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