |
386 | else |
387 | submit_bio(dio->rw, bio); |
388 | |
389 | sdio->bio = NULL; |
390 | sdio->boundary = 0; |
391 | sdio->logical_offset_in_bio = 0; |
392 | } |
393 | |
394 | /* |
395 | * Release any resources in case of a failure |
396 | */ |
397 | static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio) |
398 | { |
399 | while (dio_pages_present(sdio)) |
400 | page_cache_release(dio_get_page(dio, sdio)); |
401 | } |
402 | |
403 | /* |
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. |
408 | */ |
409 | static struct bio *dio_await_one(struct dio *dio) |
410 | { |
411 | unsigned long flags; |
412 | struct bio *bio = NULL; |
413 | |
414 | spin_lock_irqsave(&dio->bio_lock, flags); |
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; |
430 | } |
431 | if (dio->bio_list) { |
432 | bio = dio->bio_list; |
433 | dio->bio_list = bio->bi_private; |
434 | } |
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); |
445 | struct bio_vec *bvec; |
446 | unsigned i; |
447 | |
448 | if (!uptodate) |
449 | dio->io_error = -EIO; |
450 | |
451 | if (dio->is_async && dio->rw == READ) { |
452 | bio_check_pages_dirty(bio); /* transfers ownership */ |
453 | } else { |
454 | bio_for_each_segment_all(bvec, bio, i) { |
455 | struct page *page = bvec->bv_page; |
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 | } |
463 | return uptodate ? 0 : -EIO; |
464 | } |
465 | |
466 | /* |
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 |
470 | * errors are propagated through dio->io_error and should be propagated via |
471 | * dio_complete(). |
472 | */ |
473 | static void dio_await_completion(struct dio *dio) |
474 | { |
475 | struct bio *bio; |
476 | do { |
477 | bio = dio_await_one(dio); |
478 | if (bio) |
479 | dio_bio_complete(dio, bio); |
480 | } while (bio); |
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 | */ |
490 | static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio) |
491 | { |
492 | int ret = 0; |
493 | |
494 | if (sdio->reap_counter++ >= 64) { |
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 | } |
508 | sdio->reap_counter = 0; |
509 | } |
510 | return ret; |
511 | } |
512 | |
513 | /* |
514 | * Call into the fs to map some more disk blocks. We record the current number |
515 | * of available blocks at sdio->blocks_available. These are in units of the |
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 | * |
521 | * get_block() is passed the number of i_blkbits-sized blocks which direct_io |
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 |
534 | * block at a time - it will repeatedly call get_block() as it walks the hole. |
535 | */ |
536 | static int get_more_blocks(struct dio *dio, struct dio_submit *sdio, |
537 | struct buffer_head *map_bh) |
538 | { |
539 | int ret; |
540 | sector_t fs_startblk; /* Into file, in filesystem-sized blocks */ |
541 | sector_t fs_endblk; /* Into file, in filesystem-sized blocks */ |
542 | unsigned long fs_count; /* Number of filesystem-sized blocks */ |
543 | int create; |
544 | unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor; |
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) { |
552 | BUG_ON(sdio->block_in_file >= sdio->final_block_in_request); |
553 | fs_startblk = sdio->block_in_file >> sdio->blkfactor; |
554 | fs_endblk = (sdio->final_block_in_request - 1) >> |
555 | sdio->blkfactor; |
556 | fs_count = fs_endblk - fs_startblk + 1; |
557 | |
558 | map_bh->b_state = 0; |
559 | map_bh->b_size = fs_count << i_blkbits; |
560 | |
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 | */ |
572 | create = dio->rw & WRITE; |
573 | if (dio->flags & DIO_SKIP_HOLES) { |
574 | if (sdio->block_in_file < (i_size_read(dio->inode) >> |
575 | sdio->blkbits)) |
576 | create = 0; |
577 | } |
578 | |
579 | ret = (*sdio->get_block)(dio->inode, fs_startblk, |
580 | map_bh, create); |
581 | |
582 | /* Store for completion */ |
583 | dio->private = map_bh->b_private; |
584 | } |
585 | return ret; |
586 | } |
587 | |
588 | /* |
589 | * There is no bio. Make one now. |
590 | */ |
591 | static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio, |
592 | sector_t start_sector, struct buffer_head *map_bh) |
593 | { |
594 | sector_t sector; |
595 | int ret, nr_pages; |
596 | |
597 | ret = dio_bio_reap(dio, sdio); |
598 | if (ret) |
599 | goto out; |
600 | sector = start_sector << (sdio->blkbits - 9); |
601 | nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(map_bh->b_bdev)); |
602 | nr_pages = min(nr_pages, BIO_MAX_PAGES); |
603 | BUG_ON(nr_pages <= 0); |
604 | dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages); |
605 | sdio->boundary = 0; |
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 | */ |
617 | static inline int dio_bio_add_page(struct dio_submit *sdio) |
618 | { |
619 | int ret; |
620 | |
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) { |
624 | /* |
625 | * Decrement count only, if we are done with this page |
626 | */ |
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); |
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 | */ |
649 | static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio, |
650 | struct buffer_head *map_bh) |
651 | { |
652 | int ret = 0; |
653 | |
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; |
658 | |
659 | /* |
660 | * See whether this new request is contiguous with the old. |
661 | * |
662 | * Btrfs cannot handle having logically non-contiguous requests |
663 | * submitted. For example if you have |
664 | * |
665 | * Logical: [0-4095][HOLE][8192-12287] |
666 | * Physical: [0-4095] [4096-8191] |
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. |
672 | */ |
673 | if (sdio->final_block_in_bio != sdio->cur_page_block || |
674 | cur_offset != bio_next_offset) |
675 | dio_bio_submit(dio, sdio); |
676 | } |
677 | |
678 | if (sdio->bio == NULL) { |
679 | ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh); |
680 | if (ret) |
681 | goto out; |
682 | } |
683 | |
684 | if (dio_bio_add_page(sdio) != 0) { |
685 | dio_bio_submit(dio, sdio); |
686 | ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh); |
687 | if (ret == 0) { |
688 | ret = dio_bio_add_page(sdio); |
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 | */ |
713 | static inline int |
714 | submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page, |
715 | unsigned offset, unsigned len, sector_t blocknr, |
716 | struct buffer_head *map_bh) |
717 | { |
718 | int ret = 0; |
719 | |
720 | if (dio->rw & WRITE) { |
721 | /* |
722 | * Read accounting is performed in submit_bio() |
723 | */ |
724 | task_io_account_write(len); |
725 | } |
726 | |
727 | /* |
728 | * Can we just grow the current page's presence in the dio? |
729 | */ |
730 | if (sdio->cur_page == page && |
731 | sdio->cur_page_offset + sdio->cur_page_len == offset && |
732 | sdio->cur_page_block + |
733 | (sdio->cur_page_len >> sdio->blkbits) == blocknr) { |
734 | sdio->cur_page_len += len; |
735 | goto out; |
736 | } |
737 | |
738 | /* |
739 | * If there's a deferred page already there then send it. |
740 | */ |
741 | if (sdio->cur_page) { |
742 | ret = dio_send_cur_page(dio, sdio, map_bh); |
743 | page_cache_release(sdio->cur_page); |
744 | sdio->cur_page = NULL; |
745 | if (ret) |
746 | return ret; |
747 | } |
748 | |
749 | page_cache_get(page); /* It is in dio */ |
750 | sdio->cur_page = page; |
751 | sdio->cur_page_offset = offset; |
752 | sdio->cur_page_len = len; |
753 | sdio->cur_page_block = blocknr; |
754 | sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits; |
755 | out: |
756 | /* |
757 | * If sdio->boundary then we want to schedule the IO now to |
758 | * avoid metadata seeks. |
759 | */ |
760 | if (sdio->boundary) { |
761 | ret = dio_send_cur_page(dio, sdio, map_bh); |
762 | dio_bio_submit(dio, sdio); |
763 | page_cache_release(sdio->cur_page); |
764 | sdio->cur_page = NULL; |
765 | } |
766 | return ret; |
767 | } |
768 | |
769 | /* |
770 | * Clean any dirty buffers in the blockdev mapping which alias newly-created |
771 | * file blocks. Only called for S_ISREG files - blockdevs do not set |
772 | * buffer_new |
773 | */ |
774 | static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh) |
775 | { |
776 | unsigned i; |
777 | unsigned nblocks; |
778 | |
779 | nblocks = map_bh->b_size >> dio->inode->i_blkbits; |
780 | |
781 | for (i = 0; i < nblocks; i++) { |
782 | unmap_underlying_metadata(map_bh->b_bdev, |
783 | map_bh->b_blocknr + i); |
784 | } |
785 | } |
786 | |
787 | /* |
788 | * If we are not writing the entire block and get_block() allocated |
789 | * the block for us, we need to fill-in the unused portion of the |
790 | * block with zeros. This happens only if user-buffer, fileoffset or |
791 | * io length is not filesystem block-size multiple. |
792 | * |
793 | * `end' is zero if we're doing the start of the IO, 1 at the end of the |
794 | * IO. |
795 | */ |
796 | static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio, |
797 | int end, struct buffer_head *map_bh) |
798 | { |
799 | unsigned dio_blocks_per_fs_block; |
800 | unsigned this_chunk_blocks; /* In dio_blocks */ |
801 | unsigned this_chunk_bytes; |
802 | struct page *page; |
803 | |
804 | sdio->start_zero_done = 1; |
805 | if (!sdio->blkfactor || !buffer_new(map_bh)) |
806 | return; |
807 | |
808 | dio_blocks_per_fs_block = 1 << sdio->blkfactor; |
809 | this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1); |
810 | |
811 | if (!this_chunk_blocks) |
812 | return; |
813 | |
814 | /* |
815 | * We need to zero out part of an fs block. It is either at the |
816 | * beginning or the end of the fs block. |
817 | */ |
818 | if (end) |
819 | this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks; |
820 | |
821 | this_chunk_bytes = this_chunk_blocks << sdio->blkbits; |
822 | |
823 | page = ZERO_PAGE(0); |
824 | if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes, |
825 | sdio->next_block_for_io, map_bh)) |
826 | return; |
827 | |
828 | sdio->next_block_for_io += this_chunk_blocks; |
829 | } |
830 | |
831 | /* |
832 | * Walk the user pages, and the file, mapping blocks to disk and generating |
833 | * a sequence of (page,offset,len,block) mappings. These mappings are injected |
834 | * into submit_page_section(), which takes care of the next stage of submission |
835 | * |
836 | * Direct IO against a blockdev is different from a file. Because we can |
837 | * happily perform page-sized but 512-byte aligned IOs. It is important that |
838 | * blockdev IO be able to have fine alignment and large sizes. |
839 | * |
840 | * So what we do is to permit the ->get_block function to populate bh.b_size |
841 | * with the size of IO which is permitted at this offset and this i_blkbits. |
842 | * |
843 | * For best results, the blockdev should be set up with 512-byte i_blkbits and |
844 | * it should set b_size to PAGE_SIZE or more inside get_block(). This gives |
845 | * fine alignment but still allows this function to work in PAGE_SIZE units. |
846 | */ |
847 | static int do_direct_IO(struct dio *dio, struct dio_submit *sdio, |
848 | struct buffer_head *map_bh) |
849 | { |
850 | const unsigned blkbits = sdio->blkbits; |
851 | const unsigned blocks_per_page = PAGE_SIZE >> blkbits; |
852 | struct page *page; |
853 | unsigned block_in_page; |
854 | int ret = 0; |
855 | |
856 | /* The I/O can start at any block offset within the first page */ |
857 | block_in_page = sdio->first_block_in_page; |
858 | |
859 | while (sdio->block_in_file < sdio->final_block_in_request) { |
860 | page = dio_get_page(dio, sdio); |
861 | if (IS_ERR(page)) { |
862 | ret = PTR_ERR(page); |
863 | goto out; |
864 | } |
865 | |
866 | while (block_in_page < blocks_per_page) { |
867 | unsigned offset_in_page = block_in_page << blkbits; |
868 | unsigned this_chunk_bytes; /* # of bytes mapped */ |
869 | unsigned this_chunk_blocks; /* # of blocks */ |
870 | unsigned u; |
871 | |
872 | if (sdio->blocks_available == 0) { |
873 | /* |
874 | * Need to go and map some more disk |
875 | */ |
876 | unsigned long blkmask; |
877 | unsigned long dio_remainder; |
878 | |
879 | ret = get_more_blocks(dio, sdio, map_bh); |
880 | if (ret) { |
881 | page_cache_release(page); |
882 | goto out; |
883 | } |
884 | if (!buffer_mapped(map_bh)) |
885 | goto do_holes; |
886 | |
887 | sdio->blocks_available = |
888 | map_bh->b_size >> sdio->blkbits; |
889 | sdio->next_block_for_io = |
890 | map_bh->b_blocknr << sdio->blkfactor; |
891 | if (buffer_new(map_bh)) |
892 | clean_blockdev_aliases(dio, map_bh); |
893 | |
894 | if (!sdio->blkfactor) |
895 | goto do_holes; |
896 | |
897 | blkmask = (1 << sdio->blkfactor) - 1; |
898 | dio_remainder = (sdio->block_in_file & blkmask); |
899 | |
900 | /* |
901 | * If we are at the start of IO and that IO |
902 | * starts partway into a fs-block, |
903 | * dio_remainder will be non-zero. If the IO |
904 | * is a read then we can simply advance the IO |
905 | * cursor to the first block which is to be |
906 | * read. But if the IO is a write and the |
907 | * block was newly allocated we cannot do that; |
908 | * the start of the fs block must be zeroed out |
909 | * on-disk |
910 | */ |
911 | if (!buffer_new(map_bh)) |
912 | sdio->next_block_for_io += dio_remainder; |
913 | sdio->blocks_available -= dio_remainder; |
914 | } |
915 | do_holes: |
916 | /* Handle holes */ |
917 | if (!buffer_mapped(map_bh)) { |
918 | loff_t i_size_aligned; |
919 | |
920 | /* AKPM: eargh, -ENOTBLK is a hack */ |
921 | if (dio->rw & WRITE) { |
922 | page_cache_release(page); |
923 | return -ENOTBLK; |
924 | } |
925 | |
926 | /* |
927 | * Be sure to account for a partial block as the |
928 | * last block in the file |
929 | */ |
930 | i_size_aligned = ALIGN(i_size_read(dio->inode), |
931 | 1 << blkbits); |
932 | if (sdio->block_in_file >= |
933 | i_size_aligned >> blkbits) { |
934 | /* We hit eof */ |
935 | page_cache_release(page); |
936 | goto out; |
937 | } |
938 | zero_user(page, block_in_page << blkbits, |
939 | 1 << blkbits); |
940 | sdio->block_in_file++; |
941 | block_in_page++; |
942 | goto next_block; |
943 | } |
944 | |
945 | /* |
946 | * If we're performing IO which has an alignment which |
947 | * is finer than the underlying fs, go check to see if |
948 | * we must zero out the start of this block. |
949 | */ |
950 | if (unlikely(sdio->blkfactor && !sdio->start_zero_done)) |
951 | dio_zero_block(dio, sdio, 0, map_bh); |
952 | |
953 | /* |
954 | * Work out, in this_chunk_blocks, how much disk we |
955 | * can add to this page |
956 | */ |
957 | this_chunk_blocks = sdio->blocks_available; |
958 | u = (PAGE_SIZE - offset_in_page) >> blkbits; |
959 | if (this_chunk_blocks > u) |
960 | this_chunk_blocks = u; |
961 | u = sdio->final_block_in_request - sdio->block_in_file; |
962 | if (this_chunk_blocks > u) |
963 | this_chunk_blocks = u; |
964 | this_chunk_bytes = this_chunk_blocks << blkbits; |
965 | BUG_ON(this_chunk_bytes == 0); |
966 | |
967 | if (this_chunk_blocks == sdio->blocks_available) |
968 | sdio->boundary = buffer_boundary(map_bh); |
969 | ret = submit_page_section(dio, sdio, page, |
970 | offset_in_page, |
971 | this_chunk_bytes, |
972 | sdio->next_block_for_io, |
973 | map_bh); |
974 | if (ret) { |
975 | page_cache_release(page); |
976 | goto out; |
977 | } |
978 | sdio->next_block_for_io += this_chunk_blocks; |
979 | |
980 | sdio->block_in_file += this_chunk_blocks; |
981 | block_in_page += this_chunk_blocks; |
982 | sdio->blocks_available -= this_chunk_blocks; |
983 | next_block: |
984 | BUG_ON(sdio->block_in_file > sdio->final_block_in_request); |
985 | if (sdio->block_in_file == sdio->final_block_in_request) |
986 | break; |
987 | } |
988 | |
989 | /* Drop the ref which was taken in get_user_pages() */ |
990 | page_cache_release(page); |
991 | block_in_page = 0; |
992 | } |
993 | out: |
994 | return ret; |
995 | } |
996 | |
997 | static inline int drop_refcount(struct dio *dio) |
998 | { |
999 | int ret2; |
1000 | unsigned long flags; |
1001 | |
1002 | /* |
1003 | * Sync will always be dropping the final ref and completing the |
1004 | * operation. AIO can if it was a broken operation described above or |
1005 | * in fact if all the bios race to complete before we get here. In |
1006 | * that case dio_complete() translates the EIOCBQUEUED into the proper |
1007 | * return code that the caller will hand to aio_complete(). |
1008 | * |
1009 | * This is managed by the bio_lock instead of being an atomic_t so that |
1010 | * completion paths can drop their ref and use the remaining count to |
1011 | * decide to wake the submission path atomically. |
1012 | */ |
1013 | spin_lock_irqsave(&dio->bio_lock, flags); |
1014 | ret2 = --dio->refcount; |
1015 | spin_unlock_irqrestore(&dio->bio_lock, flags); |
1016 | return ret2; |
1017 | } |
1018 | |
1019 | /* |
1020 | * This is a library function for use by filesystem drivers. |
1021 | * |
1022 | * The locking rules are governed by the flags parameter: |
1023 | * - if the flags value contains DIO_LOCKING we use a fancy locking |
1024 | * scheme for dumb filesystems. |
1025 | * For writes this function is called under i_mutex and returns with |
1026 | * i_mutex held, for reads, i_mutex is not held on entry, but it is |
1027 | * taken and dropped again before returning. |
1028 | * - if the flags value does NOT contain DIO_LOCKING we don't use any |
1029 | * internal locking but rather rely on the filesystem to synchronize |
1030 | * direct I/O reads/writes versus each other and truncate. |
1031 | * |
1032 | * To help with locking against truncate we incremented the i_dio_count |
1033 | * counter before starting direct I/O, and decrement it once we are done. |
1034 | * Truncate can wait for it to reach zero to provide exclusion. It is |
1035 | * expected that filesystem provide exclusion between new direct I/O |
1036 | * and truncates. For DIO_LOCKING filesystems this is done by i_mutex, |
1037 | * but other filesystems need to take care of this on their own. |
1038 | * |
1039 | * NOTE: if you pass "sdio" to anything by pointer make sure that function |
1040 | * is always inlined. Otherwise gcc is unable to split the structure into |
1041 | * individual fields and will generate much worse code. This is important |
1042 | * for the whole file. |
1043 | */ |
1044 | static inline ssize_t |
1045 | do_blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode, |
1046 | struct block_device *bdev, const struct iovec *iov, loff_t offset, |
1047 | unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io, |
1048 | dio_submit_t submit_io, int flags) |
1049 | { |
1050 | int seg; |
1051 | size_t size; |
1052 | unsigned long addr; |
1053 | unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits); |
1054 | unsigned blkbits = i_blkbits; |
1055 | unsigned blocksize_mask = (1 << blkbits) - 1; |
1056 | ssize_t retval = -EINVAL; |
1057 | loff_t end = offset; |
1058 | struct dio *dio; |
1059 | struct dio_submit sdio = { 0, }; |
1060 | unsigned long user_addr; |
1061 | size_t bytes; |
1062 | struct buffer_head map_bh = { 0, }; |
1063 | struct blk_plug plug; |
1064 | |
1065 | if (rw & WRITE) |
1066 | rw = WRITE_ODIRECT; |
1067 | |
1068 | /* |
1069 | * Avoid references to bdev if not absolutely needed to give |
1070 | * the early prefetch in the caller enough time. |
1071 | */ |
1072 | |
1073 | if (offset & blocksize_mask) { |
1074 | if (bdev) |
1075 | blkbits = blksize_bits(bdev_logical_block_size(bdev)); |
1076 | blocksize_mask = (1 << blkbits) - 1; |
1077 | if (offset & blocksize_mask) |
1078 | goto out; |
1079 | } |
1080 | |
1081 | /* Check the memory alignment. Blocks cannot straddle pages */ |
1082 | for (seg = 0; seg < nr_segs; seg++) { |
1083 | addr = (unsigned long)iov[seg].iov_base; |
1084 | size = iov[seg].iov_len; |
1085 | end += size; |
1086 | if (unlikely((addr & blocksize_mask) || |
1087 | (size & blocksize_mask))) { |
1088 | if (bdev) |
1089 | blkbits = blksize_bits( |
1090 | bdev_logical_block_size(bdev)); |
1091 | blocksize_mask = (1 << blkbits) - 1; |
1092 | if ((addr & blocksize_mask) || (size & blocksize_mask)) |
1093 | goto out; |
1094 | } |
1095 | } |
1096 | |
1097 | /* watch out for a 0 len io from a tricksy fs */ |
1098 | if (rw == READ && end == offset) |
1099 | return 0; |
1100 | |
1101 | dio = kmem_cache_alloc(dio_cache, GFP_KERNEL); |
1102 | retval = -ENOMEM; |
1103 | if (!dio) |
1104 | goto out; |
1105 | /* |
1106 | * Believe it or not, zeroing out the page array caused a .5% |
1107 | * performance regression in a database benchmark. So, we take |
1108 | * care to only zero out what's needed. |
1109 | */ |
1110 | memset(dio, 0, offsetof(struct dio, pages)); |
1111 | |
1112 | dio->flags = flags; |
1113 | if (dio->flags & DIO_LOCKING) { |
1114 | if (rw == READ) { |
1115 | struct address_space *mapping = |
1116 | iocb->ki_filp->f_mapping; |
1117 | |
1118 | /* will be released by direct_io_worker */ |
1119 | mutex_lock(&inode->i_mutex); |
1120 | |
1121 | retval = filemap_write_and_wait_range(mapping, offset, |
1122 | end - 1); |
1123 | if (retval) { |
1124 | mutex_unlock(&inode->i_mutex); |
1125 | kmem_cache_free(dio_cache, dio); |
1126 | goto out; |
1127 | } |
1128 | } |
1129 | } |
1130 | |
1131 | /* |
1132 | * Will be decremented at I/O completion time. |
1133 | */ |
1134 | atomic_inc(&inode->i_dio_count); |
1135 | |
1136 | /* |
1137 | * For file extending writes updating i_size before data |
1138 | * writeouts complete can expose uninitialized blocks. So |
1139 | * even for AIO, we need to wait for i/o to complete before |
1140 | * returning in this case. |
1141 | */ |
1142 | dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) && |
1143 | (end > i_size_read(inode))); |
1144 | |
1145 | retval = 0; |
1146 | |
1147 | dio->inode = inode; |
1148 | dio->rw = rw; |
1149 | sdio.blkbits = blkbits; |
1150 | sdio.blkfactor = i_blkbits - blkbits; |
1151 | sdio.block_in_file = offset >> blkbits; |
1152 | |
1153 | sdio.get_block = get_block; |
1154 | dio->end_io = end_io; |
1155 | sdio.submit_io = submit_io; |
1156 | sdio.final_block_in_bio = -1; |
1157 | sdio.next_block_for_io = -1; |
1158 | |
1159 | dio->iocb = iocb; |
1160 | dio->i_size = i_size_read(inode); |
1161 | |
1162 | spin_lock_init(&dio->bio_lock); |
1163 | dio->refcount = 1; |
1164 | |
1165 | /* |
1166 | * In case of non-aligned buffers, we may need 2 more |
1167 | * pages since we need to zero out first and last block. |
1168 | */ |
1169 | if (unlikely(sdio.blkfactor)) |
1170 | sdio.pages_in_io = 2; |
1171 | |
1172 | for (seg = 0; seg < nr_segs; seg++) { |
1173 | user_addr = (unsigned long)iov[seg].iov_base; |
1174 | sdio.pages_in_io += |
1175 | ((user_addr + iov[seg].iov_len + PAGE_SIZE-1) / |
1176 | PAGE_SIZE - user_addr / PAGE_SIZE); |
1177 | } |
1178 | |
1179 | blk_start_plug(&plug); |
1180 | |
1181 | for (seg = 0; seg < nr_segs; seg++) { |
1182 | user_addr = (unsigned long)iov[seg].iov_base; |
1183 | sdio.size += bytes = iov[seg].iov_len; |
1184 | |
1185 | /* Index into the first page of the first block */ |
1186 | sdio.first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits; |
1187 | sdio.final_block_in_request = sdio.block_in_file + |
1188 | (bytes >> blkbits); |
1189 | /* Page fetching state */ |
1190 | sdio.head = 0; |
1191 | sdio.tail = 0; |
1192 | sdio.curr_page = 0; |
1193 | |
1194 | sdio.total_pages = 0; |
1195 | if (user_addr & (PAGE_SIZE-1)) { |
1196 | sdio.total_pages++; |
1197 | bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1)); |
1198 | } |
1199 | sdio.total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE; |
1200 | sdio.curr_user_address = user_addr; |
1201 | |
1202 | retval = do_direct_IO(dio, &sdio, &map_bh); |
1203 | |
1204 | dio->result += iov[seg].iov_len - |
1205 | ((sdio.final_block_in_request - sdio.block_in_file) << |
1206 | blkbits); |
1207 | |
1208 | if (retval) { |
1209 | dio_cleanup(dio, &sdio); |
1210 | break; |
1211 | } |
1212 | } /* end iovec loop */ |
1213 | |
1214 | if (retval == -ENOTBLK) { |
1215 | /* |
1216 | * The remaining part of the request will be |
1217 | * be handled by buffered I/O when we return |
1218 | */ |
1219 | retval = 0; |
1220 | } |
1221 | /* |
1222 | * There may be some unwritten disk at the end of a part-written |
1223 | * fs-block-sized block. Go zero that now. |
1224 | */ |
1225 | dio_zero_block(dio, &sdio, 1, &map_bh); |
1226 | |
1227 | if (sdio.cur_page) { |
1228 | ssize_t ret2; |
1229 | |
1230 | ret2 = dio_send_cur_page(dio, &sdio, &map_bh); |
1231 | if (retval == 0) |
1232 | retval = ret2; |
1233 | page_cache_release(sdio.cur_page); |
1234 | sdio.cur_page = NULL; |
1235 | } |
1236 | if (sdio.bio) |
1237 | dio_bio_submit(dio, &sdio); |
1238 | |
1239 | blk_finish_plug(&plug); |
1240 | |
1241 | /* |
1242 | * It is possible that, we return short IO due to end of file. |
1243 | * In that case, we need to release all the pages we got hold on. |
1244 | */ |
1245 | dio_cleanup(dio, &sdio); |
1246 | |
1247 | /* |
1248 | * All block lookups have been performed. For READ requests |
1249 | * we can let i_mutex go now that its achieved its purpose |
1250 | * of protecting us from looking up uninitialized blocks. |
1251 | */ |
1252 | if (rw == READ && (dio->flags & DIO_LOCKING)) |
1253 | mutex_unlock(&dio->inode->i_mutex); |
1254 | |
1255 | /* |
1256 | * The only time we want to leave bios in flight is when a successful |
1257 | * partial aio read or full aio write have been setup. In that case |
1258 | * bio completion will call aio_complete. The only time it's safe to |
1259 | * call aio_complete is when we return -EIOCBQUEUED, so we key on that. |
1260 | * This had *better* be the only place that raises -EIOCBQUEUED. |
1261 | */ |
1262 | BUG_ON(retval == -EIOCBQUEUED); |
1263 | if (dio->is_async && retval == 0 && dio->result && |
1264 | ((rw == READ) || (dio->result == sdio.size))) |
1265 | retval = -EIOCBQUEUED; |
1266 | |
1267 | if (retval != -EIOCBQUEUED) |
1268 | dio_await_completion(dio); |
1269 | |
1270 | if (drop_refcount(dio) == 0) { |
1271 | retval = dio_complete(dio, offset, retval, false); |
1272 | kmem_cache_free(dio_cache, dio); |
1273 | } else |
1274 | BUG_ON(retval != -EIOCBQUEUED); |
1275 | |
1276 | out: |
1277 | return retval; |
1278 | } |
1279 | |
1280 | ssize_t |
1281 | __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode, |
1282 | struct block_device *bdev, const struct iovec *iov, loff_t offset, |
1283 | unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io, |
1284 | dio_submit_t submit_io, int flags) |
1285 | { |
1286 | /* |
1287 | * The block device state is needed in the end to finally |
1288 | * submit everything. Since it's likely to be cache cold |
1289 | * prefetch it here as first thing to hide some of the |
1290 | * latency. |
1291 | * |
1292 | * Attempt to prefetch the pieces we likely need later. |
1293 | */ |
1294 | prefetch(&bdev->bd_disk->part_tbl); |
1295 | prefetch(bdev->bd_queue); |
1296 | prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES); |
1297 | |
1298 | return do_blockdev_direct_IO(rw, iocb, inode, bdev, iov, offset, |
1299 | nr_segs, get_block, end_io, |
1300 | submit_io, flags); |
1301 | } |
1302 | |
1303 | EXPORT_SYMBOL(__blockdev_direct_IO); |
1304 | |
1305 | static __init int dio_init(void) |
1306 | { |
1307 | dio_cache = KMEM_CACHE(dio, SLAB_PANIC); |
1308 | return 0; |
1309 | } |
1310 | module_init(dio_init) |
1311 |
Branches:
ben-wpan
ben-wpan-stefan
javiroman/ks7010
jz-2.6.34
jz-2.6.34-rc5
jz-2.6.34-rc6
jz-2.6.34-rc7
jz-2.6.35
jz-2.6.36
jz-2.6.37
jz-2.6.38
jz-2.6.39
jz-3.0
jz-3.1
jz-3.11
jz-3.12
jz-3.13
jz-3.15
jz-3.16
jz-3.18-dt
jz-3.2
jz-3.3
jz-3.4
jz-3.5
jz-3.6
jz-3.6-rc2-pwm
jz-3.9
jz-3.9-clk
jz-3.9-rc8
jz47xx
jz47xx-2.6.38
master
Tags:
od-2011-09-04
od-2011-09-18
v2.6.34-rc5
v2.6.34-rc6
v2.6.34-rc7
v3.9